Nowadays there is a raising interest in bridging the gap between Big Data application models and data-intensive HPC. This work explores the effects that Big Data-inspired paradigms could have in current scientific applications through the evaluation of a real-world application from the hydrology domain. This evaluation led to experience that portrayed the key aspects of the HPC and Big Data paradigms that made them successful in their respective worlds. With this information, we established a research roadmap to build a platform suitable for HPC hybrid applications, with a focus on efficient data management and fault-tolerance.

This paper explores novel techniques for improving the performance of many-task workflows based on the Swift scripting language. We propose novel programmer options for automated distributed data placement and task scheduling. These options trigger a data placement mechanism used for distributing intermediate workflow data over the servers of Hercules, a distributed key-value store that can be used to cache file system data. We demonstrate that these new mechanisms can significantly improve the aggregated throughput of many-task workflows with up to 86x, reduce the contention on the shared file system, exploit the data locality, and trade off locality and load balance.

On current large-scale HPC platforms the data path from compute nodes to final storage passes through several networks interconnecting a distributed hierarchy of nodes serving as compute nodes, I/O nodes, and file system servers. Although applications compete for resources at various system levels, the current system software offers no mechanisms for globally coordinating the data flow for attaining optimal resource usage and for reacting to overload or interference. In this paper we describe CLARISSE, a middleware designed to enhance data-staging coordination and control in the HPC software storage I/O stack. CLARISSE exposes the parallel data flows to a higher-level hierarchy of controllers, thereby opening up the possibility of developing novel cross-layer optimizations, based on the run-time information. To the best of our knowledge, CLARISSE is the first middleware that decouples the policy, control, and data layers of the software I/O stack in order to simplify the task of globally coordinating the data staging on large-scale HPC platforms. To demonstrate how CLARISSE can be used for performance enhancement, we present two case studies: an elastic load-aware collective I/O and a cross-application parallel I/O scheduling policy. The evaluation illustrates how coordination can bring a significant performance benefit with low overheads by adapting to load conditions and interference.

The optimization of parallel I/O has become challenging because of the increasing storage hierarchy, performance variability of shared storage systems, and the number of factors in the hardware and software stacks that impact performance. In this paper, we perform an in-depth study of the complexity involved in I/O autotuning and performance modeling, including the architecture, software stack, and noise. We propose a novel hybrid model combining analytical models for communication and storage operations and black-box models for the performance of the individual operations. The experimental results show that the hybrid approach performs significantly better and shows a higher robustness to noise than state-of-the-art machine learning approaches, at the cost of a higher modeling complexity.

Increasingly, large-scale computing systems are consuming more power each passing year. As power consumption is on the rise, concern has been raised over the growing implications on power bills, carbon emissions and power supply limitations for data centers. Computing systems use hardware components that are not power proportional and machines comprising large systems tend to be underutilized, resulting in great wastage of energy. Motivated by this fact, researchers aim to improve the energy efficiency of these systems by increasing resource utilization and exploiting system characteristics in order to achieve power proportionality. However, many challenges burden the task of producing a power-aware, energy-efficient large-scale system that provides the same performance as today's systems. Particularly, storage systems are of great concern since storage consumes a large amount of power in the data center. This paper outlines the main problems and challenges of delivering power-efficient storage solutions, and proposes a taxonomy of power-aware techniques. This work aims to provide a detailed exposition of current trends in power-aware storage systems in a comprehensive and organized manner, outlining and comparing current solutions and challenges.

Applications running in today's data centers show high workload variability. While seasonal patterns, trends and expected events may help building proactive resource allocation policies, this approach has to be complemented with adaptive strategies which should address unexpected events such as flash crowds and volume spikes. Additionally, the limitations of current I/O infrastructures in the face of dramatic increase of data generation require, the ability to build novel abstractions and models for robust decision making regarding data layout and data locality. In this work, we present CONDESA (CONtrolling Data distribution on Elastic Server Architectures), a framework for exploring adaptive data distribution strategies for elastic server architectures. To the best of our knowledge CONDESA is the first platform that permits to systematically study the interplay between five data related strategies: workload prediction, adaptive control of data distribution and server provisioning, adaptive data grouping, adaptive data placement, and adaptive system sizing. We demonstrate how CONDESA can be used for browsing the design space of adaptive data distribution policies. We show how prediction models can be compared in terms of overhead and accuracy. We evaluate the impact of change detection on prediction accuracy and how CONDESA can be used for choosing an adequate prediction horizon. We demonstrate how adaptive prediction can be used for sizing a server system. Finally, we show how prediction models, change detection strategies, and data placement policies can be combined and compared based on server utilization, load balance, data locality, over- and underprovisioning.

Most small-animal X-ray computed tomography (CT) scanners are based on cone-beam geometry with a flat-panel detector orbiting in a circular trajectory. Image reconstruction in these systems is usually performed by approximate methods based on the algorithm proposed by Feldkamp, Davis and Kress (FDK). Currently there is a strong need to speedup the reconstruction of X-Ray CT data in order to extend its clinical applications. The evolution of the semiconductor detector panels has resulted in an increase of detector elements density, which produces a higher amount of data to process. This work focuses on future high-resolution studies (density up to 4096 pixeles), in which multiple level of parallelism will be needed in the reconstruction. In addition, this paper addresses the future challenges of processing high-resolution images in many-core and distributed architectures. In our evaluation section we demonstrate that our solution is 17% faster than recent related works.

Geolocated social networks, combining traditional social networking features with geolocation information, have grown tremendously over the last few years. Yet, very few works have looked at implementing geolocated social networks in a fully distributed manner, a promising avenue to handle the growing scalability challenges of these systems. In this paper, we focus on georecommendation, and show that existing decentralized recommendation mechanisms perform in fact poorly on geodata. We propose a set of novel gossip-based mechanisms to address this problem, in a modular similarity framework called GEOLOGY. The resulting platform is lightweight, efficient, and scalable, and we demonstrate its advantages in terms of recommendation quality and communication overhead on a real dataset of 15,694 users from Foursquare, a leading geolocated social network.

The cloud has recently surged as a promising paradigm for hosting scalable Web systems serving a large number of users with large workload variations. It makes possible to dynamically add and remove resources to horizontally scalable architectures in order to save costs, while maintaining the quality of service. However, in order to achieve these goals the resource management of a platform must include policies and mechanisms for dynamically resizing the system, redistributing content and redirecting user requests. In this work, we address the problem of reconciling dynamic system sizing and content locality. There are three main contributions of our study. First, we address the problem of determining the system size by employing a hierarchical prediction framework that proactively provisions resources based on statistical models of the incoming workload. Second, we show how to employ the hierarchical prediction framework for designing a dispatching mechanism which can be used with any request distribution policy. Third, we propose two novel prediction-based locality-aware request distribution policies: Oblivious Locality-Aware Request Distribution (OLARD) and Affinity-Based Locality-Aware Request Distribution (ABLARD). We demonstrate the advantages of using our hierarchical prediction framework and how our approach achieves a high content locality, while adapting to unexpected workload changes.

This paper presents a methodology for efficiently solving the file transfer scheduling problem in a distributed environment. Our solution is based on the relaxation of an objective-based time-indexed formulation of a linear programming problem. The main contributions of this paper are the following. First, we introduce a novel approach to the relaxation of the time-indexed formulation of the transfer scheduling problem in multi-server and multi-user environments. Our solution consists of reducing the complexity of the optimization by transforming it into an approximation problem, whose proximity to the optimal solution can be controlled depending on practical and computational needs. Second, we present a distributed deployment of our methodology, which leverages the inherent parallelism of the divide-and-conquer approach in order to speed-up the solving process. Third, we demonstrate that our methodology is able to considerably reduce the schedule length and idle time in a computationally tractable way.

Workload variations on Internet platforms such as YouTube, Flickr, LastFM require novel approaches to dynamic resource provisioning in order to meet QoS requirements, while reducing the Total Cost of Ownership (TCO) of the infrastructures. The economy of scale promise of cloud computing is a great opportunity to approach this problem, by developing elastic large scale server infrastructures. However, a proactive approach to dynamic resource provisioning requires prediction models forecasting future load patterns. On the other hand, unexpected volume and data spikes require reactive provisioning for serving unexpected surges in workloads. When workload can not be predicted, adequate data grouping and placement algorithms may facilitate agile scaling up and down of an infrastructure. In this paper, we analyze a dynamic workload of an on-line music portal and present an elastic Web infrastructure that adapts to workload variations by dynamically scaling up and down servers. The workload is predicted by an autoregressive model capturing trends and seasonal patterns. Further, for enhancing data locality, we propose a predictive data grouping based on the history of content access of a user community. Finally, in order to facilitate agile elasticity, we present a data placement based on workload and access pattern prediction. The experimental results demonstrate that our forecasting model predicts workload with a high precision. Further, the predictive data grouping and placement methods provide high locality, load balance and high utilization of resources, allowing a server infrastructure to scale up and down depending on workload.

Infrastructures serving on-line applications experience dynamic workload variations depending on diverse factors such as popularity, marketing, periodic patterns, fads, trends, events, etc. Some predictable factors such as trends, periodicity or scheduled events allow for proactive resource provisioning in order to meet fluctuations in workloads. However, proactive resource provisioning requires prediction models forecasting future workload patterns. This paper proposes a multi-model prediction approach, in which data are grouped into bins based on content locality, and an autoregressive prediction model is assigned to each locality-preserving bin. The prediction models are shown to be identified and fitted in a computationally efficient way. We demonstrate experimentally that our multi-model approach improves locality over the uni-model approach, while achieving efficient resource provisioning and preserving a high resource utilization and load balance.

Parallel applications currently suffer from a significant imbalance between computational power and available I/O bandwidth. Additionally, the hierarchical organization of current Petascale systems contributes to an increase of the I/O subsystem latency. In these hierarchies, file access involves pipelining data through several networks with incremental latencies and higher probability of congestion. Future Exascale systems are likely to share this trait. This paper presents a scalable parallel I/O software system designed to transparently hide the latency of file system accesses to applications on these platforms. Our solution takes advantage of the hierarchy of networks involved in file accesses, to maximize the degree of overlap between computation, file I/O-related communication, and file system access. We describe and evaluate a two-level hierarchy for Blue Gene systems consisting of client-side and I/O node-side caching. Our file cache management modules coordinate the data staging between application and storage through the Blue Gene networks. The experimental results demonstrate that our architecture achieves significant performance improvements through a high degree of overlap between computation, communication, and file I/O.

This paper presents the design and implementation of a novel file I/O solution for Blue Gene systems. We propose a hierarchical I/O cache architecture based on open source software. Our solution is based on an asynchronous data staging strategy, which hides the latency of file system access from compute nodes. The performance results demonstrate the high scalability and significant performance improvements of our architecture over existing solutions.

This paper presents the design and implementation of a new file system independent collective I/O optimization based on file views: view-based collective I/O. View-based collective I/O has been implemented and evaluated inside ROMIO implementation of MPI-IO standard. The evaluation section shows that view-based I/O outperforms the original two-phase collective I/O from ROMIO in most of the cases for three well-known parallel I/O benchmarks. This is especially due to a smaller cost of scatter/gather operations, a reduction of the metadata overhead, and a smaller number of collective communication and synchronization primitives used in the implementation.

This paper presents the design and implementation of a portable ad-hoc parallel I/O system (AHPIOS). AHPIOS virtualizes on-demand available distributed storage resources and allows the files to be striped over several storage devices. Additionally, the design unifies the configuration of the MPI-IO library and the AHPIOS data servers. By a strong integration of the application, MPI-IO library and file system, a significant performance improvement can be achieved. The experimental section shows that the full MPI-IO integrated AHPIOS implementation of file access operations outperforms the existing MPI-IO implementation by as much as 495% for file writes and 522% for file reads.

This paper presents the optimization and evaluation of parallel I/O for the BIPS3D parallel irregular application, a 3-dimensional simulation of BJT and HBT bipolar devices. The parallel version of BIPS3D employs Metis, a library for partitioning graphs, finite element meshes, or sparse matrices. First, we show how the partitioning information provided by Metis can be used in order to improve the performance of parallel I/O. Second, we propose a novel technique, called Interval Data Grouping (IDG), which exploits the data replication of mesh nodes for optimizing the scheduling of the parallel file operations. Finally, we evaluate the parallel I/O version of BIPS3D for various existing parallel I/O techniques and present an in-depth analysis of the IDG performance.

This paper presents multiple-phase collective I/O, a novel collective I/O technique for distributed memory multiprocessors. Multiple-phase collective I/O is a refinement of two-phase collective I/O technique. The communication phase is structured into several steps, which progressively increase the locality of the data to be written to a file system. Besides the description of multiple-phase collective I/O, our paper addresses two additional objectives. First, the authors target to improve the efficiency of the sulphur transport Eurelian model 2 (STEM-II) application. STEM-II is an air quality model that simulates transport, chemical transformations, emission and deposition processes in a unified framework. Due to the large amount of processed data, I/O becomes a critical factor for the application performance. Multiple-phase collective I/O, considerably enhances the performance of the I/O stage in particular and, consequently, of the whole application in general. Second objective consists of evaluating and comparing the performance of multiple-phase collective I/O with that of other well known parallel I/O techniques

This paper presents the design and implementation of the MPI-IO interface for the Clusterfile parallel file system. The approach offers the opportunity of achieving a high correlation between the file access patterns of parallel applications and the physical file distribution. First, any physical file distribution can be expressed by means of MPI data types. Second, mechanisms such as views and collective I/O operations are portably implemented inside the file system, unifying the I/O scheduling strategies of the MPI-IO library and the file system. The experimental section demonstrates performance benefits of more than one order of magnitude.

Clusters are the most common solutions for high performance computing at the present time. In this kind of systems, an important challenge is the I/O subsystem design. Typically, these environments are not flexible enough and the only way to solve performance bottlenecks is adding new hardware. In this paper, we show how an I/O proxy-based architecture can improve the I/O performance of cluster applications in three ways: adapting to the application requirements, reducing the load on the I/O nodes, and finally, increasing the global performance of the storage system

This paper presents the integration of two collective I/O techniques into the Clusterfile parallel file system: disk-directed I/O and two-phase I/O. We show that global cooperative cache management improves the collective I/O performance. The solution focuses on integrating disk parallelism with other types of parallelism: memory (by buffering and caching on several nodes), network (by parallel I/O scheduling strategies) and processors (by redistributing the I/O related computation over several nodes). The performance results show considerable throughput increases over ROMIO's extended two-phase I/O.

This paper presents view I/O, a non-contiguous parallel I/O technique. We show that the linear file model may be an unsuitable abstraction for non-contiguous I/O optimizations. Additionally, the poor cooperation between a file system and an I/O library like MPI-IO may drastically affect the performance. View I/O has detailed knowledge about parallel structure of a file and about the potential access pattern and exploits it in order to improve performance. The access overhead is reduced by using a strategy "declare once, use several times" and by file off-set compaction. We compare and contrast view I/O with other non-contiguous I/O methods. Our measurements on a cluster of computers indicate a significant performance improvement over other approaches.

Abstract This paper presents Clusterfile, a parallel file system that provides parallel file access on a cluster of computers. We introduce a file partitioning model that has been used in the design of Clusterfile. The model uses a data representation that is optimized for multidimensional array partitioning while allowing arbitrary partitions. The paper shows how the file model can be employed for file partitioning into both physical subfiles and logical views. We also present how the conversion between two partitions of the same file is implemented using a general memory redistribution algorithm. We show how we use the algorithm to optimize non‐contiguous read and write operations. The experimental results include performance comparisons with the Parallel Virtual File System (PVFS) and an MPI‐IO implementation for PVFS. Copyright © 2003 John Wiley & Sons, Ltd.

This paper presents Clusterfile, a parallel file system that provides parallel file access on a cluster of computers. Existing parallel file systems offer little control over matching the I/O access patterns and file data layout. Without this matching the applications may face the following problems: contention at I/O nodes, fragmentation of file data, false sharing, small network messages, high overhead of scattering/gathering the data. Clusterfile addresses some of these inefficiencies. Parallel applications can physically partition a file in arbitrary patterns. They can also set arbitrary views on a file. Views hide the parallel structure of the file and ease the programmer's burden of computing complex access indices. The intersections between views and layouts are computed by a memory redistribution algorithm. Read and write operations are optimized by pre-computing the direct mapping between access patterns and disks. Clusterfile uses the same data representation for file layouts, access patterns, and the mappings between each other.