Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “BURST BUFFER”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 73 records · Page 4

Characterizing Machine Learning I/O Workloads on Leadership Scale HPC Systems

High performance computing (HPC) is no longer solely limited to traditional workloads such as simulation and modeling. With the increase in the popularity of machine learning (ML) and deep learning (DL) technologies, we are observing that an increasing number of HPC users are incorporating ML methods into their workflow and scientific discovery processes, across a wide spectrum of science domains such as biology, earth science, and physics. This gives rise to a diverse set of I/O patterns than the traditional checkpoint/restart-based HPC I/O behavior. The details of the I/O characteristics of such ML I/O workloads have not been studied extensively for large-scale leadership HPC systems. This paper aims to fill that gap by providing an in-depth analysis to gain an understanding of the I/O behavior of ML I/O workloads using darshan - an I/O characterization tool designed for lightweight tracing and profiling. We study the darshan logs of more than 23, 000 HPC ML I/O jobs over a time period of one year running on Summit - the second-fastest supercomputer in the world. This paper provides a systematic I/O characterization of ML I/O jobs running on a leadership scale supercomputer to understand how the I/O behavior differs across science domains and the scale of workloads, and analyze the usage of parallel file system and burst buffer by ML I/O workloads.

Paul, Arnab↗

The Design, Deployment, and Evaluation of the CORAL Pre-Exascale Systems

CORAL, the Collaboration of Oak Ridge, Argonne and Livermore, is fielding two similar IBM systems, Summit and Sierra, with NVIDIA GPUs that will replace the existing Titan and Sequoia systems. Summit and Sierra are currently ranked No. 1 and No. 3, respectively on the Top500 list. We discuss the design and key differences of the systems. Our evaluation of the systems highlights the following. Applications that fit in HBM see the most benefit and may prefer more GPUs; however, for some applications, the CPU-GPU bandwidth is more important than the number of GPUs. The node-local burst buffer scales linearly, and can achieve a 4X improvement over the parallel file system for large jobs; smaller jobs, however, may benefit from writing directly to the PFS. Finally, several CPU, network and memory bound analytics and GPU-bound deep learning codes achieve up to a 11X and 79X speedup/node, respectively over Titan.

Vazhkudai, Sudharshan↗

Enforcing End-to-end I/O Policies for Scientific Workflows using Software-Defined Storage Resource Enclaves

Data-intensive knowledge discovery requires scientific applications to run concurrently with analytics and visualization codes executing in situ for timely output inspection and knowledge extraction. Consequently, I/O pipelines of scientific workflows can be long and complex because they comprise many stages of analytics across different layers of the I/O stack of high-performance computing systems. Performance limitations at any I/O layer or stage can cause an I/O bottleneck resulting in greater than expected end-to-end I/O latency. In this paper, we present the design and implementation of a novel data management infrastructure called Software-Defined Storage Resource Enclaves (SIREN) at system level to enforce end-to-end policies that dictate an I/O pipeline's performance. SIREN provides an I/O performance interface for users to specify the desired storage resources in the context of in-situ analytics. If suboptimal performance of analytics is caused by an I/O bottleneck when data are transferred between simulations and analytics, schedulers in different layers of the I/O stack automatically provide the guaranteed lower bounds on I/O throughput. Lastly, our experimental results demonstrate that SIREN provides performance isolation among scientific workflows sharing multiple storage servers across two I/O layers (burst buffer and parallel file systems) while maintaining high system scalability and resource utilization.

97 MATHEMATICS AND COMPUTING↗

Comparative I/O workload characterization of two leadership class storage clusters. In: PDSW '15 Proceedings of the 10th Parallel Data Storage Workshop

The Oak Ridge Leadership Computing Facility (OLCF) is a leader in large-scale parallel file system development, design, deployment and continuous operation. For the last decade, the OLCF has designed and deployed two large center-wide parallel file systems. The first instantiation, Spider 1, served the Jaguar supercomputer and its predecessor, Spider 2, now serves the Titan supercomputer, among many other OLCF computational resources. The OLCF has been rigorously collecting file and storage system statistics from these Spider systems since their transition to production state. In this paper we present the collected I/O workload statistics from the Spider 2 system and compare it to the Spider 1 data. Our analysis show that the Spider 2 workload is more more write-heavy I/O compared to Spider 1 (75% vs. 60%, respectively). The data also show the OLCF storage policies such as periodic purges are effectively managing the capacity resource of Spider 2. Furthermore, due to improvements in tdm_multipath and ib_srp software, we are utilizing the Spider 2 system bandwidth and latency resources more effectively. The Spider 2 bandwidth usage statistics shows that our system is working within the design specifications. However, it is also evident that our scientific applications can be more effectively served by a burst buffer storage layer. All the data has been collected by monitoring tools developed for the Spider ecosystem. We believe the observed data set and insights will help us better design the next-generation Spider file and storage system. It will also be helpful to the larger community for building more effective large-scale file and storage systems.

Gunasekaran, Raghul↗

Performance analysis of emerging data analytics and HPC workloads

Supercomputers are increasingly being used to run a data analytics workload in addition to a traditional simulation science workload. This mixed workload must be rigorously characterized to ensure that appropriately balanced machines are deployed. In this paper we analyze a suite of applications representing the simulation science and data workload at the NERSC supercomputing center. We show how time is spent in application compute, library compute, communication and I/O, and present application performance on both the Intel Xeon and Intel Xeon-Phi partitions of the Cori supercomputer. We find commonality in the libraries used, I/O motifs and methods of parallelism, and obtain similar node-to-node performance for the base application configurations. We demonstrate that features of the Intel Xeon-Phi node architecture and a Burst Buffer can improve application performance, providing evidence that an exascale-era energy-efficient platform can support a mixed workload.

Daley, Christopher S.↗

HDF5 in the exascale era: Delivering efficient and scalable parallel I/O for exascale applications

Accurately modeling real-world systems requires scientific applications at exascale to generate massive amounts of data and manage data storage efficiently. However, parallel input and output (I/O) faces challenges due to new application workflows and the state-of-the-art memory, interconnect, and storage architectures considered in exascale designs. The storage hierarchy has expanded with node-local persistent memory, solid-state storage, and traditional disk and tape-based storage, thus requiring efficiency at each layer and much more efficient data movement among these layers. This paper discusses how the ExaHDF5 project improved the I/O performance and data management for exascale architectures by enhancing HDF5, a widely used parallel I/O library. The team developed an Asynchronous I/O Virtual Object Layer (VOL) connector that allowed overlapping I/O with computation. They also created a Cache VOL to complement asynchronous I/O by incorporating fast storage layers, such as burst buffer and node-local storage, into the parallel I/O workflow through caching and staging data. Additionally, the team enabled data aggregation and I/O at the node level by using a Subfiling Virtual File Driver (VFD). To demonstrate superior I/O performance with HDF5 at exascale, the ExaHDF5 team collaborated with several exascale applications. In this paper, we show I/O performance improvements for three applications: Cabana (a particle-based simulation library), EQSIM (a regional earthquake simulation software), and E3SM (a climate system modeling library).

Asynchronous I/Ol↗

Checkpointing Strategies for Shared High-Performance Computing Platforms

Input/output (I/O) from various sources often contend for scarcely available bandwidth. For example, checkpoint/restart (CR) protocols can help to ensure application progress in failure-prone environments. However, CR I/O alongside an application's normal, requisite I/O can increase I/O contention and might negatively impact performance. In this work, we consider different aspects (system-level scheduling policies and hardware) that optimize the overall performance of concurrently executing CR-based applications that share I/O resources. We provide a theoretical model and derive a set of necessary constraints to minimize the global waste on a given platform. Our results demonstrate that Young/Daly's optimal checkpoint interval, despite providing a sensible metric for a single, undisturbed application, is not sufficient to optimally address resource contention at scale. We show that by combining optimal checkpointing periods with contention-aware system-level I/O scheduling strategies, we can significantly improve overall application performance and maximize the platform throughput. Finally, we evaluate how specialized hardware, namely burst buffers, may help to mitigate the I/O contention problem. Altogether, these results provide critical analysis and direct guidance on how to design efficient, CR ready, large -scale platforms without a large investment in the I/O subsystem.

97 MATHEMATICS AND COMPUTING↗

LLNL Mercury Project Trinity Open Science Final Report

The Mercury Monte Carlo particle transport code developed at Lawrence Livermore National Laboratory (LLNL) is used to simulate the transport of radiation through urban environments. These challenging calculations include complicated geometries and require significant computational resources to complete. As a result, a question arises as to the level of convergence of the calculations with Monte Carlo simulation particle count. In the Trinity Open Science calculations, one main focus was to investigate convergence of the relevant simulation quantities with Monte Carlo particle count to assess the current simulation methodology. Both for this application space but also of more general applicability, we also investigated the impact of code algorithms on parallel scaling on the Trinity machine as well as the utilization of the Trinity DataWarp burst buffer technology in Mercury via the LLNL Scalable Checkpoint/Restart (SCR) library.

54 ENVIRONMENTAL SCIENCES↗

Scalable Analysis Methods and In Situ Infrastructure for Extreme Scale Knowledge Discovery

High performance computers have for many years been on a trajectory that gives them extraordinary compute power with the addition of more and more compute cores. At the same time, other system parameters such as the amount of memory per core and bandwidth to storage have remained constant or have barely increased. This creates an imbalance in the computer, giving it the ability to compute a lot of data that it cannot reasonably save out due to time and storage constraints. While technologies have been invented to mitigate this problem (burst buffers, etc.), software has been adapting to employ in situ libraries which perform data analysis and visualization on simulation data while it is still resident in memory. This avoids the need to ever have to pay the costs of writing many terabytes of data files. Instead, in situ enables the creation of more concentrated data products such as statistics, plots, and data extracts, which are all far smaller than the full-sized volume data. With the increasing popularity of in situ, multiple in situ infrastructures have been created, each with its own mechanism for integrating with a simulation. To make it easier to instrument a simulation with multiple in situ infrastructures and include custom analysis algorithms, this project created the SENSEI framework.

97 MATHEMATICS AND COMPUTING↗

Challenges and Opportunities of User-Level File Systemsfor HPC

The performance gap between magnetic disks and data processing on HPC systems has become that huge that an efficient data processing can only be achieved by introducing non-volatile memory (NVRAM) as a new storage tier. Although the benefits of hierarchical storage have been adequately demonstrated to the point that the newest leadership class HPC systems will employ burst buffers, critical questions remain for supporting hierarchical storage systems, including: How should we present hierarchical storage systems to user applications, such that they are easy to use and that application code is portable across systems? How should we manage data movement through a storage hierarchy for best performance and resilience of data? How do the particular I/O use cases mandate the way we manage data? There have been many efforts to explore this space in the form of file systems, with increasingly more implemented at the user level. This is because it is relatively easy to swap in new, specialized user-level file systems for use by applications on a case-by-case basis, as opposed to the current mainstream approach of using general-purpose, system-level file systems which may not be optimized for HPC workloads and must be installed by administrators. In contrast, file systems at the user level can be tailored for specific HPC workloads for high performance and can be used by applications without administrator intervention. Many user-level file system developers have found themselves “having to reinvent the wheel” to implement various optimizations in their file systems. Thus, a main goal of this meeting was to bring together experts in I/O performance, file systems, and storage, and collectively explore the space of current and future problems and solutions for I/O on hierarchical storage systems in order to begin a community effort in enabling user-level file system support for HPC systems. We had a lively week of learning about each other’s approaches as well as unique I/O use cases that can influence the design of a community-driven file and storage system standards. The agenda for this meeting contained talks from participants on the following high level topics: HPC storage and I/O support today; what do HPC users need for I/O; existing userlevel file system efforts; object stores and other alternative storage systems; and components for building user-level file systems. The talks were short and intended to be conversation starters for more in-depth discussions with the whole group. The participants engaged in lengthy discussions on various questions that arose from the talks including: Are we ready to program to a memory hierarchy versus block devices? Are the needs of HPC users reflected in our existing file systems and storage systems? Should we drop or keep POSIX moving forward? What do we mean when we say “user-level file system”? Do we all mean the same thing? How should the IO 500 benchmark be defined so it is fair and useful? and How are stage-in and stage-out actually going to work? The report for this seminar contains a record of the talks from the participants as well as the resulting discussions. Our hope is that the effort initiated during this seminar will result in long-term collaborations that will benefit the HPC community as a whole.

97 MATHEMATICS AND COMPUTING↗

Checkpoint-Restart and Platform Design Criteria

Long computations often use checkpoint/restart to survive the inevitable failures. Optimal checkpoint/restart strategy can be derived in terms of special functions, and approximated within the regime of interest. As a further step under nominal assumptions, platforms can be optimized for maximum progress rate, computed in terms of non-dimensional parameter JMTTI/Delta, where JMTTI is the Job Mean Time To Interrupt and Delta is the time it takes to write (or read) a checkpoint. Scaling to a larger platform of the same technology requires scaling checkpoint bandwidth quadratically to maintain the same progress rate. This was observed during the design of LANL’s Roadrunner in 2007, led to burst buffers proposed in 2009 and introduced on Trinity in 2015. The need to compute reliably continues to impact the design of future exascale platforms.

97 MATHEMATICS AND COMPUTING↗

Accelerate M-TIP on GPUs and deploy to Summit and NERSC-9 (against simulated data) WBS 2.2.4.05 ExaFEL, Milestone ADSE13-199

We evaluated the orientation matching step in the M-TIP SPI workflow for potential offloading to accelerators. We ported the code to GPUs, benchmarked it, optimized and down-selected the best versions. The accelerated version of the orientation matching code that was developed at LANL (LANL GPU v3) is 34-55X faster than sequential, 2.4-4.9X faster than the fastest OpenMP open source version we found (FAISS OpenMP) and 1.5-4X faster than the fastest GPU open source version we found (FAISS GPU). Summit single-node GPU versions were somewhat faster than Cori GPU. Image size plays a role; mid-range image sizes take more time. The LANL CUDA multi-node, multi-GPU implementation shows mostly linear strong scaling. I/O also plays a large role; splitting data into parts improves read time and burst buffers dramatically improve read times. This work will be integrated into the M-TIP workflow as part of the next milestone ADSE13-193.

97 MATHEMATICS AND COMPUTING↗

Cloud object store for archive storage of high performance computing data using decoupling middleware

Cloud object storage is enabled for archived data, such as checkpoints and results, of high performance computing applications using a middleware process. A plurality of archived files, such as checkpoint files and results, generated by a plurality of processes in a parallel computing system are stored by obtaining the plurality of archived files from the parallel computing system; converting the plurality of archived files to objects using a log structured file system middleware process; and providing the objects for storage in a cloud object storage system. The plurality of processes may run, for example, on a plurality of compute nodes. The log structured file system middleware process may be embodied, for example, as a Parallel Log-Structured File System (PLFS). The log structured file system middleware process optionally executes on a burst buffer node.

97 MATHEMATICS AND COMPUTING↗

Cloud object store for checkpoints of high performance computing applications using decoupling middleware

Cloud object storage is enabled for checkpoints of high performance computing applications using a middleware process. A plurality of files, such as checkpoint files, generated by a plurality of processes in a parallel computing system are stored by obtaining said plurality of files from said parallel computing system; converting said plurality of files to objects using a log structured file system middleware process; and providing said objects for storage in a cloud object storage system. The plurality of processes may run, for example, on a plurality of compute nodes. The log structured file system middleware process may be embodied, for example, as a Parallel Log-Structured File System (PLFS). The log structured file system middleware process optionally executes on a burst buffer node.

97 MATHEMATICS AND COMPUTING↗

Announcing Supercomputer Summit

Summit is the next leap in leadership-class computing systems for open science. With Summit we will be able to address, with greater complexity and higher fidelity, questions concerning who we are, our place on earth, and in our universe. Summit will deliver more than five times the computational performance of Titan’s 18,688 nodes, using only approximately 3,400 nodes when it arrives in 2017. Like Titan, Summit will have a hybrid architecture, and each node will contain multiple IBM POWER9 CPUs and NVIDIA Volta GPUs all connected together with NVIDIA’s high-speed NVLink. Each node will have over half a terabyte of coherent memory (high bandwidth memory + DDR4) addressable by all CPUs and GPUs plus 800GB of non-volatile RAM that can be used as a burst buffer or as extended memory. To provide a high rate of I/O throughput, the nodes will be connected in a non-blocking fat-tree using a dual-rail Mellanox EDR InfiniBand interconnect. Upon completion, Summit will allow researchers in all fields of science unprecedented access to solving some of the world’s most pressing challenges.

97 MATHEMATICS AND COMPUTING↗

Parallel checksumming of data chunks of a shared data object using a log-structured file system

Checksum values are generated and used to verify the data integrity. A client executing in a parallel computing system stores a data chunk to a shared data object on a storage node in the parallel computing system. The client determines a checksum value for the data chunk; and provides the checksum value with the data chunk to the storage node that stores the shared object. The data chunk can be stored on the storage node with the corresponding checksum value as part of the shared object. The storage node may be part of a Parallel Log-Structured File System (PLFS), and the client may comprise, for example, a Log-Structured File System client on a compute node or burst buffer. The checksum value can be evaluated when the data chunk is read from the storage node to verify the integrity of the data that is read.

97 MATHEMATICS AND COMPUTING↗

Parallel compression of data chunks of a shared data object using a log-structured file system

Techniques are provided for parallel compression of data chunks being written to a shared object. A client executing on a compute node or a burst buffer node in a parallel computing system stores a data chunk generated by the parallel computing system to a shared data object on a storage node by compressing the data chunk; and providing the data compressed data chunk to the storage node that stores the shared object. The client and storage node may employ Log-Structured File techniques. The compressed data chunk can be de-compressed by the client when the data chunk is read. A storage node stores a data chunk as part of a shared object by receiving a compressed version of the data chunk from a compute node; and storing the compressed version of the data chunk to the shared data object on the storage node.

97 MATHEMATICS AND COMPUTING↗