Privacy-Preserving Federated Learning (PPFL) is a decentralized machine learning approach where multiple clients train a model collaboratively. PPFL preserves the privacy and security of a client’s data without exchanging it. However, ensuring that data at each client is of high quality and ready for federated learning (FL) is a challenge due to restricted data access. In this paper, we introduce CADRE (Customizable Assurance of Data REadiness) for federated learning (FL), a novel framework that allows users to define custom data readiness (DR) metrics, rules, and remedies tailored to specific FL tasks. CADRE generates comprehensive DR reports based on the user-defined metrics, rules, and remedies to ensure datasets are prepared for FL while preserving privacy. We demonstrate a practical application of CADRE by integrating it into an existing PPFL framework. We conducted experiments across six datasets and addressed seven different DR issues. The results illustrate the versatility and effectiveness of CADRE in ensuring DR across various dimensions, including data quality, privacy, and fairness. This approach enhances the performance and reliability of FL models as well as utilizes valuable resources.

Efficient execution of range queries and exact queries on metadata is critical for many scientific workflows on large-scale high-performance computing (HPC) systems. Range queries in a distributed setting pose a challenge, as typical partitioning methods such as hash-based, list-based, and range-based partitioning do not work efficiently. There are also challenges in balancing the workload using locality-sensitive partitioning methods, which are capable of routing range queries to a subset of nodes to achieve efficiency. In this paper, we study prefix-hash tree and range-hash partitioning methods that fit the requirements of efficient range queries in distributed HPC systems. We also proposed a novel load-balancing method inspired by the Gossip protocol, which has similar complexity for most operations while keeping the coefficient of variation (CV) less than 7% across diverse metadata workload distributions, i.e., normal, uniform, and exponential. Our proposed methods outperform the commonly used Key-Value store, RocksDB, for exact queries by up to$15 \times$and for range queries by up to$120,000 \times$.

As the complexity of the HPC storage stack rapidly grows, domain scientists face increasing challenges in effectively utilizing HPC storage systems to achieve their desired I/O performance. To identify and address I/O issues, scientists largely rely on I/O experts to analyze their I/O traces and provide insights into potential problems. However, with a limited number of I/O experts and the growing demand for dataintensive applications, inaccessibility has become a major bottleneck, hindering scientists from maximizing their productivity. The recent rapid progress in large language models (LLMs) opens the door to creating an automated tool that democratizes trustworthy I/O performance diagnosis capabilities to domain scientists. However, LLMs face significant challenges in this task, such as the inability to handle long context windows, a lack of accurate domain knowledge about HPC I/O, and the generation of hallucinations during complex interactions. In this work, we propose IOAgent as a systematic effort to address these challenges. IOAgent integrates various new designs, including a module-based pre-processor, a RAG-based domain knowledge integrator, and a tree-based merger to accurately diagnose I/O issues from a given Darshan trace file. Similar to an I/O expert, IOAgent provides detailed justifications and references for its diagnoses and offers an interactive interface for scientists to continue asking questions about the diagnosis. To evaluate IOAgent, we collected a diverse set of labeled job traces and released the first open diagnosis test suite, TraceBench. Based on this test suite, extensive evaluations were conducted, demonstrating that IOAgent matches or outperforms state-of-the-art I/O diagnosis tools with accurate and useful diagnosis results. We also show that IOAgent is not tied to specific LLMs, performing similarly well with both proprietary and open-source LLMs. We believe IOAgent has the potential to become a powerful tool for scientists navigating complex HPC I/O subsystems in the future.

Scientific workflows are evolving from relying on a monolithic storage subsystem at a single High-Performance Computing (HPC) facility to using geographically distributed file systems, repositories, and cloud storage. As a result, storing, accessing, transferring, and managing scientific data have become highly complex and prone to performance inefficiencies. This paper delves into these challenges by exploring an optimized end-to-end interface designed to seamlessly connect various local and remote storage systems, enabling efficient data movement of objects across HPC–Cloud and HPC–HPC environments. We showcase this capability through an object-focused data management runtime system, discuss the effects of relaxed consistency semantics in distributed object scenarios, and illustrate its application in an earthquake simulation workflow. Besides reducing the amount of data by selectively transferring regions of interest, our facility-local results achieved a speedup of 45 × over an optimized HDF5 usage and 15 × over the HDF5 with caching by using the new interface in PDC-XF.

Existing similarity search methods, often limited to scalar or vector data, struggle to identify complex patterns found in scientific datasets, such as 2D seismic events or 3D magnetic flux ropes. We introduce TensorSearch, a novel parallel similarity search paradigm designed to identify known patterns in high-dimensional tensors. By directly employing tensor representations, TensorSearch captures intricate pattern structures more effectively than traditional vector-based approaches. Furthermore, its parallel architecture optimizes cache and I/O operations, enabling efficient processing of large-scale scientific data. Our performance evaluations demonstrate that TensorSearch outperforms state-of-the-art vector-based systems like Milvus by up to 10x, and achieves up to a remarkable 55x advantage over custom solution developed in Matlab used by the domain scientists. In these tests, TensorSearch exhibits linear scalability, supporting up to 2240 CPU cores.

With the rapid growth of accumulated data from various scientific domains, traditional data management systems face challenges in supporting complicated queries, such as pattern search, on massive amounts of data. To serve sophisticated queries through capturing precise features from data, recent data management systems seek to use artificial intelligence (AI) within the querying process. However, the characteristic of AI inference workflow within the querying process, such as intensive computation and expensive requirements for computing resources, becomes a bottleneck of the AI-powered query systems.In this paper, we provide a generalization of AI inference workflow in the context of AI-powered data discovery and we introduce three different caching strategies corresponding to each stage in the AI inference workflow. We provide in-depth performance evaluation on the impact of these caching strategies through a series of strong scaling experiments. Our experimental results show that the AI-powered data querying performance can be significantly improved by applying different caching strategies.

Affix-oriented metadata search is one of the essential fuzzy search capabilities that allow users to find data of interest in their voluminous data set with incomplete query conditions. With the recent transition towards object-centric data management systems in the science community, there is a paramount need for the support of such features in distributed settings. However, existing metadata search solutions either do not support efficient affix-oriented metadata search or do not suit well in a distributed setting of object-centric data management systems. To bridge this gap, we introduce IDIOMS, a metadata search solution underpinned by a distributed metadata index, meticulously designed to enable high-performance affix-oriented metadata search for parallel object-centric storage. One of the standout features of IDIOMS is its efficiency in supporting four distinct types of highly demanded metadata queries. Furthermore, IDIOMS is flexibly catering to both independent and collective metadata search operations. Our experimental comparisons with SoMeta, a state-of-the-art metadata query method, demonstrate more than 400× performance boost for independent queries and up to 300× performance improvements for collective queries, while keeping a small index management overhead.

Effectively leveraging the complex software and hardware I/O stacks of HPC systems to deliver needed I/O performance has been a challenging task for domain scientists. To identify and address I/O issues in their applications, scientists largely rely on I/O experts to analyze the recorded I/O traces of their applications and provide insights into the potential issues. However, due to the limited number of I/O experts and the growing demand for data-intensive applications across the wide spectrum of sciences, inaccessibility has become a major bottleneck hindering scientists from maximizing their productivity. Inspired by the recent rapid progress of large language models (LLMs), in this work we propose IO Navigator (ION), an LLM-based framework that takes a recorded I/O trace of an application as input and leverages the in-context learning, chain-of-thought, and code generation capabilities of LLMs to comprehensively analyze the I/O trace and provide diagnosis of potential I/O issues. Similar to an I/O expert, ION provides detailed justifications for the diagnosis and an interactive interface for scientists to ask detailed questions about the diagnosis. We illustrate ION's applicability by assessing it on a set of controlled I/O traces generated with different I/O issues. We also demonstrate that ION can match state-of-the-art I/O optimization tools and provide more insightful and adaptive diagnoses for real applications. We believe ION, with its full capabilities, has the potential to become a powerful tool for scientists to navigate through complex I/O subsystems in the future.

I/O performance monitoring tools such as Darshan and Recorder collect I/O-related metrics on production systems and help understand the applications’ behavior. However, some gaps prevent end-users from seeing the whole picture when it comes to detecting and drilling down to the root causes of I/O performance slowdowns and where those problems originate. These gaps arise from limitations in the available metrics, their collection strategy, and the lack of translation to actionable items that could advise on optimizations. This paper highlights such gaps and proposes solutions to drill down to the source code level to pinpoint the root causes of I/O bottlenecks scientific applications face by relying on cross-layer analysis combining multiple performance metrics related to I/O software layers. We demonstrate with two real applications how metrics collected in high-level libraries (which are closer to the data models used by an application), enhanced by source-code insights and natural language translations, can help streamline the understanding of I/O behavior and provide guidance to end-users, developers, and supercomputing facilities on how to improve I/O performance. Using this cross-layer analysis and the heuristic recommendations, we attained up to 6.9× speedup from run-as-is executions.

I/O operations are a known performance bottleneck of HPC applications. To achieve good performance, users often employ an iterative multistage tuning process to find an optimal I/O stack configuration. However, an I/O stack contains multiple layers, such as high-level I/O libraries, I/O middleware, and parallel file systems, and each layer has many parameters. These parameters and layers are entangled and influenced by each other. The tuning process is time-consuming and complex. In this work, we present TunIO, an AI-powered I/O tuning framework that implements several techniques to balance the tuning cost and performance gain, including tuning the high-impact parameters first. Furthermore, TunIO analyzes the application source code to extract its I/O kernel while retaining all statements necessary to perform I/O. It utilizes a smart selection of high-impact configuration parameters of the given tuning objective. Finally, it uses a novel Reinforcement Learning (RL)-driven early stopping mechanism to balance the cost and performance gain. Experimental results show that TunIO leads to a reduction of up to ≈73% in tuning time while achieving the same performance gain when compared to H5Tuner. It achieves a significant performance gain/cost of 208.4 MBps/min (I/O bandwidth for each minute spent in tuning) over existing approaches under our testing.

Various scientific applications with different I/O characteristics are executed in HPC systems. However, underlying parallel file systems are unaware of these characteristics of applications, and using a single fixed file layout for all applications can degrade the performance of HPC systems. In this paper, we propose A2FL, an autonomous and adaptive file layout adjustment scheme that optimizes parallel file system configurations by analyzing the access pattern of the applications. The key steps of A2FL are as follows: (1) A2FL initially intercepts the I/O operations of the application, recording their access patterns in real-time. (2) The access patterns are then transformed into a graphical representation used for predicting I/O performance and providing adjustment recommendations. (3) A2FL autonomously adjusts the file layout based on the prediction results, delivering an optimal file layout within the parallel file system. Moreover, we propose A2FL-Compound which analyzes an access pattern by dividing it into smaller components to optimize the file layout in a fine-grained manner. Our evaluations demonstrate that A2FL significantly enhances I/O performance, with improvements of up to 65.9× compared to the default file layout.

Garbage In Garbage Out is a universally agreed quote by computer scientists from various domains, including Artificial Intelligence (AI). As data is the fuel for AI, models trained on low-quality, biased data are often ineffective. Computer scientists who use AI invest a considerable amount of time and effort in preparing the data for AI. However, there are no standard methods or frameworks for assessing the “readiness” of data for AI. To provide a quantifiable assessment of the readiness of data for AI processes, we define parameters of AI data readiness and introduce AIDRIN (AI Data Readiness INspector). AIDRIN is a framework covering a broad range of readiness dimensions available in the literature that aid in evaluating the readiness of data quantitatively and qualitatively. AIDRIN uses metrics in traditional data quality assessment such as completeness, outliers, and duplicates for data evaluation. Furthermore, AIDRIN uses metrics specific to assess data for AI, such as feature importance, feature correlations, class imbalance, fairness, privacy, and FAIR (Findability, Accessibility, Interoperability, and Reusability) principle compliance. AIDRIN provides visualizations and reports to assist data scientists in further investigating the readiness of data. The AIDRIN framework enhances the efficiency of the machine learning pipeline to make informed decisions on data readiness for AI applications.

Summary Parallel I/O is a critical technique for moving data between compute and storage subsystems of supercomputers. With massive amounts of data produced or consumed by compute nodes, high‐performant parallel I/O is essential. I/O benchmarks play an important role in this process; however, there is a scarcity of I/O benchmarks representative of current workloads on HPC systems. Toward creating representative I/O kernels from real‐world applications, we have created h5bench , a set of I/O kernels that exercise hierarchical data format version 5 (HDF5) I/O on parallel file systems in numerous dimensions. Our focus on HDF5 is due to the parallel I/O library's heavy usage in various scientific applications running on supercomputing systems. The various tests benchmarked in the h5bench suite include I/O operations (read and write), data locality (arrays of basic data types and arrays of structures), array dimensionality (one‐dimensional arrays, two‐dimensional meshes, three‐dimensional cubes), I/O modes (synchronous and asynchronous). In this paper, we present the observed performance of h5bench executed along several of these dimensions on existing supercomputers (Cori and Summit) and pre‐exascale platforms (Perlmutter, Theta, and Polaris). h5bench measurements can be used to identify performance bottlenecks and their root causes and evaluate I/O optimizations. As the I/O patterns of h5bench are diverse and capture the I/O behaviors of various HPC applications, this study will be helpful to the broader supercomputing and I/O community.

Data provenance, or data lineage, describes the life cycle of data. In scientific workflows on HPC systems, scientists often seek diverse provenance (e.g., origins of data products, usage patterns of datasets). Unfortunately, existing provenance solutions cannot address the challenges due to their incompatible provenance models and/or system implementations. In this paper, we analyze four representative scientific workflows in collaboration with the domain scientists to identify concrete provenance needs. Based on the first-hand analysis, we propose a provenance framework called PROV-IO$^+$, which includes an I/O-centric provenance model for describing scientific data and the associated I/O operations and environments precisely. Moreover, we build a prototype of PROV-IO$^+$to enable end-to-end provenance support on real HPC systems with little manual effort. The PROV-IO$^+$framework can support both containerized and non-containerized workflows on different HPC platforms with flexibility in selecting various classes of provenance. Our experiments with realistic workflows show that PROV-IO$^+$can address the provenance needs of the domain scientists effectively with reasonable performance (e.g., less than 3.5% tracking overhead for most experiments). Moreover, PROV-IO$^+$outperforms a state-of-the-art system (i.e., ProvLake) in our experiments.

Finding relevant datasets can be a time-consuming and challenging task, especially for self-describing file formats. Current solutions use either exact or partial keyword matching approaches to extract and process metadata queries, but they fail to capture semantic relationships between the metadata content and query keywords. To address this challenge, we introduce PSQS, a novel parallel semantic search method for self-describing files. The method leverages parallel processing and kv2vec semantic similarity measures to retrieve semantically relevant data efficiently. Our evaluation against existing metadata search solutions shows that PSQS offers a new, efficient and effective semantic search functionality for various fields where large self-describing files are used, such as scientific data management, leading to more accurate and efficient data retrieval.

To alleviate bottlenecks in storing and accessing data on high-performance computing (HPC) systems, I/O libraries are enabling computation while data is in-transit, such as HDFS filters. For scientific applications that commonly use floating-point data, error-bounded lossy compression methods are a critical technique to significantly reduce the storage and bandwidth requirements. Thus far, deciding when and where to schedule in-transit data transformations, such as compression, has been outside the scope of I/O libraries. In this paper, we introduce Runway, a runtime framework that enables computation on in-transit data with an object storage abstraction. Runway is designed to be extensible to execute user-defined functions at runtime. In this effort, we focus on studying methods to offload data compression operations to available processing units based on latency and throughput. We compare the performance of running compression on multi-core CPUs, as well as offloading it to a GPU and a Data Processing Unit (DPU). We implement a state-of-the-art error-bounded lossy compression algorithm, SZ3, as a Runway function with a variant optimized for DPUs. We propose dynamic modeling to guide scheduling decisions for in-transit data compression. We evaluate Runway using four scientific datasets from the SDRBench benchmark suite on a the Perlmutter supercomputer at NERSC.

Manually diagnosing the I/O performance bottleneck for a single application (hereinafter referred to as the "job level'') is a tedious and error-prone procedure requiring domain scientists to have deep knowledge of complex storage systems. However, existing automatic methods for I/O performance bottleneck diagnosis have one major issue: the granularity of the analysis is at the platform or group level and the diagnosis results cannot be applied to the individual application. To address this issue, we designed and developed a method named "Artificial Intelligence for I/O" (AIIO), which uses AI and its interpretation technology to diagnose I/O performance bottlenecks at the job level automatically. By considering the sparsity of I/O log files, employing multiple AI models for performance prediction, merging diagnosis results across multiple models, and generalizing its performance prediction and diagnosis functions, AIIO can accurately and robustly identify the bottleneck of an even unseen application. Experimental results show that real and unseen applications can use the diagnosis results from AIIO to improve their I/O performance by at most 146 times.

Parallel I/O is an effective method to optimize data movement between memory and storage for many scientific applications. Poor performance of traditional disk-based file systems has led to the design of I/O libraries which take advantage of faster memory layers, such as on-node memory, present in high-performance computing (HPC) systems. By allowing caching and prefetching of data for applications alternating computation and I/O phases, a faster memory layer also provides opportunities for hiding the latency of I/O phases by overlapping them with computation phases, a technique called asynchronous I/O. Since asynchronous parallel I/O in HPC systems is still in the initial stages of development, there hasn't been a systematic study of the factors affecting its performance.In this paper, we perform a systematic study of various factors affecting the performance and efficacy of asynchronous I/O, we develop a performance model to estimate the aggregate I/O bandwidth achievable by iterative applications using synchronous and asynchronous I/O based on past observations, and we evaluate the performance of the recently developed asynchronous I/O feature of a parallel I/O library (HDF5) using benchmarks and real-world science applications. Our study covers parallel file systems on two large-scale HPC systems: Summit and Cori, the former with a GPFS storage and the latter with a Lustre parallel file system.

Modern-era high performance computing (HPC) systems are providing multiple levels of memory and storage layers to bridge the performance gap between fast memory and slow disk-based storage system managed by Lustre or GPFS. Several of the recent HPC systems are equipped with SSD and NVMe-based storage that is attached locally to compute nodes. A few systems are providing an SSD-based “burst buffer” intermediate storage layer that is accessible by all compute nodes as a single file system. Although these hardware layers are intended to reduce the latency gap between memory and disk-based long-term storage, how to utilize them has been left to the users. High-level I/O libraries, such as HDF5 and netCDF, can potentially take advantage of the node-local storage as a cache for reducing I/O latency from capacity storage. However, it is challenging to use node-local storage in parallel I/O especially for a single shared file. In this paper, we present an approach to integrate node-local storage as transparent caching or staging layers in a high-level parallel I/O library without placing the burden of managing these layers on users. We designed this to move data asynchronously between the caching storage layer and a parallel file system to overlap the data movement overhead in performing I/O with compute phases. We implement this approach as an external HDF5 Virtual Object Layer (VOL) connector, named Cache VOL. HDF5 VOL is a layer of abstraction in HDF5 that allows intercepting the public HDF5 application programming interface (API) and performing various optimizations to data movement after the interception. Existing HDF5 applications can use Cache VOL with minimal code modifications. We evaluated the performance of Cache VOL in HPC applications such as VPIC-10, and deep learning applications such as ImageNet and CosmoFlow. We show that using Cache VOL, one can achieve higher observed I/O performance, more scalable and stable I/O compared to direct I/O to the parallel file system, thus achieving faster time-to-solution in scientific simulations. While the caching approach is implemented in HDF5, the methods are applicable in other high-level I/O libraries.

Scientific computing workloads at HPC facilities have been shifting from traditional numerical simulations to AI/ML applications for training and inference while processing and producing ever-increasing amounts of scientific data. To address the growing need for increased storage capacity, lower access latency, and higher bandwidth, emerging technologies such as non-volatile memory are integrated into supercomputer I/O subsystems. With these emerging trends, we need a better understanding of the multilayer supercomputer I/O systems and ways to use these subsystems efficiently. In this work, we study the I/O access patterns and performance characteristics of two representative supercomputer I/O subsystems. Through an extensive analysis of year-long I/O logs on each system, we report new observations in I/O reads and writes, unbalanced use of storage system layers, and new trends in user behaviors at the HPC I/O middleware stack.

cData provenance, or data lineage, describes the life cycle of data. In scientific workflows on HPC systems, scientists often seek diverse provenance (e.g., origins of data products, usage patterns of datasets). Unfortunately, existing provenance solutions cannot address the challenges due to their incompatible provenance models and/or system implementations.

Error-bounded lossy compression is one of the most effective techniques for reducing scientific data sizes. However, the traditional trial-and-error approach used to configure lossy compressors for finding the optimal trade-off between reconstructed data quality and compression ratio is prohibitively expensive. To resolve this issue, we develop a general-purpose analytical ratio-quality model based on the prediction-based lossy compression framework, which can effectively foresee the reduced data quality and compression ratio, as well as the impact of lossy compressed data on post-hoc analysis quality. Our analytical model significantly improves the prediction-based lossy compression in three use-cases: (1) optimization of predictor by selecting the best-fit predictor; (2) memory compression with a target ratio; and (3) in-situ compression optimization by fine-grained tuning error-bounds for various data partitions. We evaluate our analytical model on 10 scientific datasets, demonstrating its high accuracy (93.47% accuracy on average) and low computational cost (up to 18.7x lower than the trial-and-error approach) for estimating the compression ratio and the impact of lossy compression on post-hoc analysis quality. We also verify the high efficiency of our ratio-quality model using different applications across the three use-cases. In addition, our experiment demonstrates that our modeling-based approach reduces the time to store the 3D RTM data with HDF5 by up to 3.4 x with 128 CPU cores over the traditional solution.

Lossy compression is one of the most efficient solutions to reduce storage overhead and improve I/O performance for HPC applications. However, existing parallel I/O libraries cannot fully utilize lossy compression to accelerate parallel write due to the lack of deep understanding on compression-write performance. To this end, we propose to deeply integrate predictive lossy compression with HDF5 to significantly improve the parallel-write performance. Specifically, we propose analytical models to predict the time of compression and parallel write before the actual compression to enable compression-write overlapping. We also introduce an extra space in the process to handle possible data overflows resulting from prediction uncertainty in compression ratios. Moreover, we propose an optimization to reorder the compression tasks to increase the overlapping efficiency. Experiments with up to 4,096 cores from Summit show that our solution improves the write performance by up to$4.5\times$and$2.9\times$over the non-compression and lossy compression solutions, respectively, with only 1.5% storage overhead (compared to original data) on two real-world HPC applications.

Moving toward exascale computing, the size of data stored and accessed by applications is ever increasing. However, traditional disk-based storage has not seen improvements that keep up with the explosion of data volume or the speed of processors. Multiple levels of non-volatile storage devices are being added to handle bursty I/O, however, moving data across the storage hierarchy can take longer than the data generation or analysis. Asynchronous I/O can reduce the impact of I/O latency as it allows applications to schedule I/O early and to check their status later. I/O is thus overlapped with application communication or computation or both, effectively hiding some or all of the I/O latency. POSIX and MPI-I/O provide asynchronous read and write operations, but lack the support for non-data operations such as file open and close. Users also have to manually manage data dependencies and use low-level byte offsets, which requires significant effort and expertise to adopt. In this article, we present an asynchronous I/O framework that supports all types of I/O operations, manages data dependencies transparently and automatically, provides implicit and explicit modes for application flexibility, and error information retrieval. We implemented these techniques in HDF5. Our evaluation of several benchmarks and application workloads demonstrates it effectiveness on hiding the I/O cost from the application.

Accessing non-contiguous blocks in multiple array variables is a challenging I/O pattern for parallel applications to obtain good I/O performance. High-level I/O libraries such as HDF5 allow users to implement this pattern conveniently, but users have observed significant performance bottlenecks in the two-phase I/O implementation of MPI-IO. Recent studies have advanced the two-phase I/O performance by novel communication algorithms, but such improvements still have limitations. Two-phase I/O has to faithfully process inputs from high-level I/O libraries, so that implementation overheads can accumulate for improper usage of high-level I/O libraries. In this paper, we propose approaches for efficient usage of high-level I/O libraries that can circumvent major collective I/O overheads. We adopt a multi-dataset implementation of HDF5 dataset I/O to aggregate non-contiguous requests for array blocks and provide corresponding parameter assignment strategies. These approaches reduce the overheads caused by communication straggler effects in two-phase I/O. We show that our proposed methods can improve the parallel I/O performance up to 8× on two supercomputing systems for the HDF5 implementations of an I/O kernel extracted from climate simulation code compared with its baseline implementations.

Scientific applications, such as those simulating earthquakes, the origins of universe, etc., often produce massive amounts of data as high-performance computing (HPC) systems are moving toward exascale. The ever-increasing volumes of data are posing challenges for scientists to store, share, analyze, and visualize. Compression algorithms have become a crucial component for data management in scientific workflows. Data reduction enables simulations to output more data without worrying about exceeding storage quotas, and could capture more insights in the simulation. However, due to the complexity and poor performance of I/O and compression libraries as well as parallel file systems, the overall compression and I/O performance varies significantly. In this paper, we explore tuning parallel compression of data produced by a large-scale earthquake simulation. We show that our strategies achieve up to 13X performance improvement and a compression ratio of up to 251.

Popular parallel I/O libraries, such as HDF5, provide tuning parameters to obtain superior performance. However, the selection of effective parameters on production systems is complex due to the interdependence of I/O software and file system layers. Hence, application developers typically use the default parameters and often experience poor I/O performance. This work conducts a benchmarking-based analysis on the HDF5 behaviors with a wide variety of I/O patterns to extract performance characteristics under the production workload. To make the analysis well controlled, we exercise I/O benchmarks on POSIX-IO, MPI-IO, and HDF5 using the same I/O patterns and in the same jobs. To address high performance variability in production environments, we repeat the benchmarks across I/O patterns, storage devices, and time intervals. Based on the results, we identified consistent HDF5 behaviors that appropriate configurations and operations on dataset layout and file-metadata placement can improve performance significantly. We apply our findings and evaluate the tuned I/O library on two supercomputers: Summit and Cori. The results show that our tuned parameters can achieve more than 10× I/O performance speedup than that with default parameters on both systems, suggesting the effectiveness, stability, and generality of our solution.

In recent years, the increasing complexity in scientific simulations and emerging demands for training heavy artificial intelligence models require massive and fast data accesses, which urges high-performance computing (HPC) platforms to equip with more advanced storage infrastructures such as solid-state disks (SSDs). While SSDs offer high-performance I/O, the reliability challenges faced by the HPC applications under the SSD-related failures remains unclear, in particular for failures resulting in data corruptions. The goal of this paper is to understand the impact of SSD-related faults on the behaviors of complex HPC applications. To this end, we propose FFIS, a FUSE-based fault injection framework that systematically introduces storage faults into the application layer to model the errors originated from SSDs. FFIS is able to plant different I/O related faults into the data returned from underlying file systems, which enables the investigation on the error resilience characteristics of the scientific file format. We demonstrate the use of FFIS with three representative real HPC applications, showing how each application reacts to the data corruptions, and provide insights on the error resilience of the widely adopted HDF5 file format for the HPC applications.

High-performance computing (HPC) systems consist of thousands of compute nodes, storage systems and high-speed networks, providing multiple layers of I/O stack with high complexity. By adjusting the diverse configuration settings that HPC systems provide, the I/O performance of applications can be improved. However, it is challenging to identify the optimal configuration settings without a thorough knowledge of the system, as each of the different I/O characteristics of applications can be an important factor for parameter decision. In this paper, we use multiple machine learning approaches to perform an in-depth analysis on I/O behaviors of HPC applications and to search for the optimal configuration settings for jobs sharing similar I/O characteristics. Improved by maximum 0.07 R-squared score, our results in overall show that jobs run on the HPC systems can obtain the predicted I/O performance for different configuration parameters with a high accuracy, using the proposed machine learning-based prediction models.

HPC systems typically rely on the fixed-lifetime (FLT) data retention strategy, which only considers temporal locality of data accesses to parallel file systems. However, our extensive analysis based on the leadership-class HPC system traces suggests that the FLT approach often fails to capture the dynamics in users' behavior and leads to undesired data purge. In this study, we propose an activeness-based data retention (ActiveDR) solution, which advocates considering the data retention approach from a holistic activeness-based perspective. By evaluating the frequency and impact of users' activities, ActiveDR prioritizes the file purge process for inactive users and rewards active users with extended file lifetime on parallel storage. Our extensive evaluations based on the traces of the prior Titan supercomputer show that, when reaching the same purge target, ActiveDR achieves up to 37% file miss reduction as compared to the current FLT retention methodology.

Large-scale high performance computing (HPC) systems typically consist of many thousands of CPUs and storage units, while used by hundreds to thousands of users at the same time. Applications from these large numbers of users have diverse characteristics, such as varying compute, communication, memory, and I/O intensiveness. A good understanding of the performance characteristics of each user application is important for job scheduling and resource provisioning. Among these performance characteristics, the I/O performance is difficult to predict because the I/O system software is complex, the I/O system is shared among all users, and the I/O operations also heavily rely on networking systems. To improve the prediction of the I/O performance on HPC systems, we propose to integrate information from a number of different system logs and develop a regression-based approach that dynamically selects the most relevant features from the most recent log entries, and automatically select the best regression algorithm for the prediction task. Evaluation results show that our proposed scheme can predict the I/O performance with up to 84% prediction accuracy in the case of the I/O-intensive applications using the logs from CORI supercomputer at NERSC.

Recently developed distributed acoustic sensing (DAS) technologies convert fiber-optic cables into large arrays of subsurface sensors, enabling a variety of applications including earthquake detection and environmental characterization. However, DAS systems produce voluminous datasets sampled at high spatial-temporal resolution and consequently, discovering useful geophysical knowledge within these large-scale data becomes a nearly impossible task for geophysicists. It is appealing to use supercomputers for DAS data analysis, as modern supercomputers are capable of performing over a hundred quadrillion FLOPS operations and have access to exabytes of storage space. Unfortunately, the majority of geophysical data processing libraries are not geared towards these supercomputer environments. This paper introduces a parallel DAS Data Storage and Analysis (DASSA) framework to enable easy-to-use and parallel DAS data analysis on modern supercomputers. DASSA uses a hybrid (i.e., MPI and OpenMP) data analysis execution engine that supports a user-defined function (UDF) interface for various operations and automatically parallelizes them for supercomputer execution. DASSA also provides novel data storage and access strategies, such as communication-avoiding parallel I/O, to reduce the cost of retrieving large DAS data for analysis. Compared with existing data analysis pipelines used by the geophysical community, DASSA is 16× faster and can efficiently scale up to 1456 computing nodes with 11648 CPU cores.

Many applications are increasingly becoming I/O-bound. To improve scalability, analytical models of parallel I/O performance are often consulted to determine possible I/O optimizations. However, I/O performance modeling has predominantly focused on applications that directly issue I/O requests to a parallel file system or a local storage device. These I/O models are not directly usable by applications that access data through standardized I/O libraries, such as HDF5, FITS, and NetCDF, because a single I/O request to an object can trigger a cascade of I/O operations to different storage blocks. The I/O performance characteristics of applications that rely on these libraries is a complex function of the underlying data storage model, user-configurable parameters and object-level access patterns. As a consequence, I/O optimization is predominantly an ad-hoc process that is performed by application developers, who are often domain scientists with limited desire to delve into nuances of the storage hierarchy of modern computers. This paper presents an analytical cost model to predict the end-to-end execution time of applications that perform I/O through established array management libraries. The paper focuses on the HDF5 and Zarr array libraries, as examples of I/O libraries with radically different storage models: HDF5 stores every object in one file, while Zarr creates multiple files to store different objects. We find that accessing array objects via these I/O libraries introduces new overheads and optimizations. Specifically, in addition to I/O time, it is crucial to model the cost of transforming data to a particular storage layout (memory copy cost), as well as model the benefit of accessing a software cache. We evaluate the model on real applications that process observations (neuroscience) and simulation results (plasma physics). The evaluation on three HPC clusters reveals that I/O accounts for as little as 10% of the execution time in some cases, and hence models that only focus on I/O performance cannot accurately capture the performance of applications that use standard array storage libraries. In parallel experiments, our model correctly predicts the fastest storage library between HDF5 and Zarr 94% of the time, in contrast with 70% of the time for a cutting-edge I/O model.

Summary Object storage technologies that take advantage of multitier storage on HPC systems are emerging. However, to use these technologies at present, applications have to be modified significantly from current I/O libraries. HDF5, a widely used I/O middleware on HPC systems, provides a virtual object layer (VOL) that allows applications to connect to different storage mechanisms transparently without requiring significant code modifications. We recently designed the proactive data containers (PDC) object‐centric storage system that provides the capabilities of transparent, asynchronous, and autonomous data movement taking advantage of multiple storage tiers—a decision that has so far been left upon the user on most current systems. To enable PDC's features through HDF5 without modifying application codes, we have developed an HDF5 VOL connector that interfaces with PDC. We present in this article the connector interface and evaluate its performance on Cori, a Cray XC40 supercomputer located at the National Energy Research Scientific Computing Center (NERSC). Our evaluation demonstrates up to an 8× improvement compared with HDF5 that has the most recent optimizations.

In high-performance computing, storage is a shared resource and used by all users with many different application requirements and knowledge of storage. Consequently, the optimal storage configuration varies according to the I/O behavior of each application. While system logs are helpful resources in understanding the storage behavior, it is non-trivial for each user to analyze the logs and adjust complex configurations. Even for experienced users, it is difficult to understand the full stack of I/O systems and find the optimal configuration for the specific application. In this work, we analyzed the I/O activities of CORI which is an HPC system in National Energy Research Scientific Computing Center (NERSC). The result of our analysis shows that most users do not adjust storage configurations and use the default settings. Also, it shows that only a few applications are executed repeatedly in the HPC environment. Based on this result, we have developed DCA-IO, a dynamic distributed file system configuration adjustment algorithm, which utilizes system log information and widely adapted rules to adjust storage configurations automatically without any user intervention. DCA-IO utilizes existing system logs and does not require any modifications in code or an additional library. To demonstrate the effectiveness of DCA-IO, we have performed experiments using I/O kernels of the real applications in both isolated small-sized Lustre environment and CORI. Our experimental result shows that the use of our scheme can lead to improvements in the performance of HPC applications by up to 75% in an isolated environment and 50% in a real HPC environment without user intervention.

Scientific applications frequently spend a large fraction of their execution time in reading and writing data on parallel file systems. Identifying these I/O performance bottlenecks and attributing root causes are critical steps toward devising optimization strategies. Several existing studies analyze I/O logs of a set of benchmarks or applications that were run with controlled behaviors. However, there is still a lack of general approach that systematically identifies I/O performance bottlenecks for applications running "in the wild" on production systems. In this study, we have developed an analysis approach of "zooming in" from platform-wide to application-wide to job-level I/O logs for identifying I/O bottlenecks in arbitrary scientific applications. We analyze the logs collected on a Cray XC40 system in production over a two-month period. This study results in several insights for application developers to use in optimizing I/O behavior.

Efficient management of scientific data on high-performance computing (HPC) systems has been a challenge, as it often requires knowledge of various hardware and software components of the system, as well as tedious manual effort in optimizing parallel I/O for each application. This situation is exacerbated by the fact that storage systems on upcoming exascale supercomputers are equipped with an unprecedented level of complexity due to a deep storage and memory hierarchy with heterogeneous hardware and their management software. Simple and effective data management methods are critical for numerous scientific applications that are storing and analyzing massive amounts of data on HPC systems. Object-centric data management systems (ODMS) provide an easy-to-use interface, allow for massive scalability with relaxed consistency, and have been gaining popularity in the HPC community. However, tuning an ODMS to achieve its full potential on existing HPC systems with large-scale science use cases still remains a challenging task. In this paper, we explore and evaluate various well-known I/O tuning techniques on a new ODMS called Proactive Data Containers (PDC). Our experiments using real science applications and I/O kernels demonstrate that the benefits of these tuning methods with up to 9X I/O performance speedup over the previous version of PDC, and 47X over a highly optimized HDF5 implementation.

Emerging high performance computing (HPC) systems are expected to be deployed with an unprecedented level of complexity due to a deep system memory and storage hierarchy. Efficient and scalable methods of data management and movement through the multi-level storage hierarchy of upcoming HPC systems will be critical for scientific applications at exascale. In this paper, we propose in locus analysis that allows registering user-defined functions (UDFs) and running those functions automatically while the data is moving between levels of a storage hierarchy. We implement this analysis in the data path approach in our object-centric data management system, called Proactive Data Containers (PDC). The transparent invocation of analysis functions as part of PDC object mapping is an optimized approach to minimize latency to access data as it moves within the storage hierarchy. Because a user defined analysis or transform function will be invoked automatically by the PDC runtime, the user simply registers their functions for PDC to identify the function name as well as the required list of actual parameters. To demonstrate the validity and flexibility of this analysis approach, we have implemented several scientific analysis kernels to compare against other HPC analysis-oriented approaches.

Scientific experiments and observations store massive amounts of data in various scientific file formats. Metadata, which describes the characteristics of the data, is commonly used to sift through massive datasets in order to locate data of interest to scientists. Several indexing data structures (such as hash tables, trie, self-balancing search trees, sparse array, etc.) have been developed as part of efforts to provide an efficient method for locating target data. However, efficient determination of an indexing data structure remains unclear in the context of scientific data management, due to the lack of investigation on metadata, metadata queries, and corresponding data structures. In this study, we perform a systematic study of the metadata search essentials in the context of scientific data management. We study a real-world astronomy observation dataset and explore the characteristics of the metadata in the dataset. We also study possible metadata queries based on the discovery of the metadata characteristics and evaluate different data structures for various types of metadata attributes. Our evaluation on real-world dataset suggests that trie is a suitable data structure when prefix/suffix query is required, otherwise hash table should be used. We conclude our study with a summary of our findings. These findings provide a guideline and offers insights in developing metadata indexing methodologies for scientific applications.

Scientific applications often store datasets in self-describing data file formats, such as HDF5 and netCDF. Regrettably, to efficiently search the metadata within these files remains challenging due to the sheer size of the datasets. Existing solutions extract the metadata and store it in external database management systems (DBMS) to locate desired data. However, this practice introduces significant overhead and complexity in extraction and querying. In this research, we propose a novel Metadata Indexing and Querying Service (MIQS), which removes the external DBMS and utilizes in-memory index to achieve efficient metadata searching. MIQS follows the self-contained data management paradigm and provides portable and schema-free metadata indexing and querying functionalities for self-describing file formats. We have evaluated MIQS with the state-of-the-art MongoDB-based metadata indexing solution. MIQS achieved up to 99% time reduction in index construction and up to 172kx search performance improvement with up to 75% reduction in memory footprint.

Large-scale applications typically spend a large fraction of their execution time performing I/O to a parallel storage system. However, with rapid progress in compute and storage system stack of large-scale systems, it is critical to investigate and update our understanding of the I/O behavior of large-scale applications. Toward that end, in this work, we monitor, collect and analyze a year worth of storage system data from a large-scale production parallel storage system. We perform temporal, spatial and correlative analysis of the system and uncover surprising patterns which defy existing assumptions and have important implications for future systems.

A prime question for plasma physicists is how a fraction of charged particles is accelerated to very high energy.To answer this question, physicists simulate trillions of particles with detailed dynamics and analyze their trajectories. This process requires a range of data analysis tasks with high diversity. In this paper, we present a use case of formulating various analysis tasks on terabyte-scale particle data with a novel data analysis framework called ArrayUDF. The flexibility of ArrayUDF allows it to compose a wide range of particle data operations. We also present optimization strategies to avoid frequent global reduction and to take full advantage of the data locality. Tests show that our optimization methods could accelerate these particle data analysis operations by up to 1,600 times.

Summary Many scientific applications produce very large amounts of data as advances in hardware fuel computing and experimental facilities. Managing and analyzing massive quantities of scientific data is challenging as data are often stored in specific formatted files, such as HDF5 and NetCDF, which do not offer appropriate search capabilities. In this research, we investigated a special class of search capability, called membership query, to identify whether queried elements of a set are members of an attribute. Attributes that naturally have classification values appear frequently in scientific domains such as category and object type as well as in daily life such as zip code and occupation. Because classification attribute values are discrete and require random data access, performing a membership query on a large scientific data set creates challenges. We applied bitmap indexing and parallelization to membership queries to overcome these challenges. Bitmap indexing provides high performance not only for low cardinality attributes but also for high cardinality attributes, such as floating‐point variables, electric charge, or momentum in a particle physics data set, due to compression algorithms such as Word‐Aligned Hybrid. We conducted experiments, in a highly parallelized environment, on data obtained from a particle accelerator model and a synthetic data set.

Affix-based search is a fundamental functionality for storage systems. It allows users to find desired datasets, where attributes of a dataset match an affix. While building inverted index to facilitate efficient affix-based keyword search is a common practice for standalone databases and for desktop file systems, building local indexes or adopting indexing techniques used in a standalone data store is insufficient for high-performance computing (HPC) systems due to the massive amount of data and distributed nature of the storage devices within a system. In this paper, we propose Distributed Adaptive Radix Tree (DART), to address the challenge of distributed affix-based keyword search on HPC systems. This trie-based approach is scalable in achieving efficient affix-based search and alleviating imbalanced keyword distribution and excessive requests on keywords at scale. Our evaluation at different scales shows that, comparing with the "full string hashing" use case of the most popular distributed indexing technique - Distributed Hash Table (DHT), DART achieves up to 55× better throughput with prefix search and with suffix search, while achieving comparable throughput with exact and infix searches. Also, comparing to the "initial hashing" use case of DHT, DART maintains a balanced keyword distribution on distributed nodes and alleviates excessive query workload against popular keywords.

Scientific data analysis typically involves reading massive amounts of data that was generated by simulations, experiments, and observations. Performance of reading such large volumes of data from disk-based file systems is often poor because of the slow and mechanical components in the disks. Recent supercomputing systems are adding non-volatile storage layers in a hierarchy to handle the performance gap between fast main memory and slow disk-based storage. Software libraries for managing this hierarchy not only need efficient reading of data but also reduce user-involvement for cross-layer data movement. Furthermore, these libraries need to support array data access patterns into hierarchical storage management as scientific data is often organized in array-based data structures. Existing software typically manage individual storage layers requiring significant manual process in moving data among them. In this paper, we introduce a new array caching in hierarchical storage (ARCHIE) to accelerate array data analysis in a seamless fashion. ARCHIE evaluates array access patterns and prefetches data with array semantics between storage layers. Our evaluation shows that ARCHIE outperforms state-of-the-art file systems, i.e., Lustre and DataWarp, on a production supercomputing system by up to 5.8× in accessing data by scientific analysis applications.

Emerging high performance computing (HPC) systems are expected to be deployed with an unprecedented level of complexity due to a deep system memory and storage hierarchy. Efficient and scalable methods of data management and movement through this hierarchy is critical for scientific applications using exascale systems. Moving toward new paradigms for scalable I/O in the extreme-scale era, we introduce novel object-centric data abstractions and storage mechanisms that take advantage of the deep storage hierarchy, named Proactive Data Containers (PDC). In this paper, we formulate object-centric PDCs and their mappings in different levels of the storage hierarchy. PDC adopts a client-server architecture with a set of servers managing data movement across storage layers. To demonstrate the effectiveness of the proposed PDC system, we have measured performance of benchmarks and I/O kernels from scientific simulation and analysis applications using PDC programming interface, and compared the results with existing highly tuned I/O libraries. Using asynchronous I/O along with data and metadata optimizations, PDC demonstrates up to 23× speedup over HDF5 and PLFS in writing and reading data from a plasma physics simulation. PDC achieves comparable performance with HDF5 and PLFS in reading and writing data of a single timestep at small scale, and outperforms them at a scale of larger than 10K cores. In contrast to existing storage systems, PDC offers user-space data management with the flexibility to allocate the number of PDC servers depending on the workload.

On the road to exascale, the high-performance computing (HPC) community is seeing the emergence of multi-tier storage systems. However, existing data management solutions for HPC applications are no longer suitable for handling the increased level of storage complexity and currently delegate that task back to the user. We describe a novel object-based data abstraction that takes advantage of deep memory hierarchies by providing a simplified programming interface that enables autonomous, asynchronous, and transparent data movement with a server-driven architecture. Users can define a mapping between the application memory and abstract storage objects, creating a linkage between either all or part of an object's content without data copy or transfer, avoiding explicit management of complex data movement across multiple storage hierarchies. We evaluate our system by storing plasma physics simulation data with different storage layouts.

High performance computing (HPC) architectures have been adding new layers of storage, such as burst buffers, to tolerate latency between memory and disk-based file systems. However, existing file system and burst buffer management software typically manage each storage layer separately. As a result, the burden of moving data across multiple layers falls upon HPC system users. To hide the complexity of managing the scattered storage devices from applications, we introduce UniviStor, a data management service offering a unified view of storage layers. By considering each layer's distinct characteristics, UniviStor provides performance optimizations and data structures tailored for distributed and hierarchical data placement, interference aware data movement scheduling, adaptive data striping, and lightweight workflow management. UniviStor supports parallel I/O library APIs, such as MPI-IO and HDF5. Our evaluations on a large-scale supercomputer demonstrated that UniviStor outperforms Data Elevator, a state-of-the-art transparent caching solution for burst buffers by up to 17x, and Lustre by up to 46x.

The following topics are dealt with: parallel processing; message passing; parallel machines; application program interfaces; multiprocessing systems; resource allocation; scheduling; cache storage; data analysis; graphics processing units.

Upcoming exascale high performance computing (HPC) systems are expected to comprise multi-tier storage hierarchy, and thus will necessitate innovative storage and I/O mechanisms. Traditional disk and block-based interfaces and file systems face severe challenges in utilizing capabilities of storage hierarchies due to the lack of hierarchy support and semantic interfaces. Object-based and semantically-rich data abstractions for scientific data management on large scale systems offer a sustainable solution to these challenges. Such data abstractions can also simplify users involvement in data movement. In this paper, we take the first steps of realizing such an object abstraction and explore storage mechanisms for these objects to enhance I/O performance, especially for scientific applications. We explore how an object-based interface can facilitate next generation scalable computing systems by presenting the mapping of data I/O from two real world HPC scientific use cases: a plasma physics simulation code (VPIC) and a cosmology simulation code (HACC). Our storage model stores data objects in different physical organizations to support data movement across layers of memory/storage hierarchy. Our implementation sclaes well to 16K parallel processes, and compared to the state of the art, such as MPI-IO and HDF5, our object-based data abstractions and data placement strategy in multi-level storage hierarchy achieves up to 7× I/O performance improvement for scientific data.

Scientists are increasingly turning to datacenter-scale computers to analyze massive arrays. Despite decades of database research that extols the virtues of declarative query processing, scientists still write, debug and parallelize imperative HPC programs even for the most mundane queries. This impedance mismatch is due to the cumbersome and costly data format conversions that are needed to use scientific data management tools, such as SciDB, in an HPC setting. Our goal is to make declarative array manipulations from SciDB interoperable with imperative, file-centric analyses from HDF5-based programs. This paper describes ArrayBridge, a bi-directional array view mechanism for the HDF5 file format, that allows scientists to use SciDB, TensorFlow and HDF5-based analysis code in the same file-centric pipeline without converting between file formats. In addition to fast querying over HDF5 array objects, ArrayBridge produces arrays in the HDF5 file format as easily as it can read from it. ArrayBridge also supports time travel queries from imperative codes through the unmodified HDF5 API, and automatically deduplicates between versions for space efficiency. Our performance evaluation in a large scientific computing facility shows that ArrayBridge exhibits statistically indistinguishable performance and I/O scalability to the native SciDB storage engine and is 3× faster than TileDB.

Scientific data sets, which grow rapidly in volume, are often attached with plentiful metadata, such as their associated experiment or simulation information. Thus, it becomes difficult for them to be utilized and their value is lost over time. Ideally, metadata should be managed along with its corresponding data by a single storage system, and can be accessed and updated directly. However, existing storage systems in high-performance computing (HPC) environments, such as Lustre parallel file system, still use a static metadata structure composed of non-extensible and fixed amount of information. The burden of metadata management falls upon the end-users and require ad-hoc metadata management software to be developed.With the advent of "object-centric" storage systems, there is an opportunity to solve this issue. In this paper, we present SoMeta, a scalable and decentralized metadata management approach for object-centric storage in HPC systems. It provides a flat namespace that is dynamically partitioned, a tagging approach to manage metadata that can be efficiently searched and updated, and a light-weight and fault tolerant management strategy. In our experiments, SoMeta achieves up to 3.7X speedup over Lustre in performing common metadata operations, and up to 16X faster than SciDB and MongoDB for advanced metadata operations, such as adding and searching tags. Additionally, in contrast to existing storage systems, SoMeta offers scalable user-space metadata management by allowing users with the capability to specify the number of metadata servers depending on their workload.

User-Defined Functions (UDF) allow application programmers to specify analysis operations on data, while leaving the data management tasks to the system. This general approach enables numerous custom analysis functions and is at the heart of the modern Big Data systems. Even though the UDF mechanism can theoretically support arbitrary operations, a wide variety of common operations -- such as computing the moving average of a time series, the vorticity of a fluid flow, etc., -- are hard to express and slow to execute. Since these operations are traditionally performed on multi-dimensional arrays, we propose to extend the expressiveness of structural locality for supporting UDF operations on arrays. We further propose an in situ UDF mechanism, called ArrayUDF, to implement the structural locality. ArrayUDF allows users to define computations on adjacent array cells without the use of join operations and executes the UDF directly on arrays stored in data files without requiring to load their content into a data management system. Additionally, we present a thorough theoretical analysis of the data access cost to exploit the structural locality, which enables ArrayUDF to automatically select the best array partitioning strategy for a given UDF operation. In a series of performance evaluations on large scientific datasets, we have observed that -- using the generic UDF interface -- ArrayUDF consistently outperforms Spark, SciDB, and RasDaMan.

Runtime data sharing across applications is of great importance for avoiding high I/O overhead for scientific data analytics. Sharing data on a staging space running on a set of dedicated compute nodes is faster than writing data to a slow disk-based parallel file system (PFS) and then reading it back for post-processing. Originally, the staging space has been purely based on main memory (DRAM), and thus was several orders of magnitude faster than the PFS approach. However, storing all the data produced by large-scale simulations on DRAM is impractical. Moving data from memory to SSD-based burst buffers is a potential approach to address this issue. However, SSDs are about one order of magnitude slower than DRAM. To optimize data access performance over the staging space, methods such as prefetching data from SSDs according to detected spatial access patterns and distributing data across the network topology have been explored. Although these methods work well for uniform mesh data, which they were designed for, they are not well suited for adaptive mesh refinement (AMR) data. Two major issues must be addressed before constructing such a memory hierarchy and topology-aware runtime AMR data sharing framework: (1) spatial access pattern detection and prefetching for AMR data; (2) AMR data distribution across the network topology at runtime. We propose a framework that addresses these challenges and demonstrate its effectiveness with extensive experiments on AMR data. Our results show the framework's spatial access pattern detection and prefetching methods demonstrate about 26% performance improvement for client analytical processes. Moreover, the framework's topology-aware data placement can improve overall data access performance by up to 18%.

As scientific simulations and experiments move toward extremely large scales and generate massive amounts of data, the data access performance of analytic applications becomes crucial. A mismatch often happens between write and read patterns of data accesses, typically resulting in poor read performance. Data layout reorganization has been used to improve the locality of data accesses. However, current data reorganizations are static and focus on generating a single (or set of) optimized layouts that rely on prior knowledge of exact future access patterns. We propose a framework that dynamically recognizes the data usage patterns, replicates the data of interest in multiple reorganized layouts that would benefit common read patterns, and makes runtime decisions on selecting a favorable layout for a given read pattern. This framework supports reading individual elements and chunks of a multi-dimensional array of variables. Our pattern-driven layout selection strategy achieves multi-fold speedups compared to reading from the original dataset.

Frameworks that facilitate runtime data sharingacross multiple applications are of great importance for scientificdata analytics. Although existing frameworks work well overuniform mesh data, they can not effectively handle adaptive meshrefinement (AMR) data. Among the challenges to construct anAMR-capable framework include: (1) designing an architecturethat facilitates online AMR data management, (2) achievinga load-balanced AMR data distribution for the data stagingspace at runtime, and (3) building an effective online indexto support the unique spatial data retrieval requirements forAMR data. Towards addressing these challenges to supportruntime AMR data sharing across scientific applications, wepresent the AMRZone framework. Experiments over real-worldAMR datasets demonstrate AMRZone's effectiveness at achievinga balanced workload distribution, reading/writing large-scaledatasets with thousands of parallel processes, and satisfyingqueries with spatial constraints. Moreover, AMRZone's performance and scalability are even comparable with existing state-of-the-art work when tested over uniform mesh data with up to16384 cores, in the best case, our framework achieves a 46% performance improvement.

Hierarchical storage subsystems that include multiple layers of burst buffers (BB) and disk-based parallel file systems (PFS), are becoming an essential part of HPC systems to address the I/O performance gap. However, the state-of-the-art software for managing these hierarchical storage subsystems, such as Cray DataWarp, requires user involvement in moving data among storage layers. Such manual data movement may experience poor performance because of resource contention on the I/O servers of a layer for serving data movement in the hierarchy as well as regular read/write requests. In this paper, we propose a new system, named Data Elevator, for transparently and efficiently moving data in hierarchical storage. Users specify the final destination for their data, typically a PFS. Data Elevator intercepts the I/O calls, stages data on a fast persistent storage layer (for example, an SSD-based burst buffer), and then asynchronously transfers the data to the final destination in the background. Data Elevator reduces the resource contention on BB servers via offloading the data movement from a fixed number of BB server nodes to compute nodes. The number of the compute nodes is configurable based on the data movement load. Data Elevator also allows optimizations, such as overlapping read and write operations, choosing I/O modes, and aligning buffer boundaries. In our tests with large-scale scientific applications, Data Elevator is as much as 4.2X faster than Cray DataWarp, and 4X faster than directly writing data to PFS.

Parallel sorting is an essential algorithm in large-scale data analytics using distributed memory systems. As the number of processes increases, existing parallel sorting algorithms could become inefficient because of the unbalanced workload. A common cause of load imbalance is the skewness of data, which is common in application data sets from physics, biology, earth and planetary sciences. In this work, we introduce a new scalable dynamic skew-aware parallel sorting algorithm, named SDS-Sort. It uses a skew-aware partition method to guarantee a tighter upper bound on the workload of each process. To improve load balance among parallel processes, existing algorithms usually add extra variables to the sorting key, which increase the time needed to complete the sorting operation. SDS-Sort allows a user to select any sorting key without sacrificing performance. SDS-Sort also provides optimizations, including adaptive local merging, overlapping of data exchange and data processing, and dynamic selection of data processing algorithms for different hardware configurations and for partially ordered data. SDS-Sort uses local-sampling based partitioning to further reduce its overhead. We tested SDS-Sort extensively on Edison, a Cray XC30 supercomputer. Timing measurements show that SDS-Sort can scale to 130K CPU cores and deliver a sorting throughput of 117TB/min. In tests with real application data from large science projects, SDS-Sort outperforms HykSort, a state-of-art parallel sorting algorithm, by 3.4X.

Analyses of large simulation data often concentrate on regions in space and in time that contain important information. As simulations adopt Adaptive Mesh Refinement (AMR), the data records from a region of interest could be widely scattered on storage devices and accessing interesting regions results in significantly reduced I/O performance. In this work, we study the organization of block-structured AMR data on storage to improve performance of spatio-temporal data accesses. AMR has a complex hierarchical multi-resolution data structure that does not fit easily with the existing approaches that focus on uniform mesh data. To enable efficient AMR read accesses, we develop an in situ data layout optimization framework. Our framework automatically selects from a set of candidate layouts based on a performance model, and reorganizes the data before writing to storage. We evaluate this framework with three AMR datasets and access patterns derived from scientific applications. Our performance model is able to identify the best layout scheme and yields up to a 3X read performance improvement compared to the original layout. Though it is not possible to turn all read accesses into contiguous reads, we are able to achieve 90% of contiguous read throughput with the optimized layouts on average.

Computing k-Nearest Neighbors (KNN) is one of the core kernels used in many machine learning, data mining and scientific computing applications. Although kd-tree based O(log n) algorithms have been proposed for computing KNN, due to its inherent sequentiality, linear algorithms are being used in practice. This limits the applicability of such methods to millions of data points, with limited scalability for Big Data analytics challenges in the scientific domain. In this paper, we present parallel and highly optimized kd*tree based KNN algorithms (both construction and querying) suitable for distributed architectures. Our algorithm includes novel approaches for pruning search space and improving load balancing and partitioning among nodes and threads. Using TB-sized datasets from three science applications: astrophysics, plasma physics, and particle physics, we show that our implementation can construct kd-tree of 189 billion particles in 48 seconds on utilizing ~50,000 cores. We also demonstrate computation of KNN of 19 billion queries in 12 seconds. We demonstrate almost linear speedup both for shared and distributed memory computers. Our algorithms outperforms earlier implementations by more than order of magnitude, thereby radically improving the applicability of our implementation to state-of-the-art Big Data analytics problems.

In the era of data-intensive scientific discovery, data analysis is critical for scientists to identify essential information from the mountains of data generated by large-scale simulations or experiments. A generic operation in scientific data analysis is to combine information from multiple data sets, which are stored in heterogeneous ile formats. This operation is typically known as a Join in database management field. Currently, a join operation involving multiple data sets in different file formats is time-consuming because of the need to prepare data (i.e., to convert data into a uniform format or to ingest into a database) and to run the join algorithms. Furthermore, data processing languages, such as SQL (Structured Query Language), can not easily express typical scientific analysis tasks such as interpolation. In this paper, we propose three techniques to address these challenges: a two-level data model to process data from different file formats without converting to a uniform format, a data organization structure known as Multi-Dimensional Binning (MDBin), and a join processing algorithm known as Spatially Clustered Join (SCJoin). Together, these techniques allow scientific data files to be used for query processing with less I/O cost and fast query response time without the extra cost to perform ile format conversion and data ingestion. Evaluation of our proposed techniques in joining and interpolating data sets generated by a plasma physics simulation studying space weather phenomenon showed up to 8X improvement over FastQuery. Querying with our solution outperforms SciDB, a popular array data management system for scientific data, by 43X-143X. We also demonstrate that our methods scale to 64K CPU cores in analyzing 32TB data on a large-scale supercomputing system.

Scientific data is often shared among researchers and even reorganized by colleagues or third-party users. Thus, it is essential to provide an adequate degree of access control for such shared data to preserve a desired level of security requirements. In this work, we develop an essential, lightweight access control model for secure data services in a limited distributed setting such as an HPC cluster. In particular, we consider SDS (the Scientific Data Services framework) as a use case system, which is a framework offering performance-optimized data access, reorganization, and analysis. We outline the requirements and challenges for access control for effective data services, and develop an authorization service model based on the defined requirements. We also present an initial prototyping model.

Adaptive Mesh Refinement (AMR) represents a significant advance for scientific simulation codes, greatly reducing memory and compute requirements by dynamically varying simulation resolution over space and time. As simulation codes transition to AMR, existing analysis algorithms must also make this transition. One such algorithm, connected component detection, is of vital importance in many simulation and analysis contexts, with some simulation codes even relying on parallel, in situ connected component detection for correctness. Yet, current detection algorithms designed for uniform meshes are not applicable to hierarchical, non-uniform AMR, and to the best of our knowledge, AMR connected component detection has not been explored in the literature. Therefore, in this paper, we formally define the general problem of connected component detection for AMR, and present a general solution. Beyond solving the general detection problem, achieving viable in situ detection performance is even more challenging. The core issue is the conflict between the communication-intensive nature of connected component detection (in general, and especially for AMR data) and the requirement that in situ processes incur minimal performance impact on the co-located simulation. We address this challenge by presenting the first connected component detection methodology for structured AMR that is applicable in a parallel, in situ context. Our key strategy is the incorporation of an multi-phase AMR-aware communication pattern that synchronizes connectivity information across the AMR hierarchy. In addition, we distil our methodology to a generic framework within the Combo AMR infrastructure, making connected component detection services available for many existing applications. We demonstrate our method's efficacy by showing its ability to detect ice calving events in real time within the real-world BISICLES ice sheet modelling code. Results show up to a 6.8x speedup of our algorithm over the existing specialized BISICLES algorithm. We also show scalability results for our method up to 4,096 cores using a parallel Combo-based benchmark.

Parallel I/O performance depends highly on the interactions among multiple layers of the parallel I/O stack. The most common layers include high-level I/O libraries, MPI-IO middleware, and parallel file system. Each of these layers offers various tunable parameters to control intermediary data transfer points and the final data layout. Due to the interdependencies and the number of combinations of parameters, finding a good set of parameter values for a specific application's I/O pattern is challenging. Recent efforts, such as autotuning with genetic algorithms (GAs) and analytical models, have several limitations. For instance, analytical models fail to capture the dynamic nature of shared supercomputing systems and are application-specific. GA-based tuning requires running many time-consuming experiments for each input size. In this paper, we present a strategy to generate automatically an empirical model for a given application pattern. Using a set of real measurements from running an I/O kernel as training set, we generate a nonlinear regression model. We use this model to predict the top-20 tunable parameter values that give efficient I/O performance and rerun the I/O kernel to select the best set of parameter under the current conditions as tunable parameters for future runs of the same I/O kernel. Using this approach, we demonstrate 6X - 94X speedup over default I/O time for different I/O kernels running on multiple HPC systems. We also evaluate performance by identifying interdependencies among different sets of tunable parameters.

We examine the I/O behavior of thousands of supercomputing applications "in the wild," by analyzing the Darshan logs of over a million jobs representing a combined total of six years of I/O behavior across three leading high-performance computing platforms. We mined these logs to analyze the I/O behavior of applications across all their runs on a platform; the evolution of an application's I/O behavior across time, and across platforms; and the I/O behavior of a platform's entire workload. Our analysis techniques can help developers and platform owners improve I/O performance and I/O system utilization, by quickly identifying underperforming applications and offering early intervention to save system resources. We summarize our observations regarding how jobs perform I/O and the throughput they attain in practice.

Modern cosmology and plasma physics codes are now capable of simulating trillions of particles on petascale systems. Each timestep output from such simulations is on the order of 10s of TBs. Summarizing and analyzing raw particle data is challenging, and scientists often focus on density structures, whether in the real 3D space, or a high-dimensional phase space. In this work, we develop a highly scalable version of the clustering algorithm Dbscan, and apply it to the largest datasets produced by state-of-the-art codes. Our system, called Bd-Cats, is the first one capable of performing end-to-end analysis at trillion particle scale (including: loading the data, geometric partitioning, computing kd-trees, performing clustering analysis, and storing the results). We show analysis of 1.4 trillion particles from a plasma physics simulation, and a 10,2403 particle cosmological simulation, utilizing ~100,000 cores in 30 minutes. Bd-Cats is helping infer mechanisms behind particle acceleration in plasma physics and holds promise for qualitatively superior clustering in cosmology. Both of these results were previously intractable at the trillion particle scale.

Scientific experiments and simulations produce mountains of data in file formats, such as HDF5, NetCDF, and FITS. Often, a relatively small amount of data holds the key to new scientific insight. Locating that critical information in these large files is challenging because existing solutions need significant user involvement in preparing the data, generating indexes, and answering queries. Data management systems that support querying, such as SciDB, require a costly process of loading data from scientific data formats to these systems. The search results also need to be converted back to a format needed by the subsequent data analysis and visualization tools. These steps are time-consuming, tedious, and possibly error-prone. Toward providing efficient data management directly on these scientific file formats, we introduce a framework called Scientific Data Services (SDS). SDS targets to provide efficient data management optimizations as services. In this paper, we introduce the design and implementation of one such service, the parallel querying service. To answer the queries efficiently, we transparently augment user data with bitmap indexes and ordered datasets. We design the querying service to manage these augmented datasets and to redirect queries automatically to bitmap indexes or to ordered datasets based on their availability and the expected query response time. The generation of bitmap indexes and sorted datasets and querying are parallelized to work on large supercomputers. We show that SDS achieves 22X, 55X, and 62X speedups compared to conventional full-scan approach of sifting through data in answering three queries from a plasma physics analysis application.

Various layers of the parallel I/O subsystem offer tunable parameters for improving I/O performance on large-scale computers. However, searching through a large parameter space is challenging. We are working towards an autotuning framework for determining the parallel I/O parameters that can achieve good I/O performance for different data write patterns. In this paper, we characterize parallel I/O and discuss the development of predictive models for use in effectively reducing the parameter space. Applying our technique on tuning an I/O kernel derived from a large-scale simulation code shows that the search time can be reduced from 12 hours to 2 hours, while achieving 54X I/O performance speedup.

Complex indexing techniques are needed to reduce the time of analyzing massive scientific datasets, but generating these indexing data structures can be very time consuming. In this work, we propose a set of strategies to simplify the index file structure and to improve the I/O performance during index construction using FastQuery, which is a parallel indexing and querying system for scientific data. FastQuery has been used to analyze data from various scientific applications, including a trillion plasma particles simulation. To accelerate query process, FastQuery uses FastBit to build indexes, and then stores the indexes into file system through parallel scientific data format libraries, such as HDF5. Although these data format libraries are designed to support more complex multi-dimensional arrays, we observed that it still takes considerable work to map the indexing data structures into arrays, especially on parallel machines. To address this problem, in this paper, we attempt to minimize the I/O time by storing indexes into our self-defined binary data format. By fully controlling the data structure, we can minimize the I/O synchronization overhead and explore more efficient I/O strategy for storing indexes. Our experiments of indexing a trillion particle dataset using 20,000 cores of a supercomputer show that the proposed binary I/O driver can reach 85% of the peak I/O bandwidth on the system, and achieves a speedup of up to 4X in terms of the total execution time comparing to the previous FastQuery implementation with HDF5 I/O driver.

Scientific experiments and large-scale simulations produce massive amounts of data. Many of these scientific datasets are arrays, and are stored in file formats such as HDF5 and NetCDF. Although scientific data management systems, such as SciDB, are designed to manipulate arrays, there are challenges in integrating these systems into existing analysis workflows. Major barriers include the expensive task of preparing and loading data before querying, and converting the final results to a format that is understood by the existing post-processing and visualization tools. As a consequence, integrating a data management system into an existing scientific data analysis workflow is time-consuming and requires extensive user involvement. In this paper, we present the design of a new scientific data analysis system that efficiently processes queries directly over data stored in the HDF5 file format. This design choice eliminates the tedious and error-prone data loading process, and makes the query results readily available to the next processing steps of the analysis workflow. Our design leverages the increasing main memory capacities found in supercomputers through bitmap indexing and in-memory query execution. In addition, query processing over the HDF5 data format can be effortlessly parallelized to utilize the ample concurrency available in large-scale supercomputers and modern parallel file systems. We evaluate the performance of our system on a large supercomputing system and experiment with both a synthetic dataset and a real cosmology observation dataset. Our system frequently outperforms the relational database system that the cosmology team currently uses, and is more than 10X faster than Hive when processing data in parallel. Overall, by eliminating the data loading step, our query processing system is more effective in supporting in situ scientific analysis workflows.

Many scientific applications nowadays generate a few terabytes (TB) of data in a single run and the data sizes are expected to reach petabytes (PB) in the near future. Enabling fast extraction of knowledge through analyzing these large datasets holds the key to faster scientific discoveries. However, reading data from traditional storage subsystem is a slow process as the I/O performance lags far behind computational performance. Reducing data movement from the storage subsystem is widely considered a viable option for improving performance of data analysis. In this paper, we propose Segmented Analysis, a data movement reduction strategy through reusing results, where multiple similar analysis tasks process the same segments of data. The basic idea is to segment the data accessed in an analysis task, to process the data segments with a given analysis task, and to store the results of segments in a cache for future use. In future, when an analysis task needs to perform the same process on the data segments for which the results are available in the cache, the task can avoid moving data and performing computation for the available results. The Segmented Analysis framework contains modules for computation and I/O access overlap detection, in situ segmentation, and segment result caching. We evaluate the Segmented Analysis strategy by varying factors like the overlap rate among analysis tasks, the request size and the granularity of segmentation. We observed 2X to 13X I/O and to 2X to 8X computation speedups when the overlap is above 50%.

Data producers typically optimize the layout of data files to minimize the write time. In most cases, data analysis tasks read these files in access patterns different from the write patterns causing poor read performance. In this paper, we introduce Scientific Data Services (SDS), a framework for bridging the performance gap between writing and reading scientific data. SDS reorganizes data to match the read patterns of analysis tasks and enables transparent data reads from the reorganized data. We implemented a HDF5 Virtual Object Layer (VOL) plugin to redirect the HDF5 dataset read calls to the reorganized data. To demonstrate the effectiveness of SDS, we applied two parallel data organization techniques: a sort-based organization on a plasma physics data and a transpose-based organization on mass spectrometry imaging data. We also extended the HDF5 data access API to allow selection of data based on their values through a query interface, called SDS Query. We evaluated the execution time in accessing various subsets of data through existing HDF5 Read API and SDS Query.We showed that reading the reorganized data using SDS is up to 55X faster than reading the original data.

We present an auto-tuning system for optimizing I/O performance of HDF5 applications and demonstrate its value across platforms, applications, and at scale. The system uses a genetic algorithm to search a large space of tunable parameters and to identify effective settings at all layers of the parallel I/O stack. The parameter settings are applied transparently by the auto-tuning system via dynamically intercepted HDF5 calls.

FastQuery is a parallel indexing and querying system we developed for accelerating analysis and visualization of scientific data. We have applied it to a wide variety of HPC applications and demonstrated its capability and scalability using a petascale trillion-particle simulation in our previous work. Yet, through our experience, we found that performance of reading and writing data with FastQuery, like many other HPC applications, could be significantly affected by various tunable parameters throughout the parallel I/O stack. In this paper, we describe our success in tuning the performance of FastQuery on a Lustre parallel file system. We study and analyze the impact of parameters and tunable settings at file system, MPI-IO library, and HDF5 library levels of the I/O stack. We demonstrate that a combined optimization strategy is able to improve performance and I/O bandwidth of FastQuery significantly. In our tests with a trillion-particle dataset, the time to index the dataset reduced by more than one half.

Many scientific applications spend a significant portion of their execution time in accessing data from files. Various optimization techniques exist to improve data access performance, such as data prefetching and data layout optimization. However, optimization process is usually a difficult task due to the complexity involved in understanding I/O behavior. Tools that can help simplify the optimization process have a significant importance. In this paper, we introduce a tool, called IOSIG, for providing a better understanding of parallel I/O accesses and information to be used for optimization techniques. The tool enables tracing parallel I/O calls of an application and analyzing the collected information to provide a clear understanding of I/O behavior of the application. We show that performance overheads of the tool in trace collection and analysis are negligible. The analysis step creates I/O signatures that various optimizations can use for improving I/O performance. I/O signatures are compact, easy-to-understand, and parameterized representations containing data access pattern information such as size, strides between consecutive accesses, repetition, timing, etc. The signatures include local I/O behavior for each process and global behavior for an overall application. We illustrate the usage of the IOSIG tool in data prefetching and data layout optimizations.

Petascale plasma physics simulations have recently entered the regime of simulating trillions of particles. These unprecedented simulations generate massive amounts of data, posing significant challenges in storage, analysis, and visualization. In this paper, we present parallel I/O, analysis, and visualization results from a VPIC trillion particle simulation running on 120,000 cores, which produces ~30TB of data for a single timestep. We demonstrate the successful application of H5Part, a particle data extension of parallel HDF5, for writing the dataset at a significant fraction of system peak I/O rates. To enable efficient analysis, we develop hybrid parallel FastQuery to index and query data using multi-core CPUs on distributed memory hardware. We show good scalability results for the FastQuery implementation using up to 10,000 cores. Finally, we apply this indexing/query-driven approach to facilitate the first-ever analysis and visualization of the trillion particle dataset.

It is widely believed that most Recognition and Mining (RM) workloads can easily take advantage of parallel computing platforms because these workloads are dataparallel. Contrary to this popular belief, we present RM workloads for which conventional parallel implementations scale poorly on multi-core platforms. We identify off-chip memory transfers and overheads in the parallel runtime library as the primary bottlenecks that limit speedups to be well below the ideal linear speedup expected for data-parallel workloads. To achieve improved parallel scalability, we identify and exploit several interesting properties of RM workloads - sparsity of model updates, low spatial locality among model updates, presence of insignificant computations, and the inherently self-healing nature of these algorithms in the presence of errors. We leverage these domain-specific characteristics to improve parallel scalability in two major ways. First, we utilize data dependency relaxation to simultaneously execute multiple training iterations in parallel, thereby increasing the granularity of the parallel tasks and significantly lowering the run-time overheads of fine-grained threading. Second, we strategically drop selected computations that are insignificant to the accuracy of the final result, but account for a disproportionately large amount of off-chip (memory and coherence) traffic. Through the application of the proposed techniques, we show that much higher speedups are possible on multi-core platforms for two important RM applications - document search using semantic indexing, and eye detection in images using generalized learning vector quantization. On an 8-core platform, we achieve application speedups of 5.5X and 7.3X compared to sequential implementations. Compared to conventional parallel implementations of these applications using Intel's TBB, the proposed techniques result in 4.3X and 4.9X improvements. Although the optimized parallel implementations are not numerically equivalent to the sequential implementations, the output quality is shown to be comparable (and within the margin of variation produced by processing the input data in a different order). We also explore error mitigation techniques that can be used to ensure that the accuracy of results is not compromised.

In this paper, we describe a runtime to automatically enhance the performance of applications running on heterogeneous platforms consisting of a multi-core (CPU) and a throughput-oriented many-core (GPU). The CPU and GPU are connected by a non-coherent interconnect such as PCI-E, and as such do not have shared memory. Heterogeneous platforms available today such as [9] are of this type. Our goal is to enable the programmer to seamlessly use such a system without rewriting the application and with minimal knowledge of the underlying architectural details. Assuming that applications perform function calls to computational kernels with available CPU and GPU implementations, our runtime achieves this goal by automatically scheduling the kernels and managing data placement. In particular, it intercepts function calls to well-known computational kernels and schedules them on CPU or GPU based on their argument size and location. To improve performance, it defers all data transfers between the CPU and the GPU until necessary. By managing data placement transparently to the programmer, it provides a unified memory view despite the underlying separate memory sub-systems.

Multicore processors are now part of mainstream computing. However, data access contention among multiple cores is a significant performance bottleneck in utilizing these processors. Typically, memory hierarchies in multicore architectures use shared last level cache or shared memory. As multiple cores concurrently send requests to access data from these shared memory hierarchy levels, their capacity to serve all the requests is overwhelming and causes performance bottlenecks. In this paper, we introduce simple analytical models for predicting the occurrence of data access contention and provide a guideline for choosing optimal number of cores in running an application without causing data access contention. We verify our models by comparing the predicted optimal number of cores without causing data contention with the measured value in running MIMD Lattice Computation (MILC) application. The proposed analytical models are accurate and promising in guiding data access optimizations to improve multicore utilization.

Multi-core processors have changed the conventional hardware structure and require a rethinking of system scheduling and resource management to utilize them efficiently. However, current multi-core systems are still using conventional single-core memory scheduling. In this study, we investigate and evaluate traditional memory access scheduling techniques, and propose a core-aware memory scheduling for multi-core environments. Since memory requests from the same source exhibit better locality, it is reasonable to schedule the requests by taking the source of the requests into consideration. Motivated from this principle of locality, we propose two core-aware policies based on traditional bank-first and row-first schemes. Simulation results show that the core-aware policies can effectively improve the performance. Compared with the bank-first and row-first policies, the proposed core-aware policies reduce the execution time of certain NAS Parallel Benchmarks by up to 20% in running the benchmarks separately, and by 11% in running them concurrently.

Parallel applications can benefit greatly from massive computational capability, but their performance usually suffers due to large latency in I/O accesses. Conventional I/O prefetching techniques are conservative and are limited by low accuracy and coverage. As the processor performance has been increasing rapidly and the computing power is virtually free, we introduce a novel speculative approach for comprehensive and aggressive parallel I/O prefetching in this study. We present the design of our approach as well as challenges, solutions, and our prototype implementation. The experiments have shown promising results in reducing I/O access latency.

Parallel I/O prefetching is considered to be effective in improving I/O performance. However, the effectiveness depends on determining patterns among future I/O accesses swiftly and fetching data in time, which is difficult to achieve in general. In this study, we propose an I/O signature-based prefetching strategy. The idea is to use a predetermined I/O signature of an application to guide prefetching. To put this idea to work, we first derived a classification of patterns and introduced a simple and effective signature notation to represent patterns. We then developed a toolkit to trace and generate I/O signatures automatically. Finally, we designed and implemented a thread-based client-side collective prefetching cache layer for MPI-IO library to support prefetching. A prefetching thread reads I/O signatures of an application and adjusts them by observing I/O accesses at runtime. Experimental results show that the proposed prefetching method improves I/O performance significantly for applications with complex patterns.

Parallel applications are usually able to achieve high computational performance but suffer from large latency in I/O accesses. I/O prefetching is an effective solution for masking the latency. Most of existing I/O prefetching techniques, however, are conservative and their effectiveness is limited by low accuracy and coverage. As the processor-I/O performance gap has been increasing rapidly, data-access delay has become a dominant performance bottleneck. We argue that it is time to revisit the ldquoI/O wallrdquo problem and trade the excessive computing power with data-access speed. We propose a novel pre-execution approach for masking I/O latency. We describe the pre-execution I/O prefetching framework, the pre-execution thread construction methodology, the underlying library support, and the prototype implementation in the ROMIO MPI-IO implementation in MPICH2. Preliminary experiments show that the pre-execution approach is promising in reducing I/O access latency and has real potential.

Data prefetching, where data is fetched before CPU demands for it, has been considered as an effective solution to mask data access latency. However, the current client-initiated prefetching strategies do not work well for applications with complex, non-contiguous data access patterns. While technology advances continue to enlarge the gap between computing and data access performance, trading computing power for data access delay has become a natural choice. We propose a server-based data-push approach. In this server-push architecture, a dedicated server named data push server (DPS) initiates and proactively pushes data closer to the client in time. We present the DPS architecture and study the issues such as what data to fetch, when to fetch, how to push, and data access modeling.

Data prefetching is an effective way to bridge the increasing performance gap between processor and memory. As computing power is increasing much faster than memory performance, we suggest that it is time to have a dedicated cache to store data access histories and to serve prefetching to mask data access latency effectively. We thus propose a new cache structure, named Data Access History Cache (DAHC), and study its associated prefetching mechanisms. The DAHC behaves as a cache for recent reference information instead of as a traditional cache for instructions or data. Theoretically, it is capable of supporting many well known history-based prefetching algorithms, especially adaptive and aggressive approaches. We have carried out simulation experiments to validate DAHC design and DAHC-based data prefetching methodologies and to demonstrate performance gains. The DAHC provides a practical approach to reaping data prefetching benefits and its associated prefetching mechanisms are proven more effective than traditional approaches.

Improving memory performance at software level is more effective in reducing the rapidly expanding gap between processor and memory performance. Loop transformations (e.g. loop unrolling, loop tiling) and array restructuring optimizations improve the memory performance by increasing the locality of memory accesses. To find the best optimization parameters at runtime, we need a fast and simple analytical model to predict the memory access cost. Most of the existing models are complex and impractical to be integrated in the runtime tuning systems. In this paper, we propose a simple, fast and reasonably accurate model that is capable of predicting the memory access cost based on a wide range of data access patterns that appear in many scientific applications.

The MPI Standard supports derived datatypes, which allow users to describe noncontiguous memory layout and communicate noncontiguous data with a single communication function. This feature enables an MPI implementation to optimize the transfer of noncontiguous data. In practice, however, few MPI implementations implement derived datatypes in a way that performs better than what the user can achieve by manually packing data into a contiguous buffer and then calling an MPI function. In this paper, we present a technique for improving the performance of derived datatypes by automatically using packing algorithms that are optimized for memory-access cost. The packing algorithms use memory-optimization techniques that the user cannot apply easily without advanced knowledge of the memory architecture. We present performance results for a matrix-transpose example that demonstrate that our implementation of derived datatypes significantly outperforms both manual packing by the user and the existing derived-datatype code in the MPI implementation (MPICH).