Scientific workflows increasingly need to train a DNN model in real-time during an experiment (e.g. using ground truth from a simulation), while using it at the same time for inferences. Instead of sharing the same model instance, the training (producer) and inference server (consumer) often use different model replicas that are kept synchronized. In addition to efficient I/O techniques to keep the model replica of the producer and consumer synchronized, there is another important trade-off: frequent model updates enhance inference quality but may slow down training; infrequent updates may lead to less precise inference results. To address these challenges, we introduce Viper: a new I/O framework designed to determine a near-optimal checkpoint schedule and accelerate the delivery of the latest model updates. Viper builds an inference performance predictor to identify the optimal checkpoint schedule to balance the trade-off between training slowdown and inference quality improvement. It also creates a memory-first model transfer engine to accelerate model delivery through direct memory-to-memory communication. Our experiments show that Viper can reduce the model update latency by ≈ 9x using the GPU-to-GPU data transfer engine and ≈ 3x using the DRAM-to-DRAM host data transfer. The checkpoint schedule obtained from Viper’s predictor also demonstrates improved cumulative inference accuracy compared to the baseline of epoch-based solutions.

We describe LowFive, a new data transport layer based on the HDF5 data model, for in situ workflows. Executables using LowFive can communicate in situ (using in-memory data and MPI message passing), reading and writing traditional HDF5 files to physical storage, and combining the two modes. Minimal and often no source-code modification is needed for programs that already use HDF5. LowFive maintains deep copies or shallow references of datasets, configurable by the user. More than one task can produce (write) data, and more than one task can consume (read) data, accommodating fan-in and fan-out in the workflow task graph. LowFive supports data redistribution from n producer processes to m consumer processes. We demonstrate the above features in a series of experiments featuring both synthetic benchmarks as well as a representative use case from a scientific workflow, and we also compare with other data transport solutions in the literature.

Scientific workflows drive most modern large-scale science breakthroughs by allowing scientists to define their computations as a set of jobs executed in a given order based on their data dependencies. Workflow management systems (WMSs) have become key to automating scientific workflows-executing computational jobs and orchestrating data transfers between those jobs running on complex high-performance computing (HPC) platforms. Traditionally, WMSs use files to communicate between jobs: a job writes out files that are read by other jobs. However, HPC machines face a growing gap between their storage and compute capabilities. To address that concern, the scientific community has adopted a new approach called in situ, which bypasses costly parallel filesystem I/O operations with faster in-memory or in-network communications. When using in situ approaches, communication and computations can be interleaved. In this work, we leverage the Decaf in situ dataflow framework to accelerate task-based scientific workflows managed by the Pegasus WMS, by replacing file communications with faster MPI messaging. We propose a new execution engine that uses Decaf to manage communications within a sub-workflow (i.e., set of jobs) to optimize inter-job communications. We consider two workflows in this study: (i) a synthetic workflow that benchmarks and compares file- and MPI-based communication; and (ii) a realistic bioinformatics workflow that computes mu-tational overlaps in the human genome. Experiments show that in situ communication can improve the bioinformatics workflow execution time by 22% to 30% compared with file communication. Our results motivate further opportunities and challenges for bridging traditional WMSs with in situ frameworks.

Data parallel techniques have been widely adopted both in academia and industry as a tool to enable scalable training of deep learning models. At scale, DL training jobs can fail due to software or hardware bugs, may need to be preempted or terminated due to unexpected events, or may perform suboptimally because they were misconfigured. Under such circumstances, there is a need to recover and/or reconfigure data-parallel DL training jobs on-the-fly, while minimizing the impact on the accuracy of the DNN model and the runtime overhead. In this regard, state-of-art techniques adopted by the HPC community mostly rely on checkpoint-restart, which inevitably leads to loss of progress, thus increasing the runtime overhead. In this paper we explore alternative techniques that exploit the properties of modern deep learning frameworks (overlapping of gradient averaging and weight updates with local gradient computations through pipeline parallelism) to reduce the overhead of resilience/elasticity. To this end we introduce a failure simulation framework and two resilience strategies (immediate mini-batch rollback and lossy forward recovery), which we study compared with checkpoint-restart approaches in a variety of settings in order to understand the trade-offs between the accuracy loss of the DNN model and the runtime overhead.

Scientific computation increasingly consists of a workflow of interrelated tasks. Containerization can make workflow systems more manageable, reproducible, and portable, but containers can impede communication due to their focus on encapsulation. In some circumstances, shared-memory regions are an effective way to improve performance of workflows; however sharing memory between containerized workflow tasks is difficult. In this work, we have created a software library called Dhmem that manages shared memory between workflow tasks in separate containers, with minimal code change and performance overhead. Instead of all code being in the same container, Dhmem allows a separate container for each workflow task to be constructed completely independently. Dhmem enables additional functionality: easy integration in existing workflow systems, communication configuration at runtime based on the environment, and scalable performance.

Today's scientific applications are increasingly relying on a variety of data sources, storage facilities, and computing infrastructures, and there is a growing demand for data analysis and visualization for these applications. In this context, exploiting Big Data frameworks for scientific computing is an opportunity to incorporate high-level libraries, platforms, and algorithms for machine learning, graph processing, and streaming; inherit their data awareness and fault-tolerance; and increase productivity. Nevertheless, limitations exist when Big Data platforms are integrated with an HPC environment, namely poor scalability, severe memory overhead, and huge development effort. This paper focuses on a popular Big Data framework -Apache Spark- and proposes an architecture to support the integration of highly scalable MPI block-based data models and communication patterns with a map-reduce-based programming model. The resulting platform preserves the data abstraction and programming interface of Spark, without conducting any changes in the framework, but allows the user to delegate operations to the MPI layer. The evaluation of our prototype shows that our approach integrates Spark and MPI efficiently at scale, so end users can take advantage of the productivity facilitated by the rich ecosystem of high-level Big Data tools and libraries based on Spark, without compromising efficiency and scalability.

Burst Buffer is an effective solution for reducing the data transfer time and the I/O interference in HPC systems. Extending Burst Buffers (BBs) to handle Big Data applications is challenging because BBs must account for the large data inputs of Big Data applications and the performance guarantees of HPC applications - which are considered as first-class citizens in HPC systems. Existing BBs focus on only intermediate data of Big Data applications and incur a high performance degradation of both Big Data and HPC applications. We present Eley, a burst buffer solution that helps to accelerate the performance of Big Data applications while guaranteeing the performance of HPC applications. In order to improve the performance of Big Data applications, Eley employs a prefetching technique that fetches the input data of these applications to be stored close to computing nodes thus reducing the latency of reading data inputs. Moreover, Eley is equipped with a full delay operator to guarantee the performance of HPC applications - as they are running independently on a HPC system. The experimental results show the effectiveness of Eley in obtaining shorter execution time of Big Data applications (shorter map phase) while guaranteeing the performance of HPC applications.

As we move toward the exascale era, performance variability in HPC systems remains a challenge. I/O interference, a major cause of this variability, is becoming more important every day with the growing number of concurrent applications that share larger machines. Earlier research efforts on mitigating I/O interference focus on a single potential cause of interference (e.g., the network). Yet the root causes of I/O interference can be diverse. In this work, we conduct an extensive experimental campaign to explore the various root causes of I/O interference in HPC storage systems. We use microbenchmarks on the Grid'5000 testbed to evaluate how the applications' access pattern, the network components, the file system's configuration, and the backend storage devices influence I/O interference. Our studies reveal that in many situations interference is a result of bad flow control in the I/O path, rather than being caused by some single bottleneck in one of its components. We further show that interference-free behavior is not necessarily a sign of optimal performance. To the best of our knowledge, our work provides the first deep insight into the role of each of the potential root causes of interference and their interplay. Our findings can help developers and platform owners improve I/O performance and motivate further research addressing the problem across all components of the I/O stack.

Hadoop emerged as the de facto state-of-the-art system for MapReduce-based data analytics. The reliability of Hadoop systems depends in part on how well they handle failures. Currently, Hadoop handles machine failures by re-executing all the tasks of the failed machines (i.e., executing recovery tasks). Unfortunately, this elegant solution is entirely entrusted to the core of Hadoop and hidden from Hadoop schedulers. The unawareness of failures therefore may prevent Hadoop schedulers from operating correctly towards meeting their objectives (e.g., fairness, job priority) and can significantly impact the performance of MapReduce applications. This paper presents Chronos, a failure-aware scheduling strategy that enables an early yet smart action for fast failure recovery while still operating within a specific scheduler objective. Upon failure detection, rather than waiting an uncertain amount of time to get resources for recovery tasks, Chronos leverages a lightweight preemption technique to carefully allocate these resources. In addition, Chronos considers data locality when scheduling recovery tasks to further improve the performance. We demonstrate the utility of Chronos by combining it with Fifo and Fair schedulers. The experimental results show that Chronos recovers to a correct scheduling behavior within a couple of seconds only and reduces the job completion times by up to 55% compared to state-of-the-art schedulers.