Evaluating Dynamic File Striping For Lustre
Explore the source record for details and available documents.
SEARCH · Search NASA
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.
Explore the source record for details and available documents.
Combining the strengths of MPIProf and IOT, an efficient and systematic method is devised for I/O characterization at the per-job, per-rank, per-file and per-call levels of HPC programs running on the NASA Advanced Supercomputing Center. This method is applied to answer four I/O questions in this paper. A total of 13 MPI programs and 15 cases, ranging from 24 to 5968 ranks, are analyzed to establish the I/O landscape from answers to the four questions. Four of the 13 programs use MPI I/O and the behavior of their collective writes depends on the specific implementation of the MPI library used. The SGI MPT library, the prevailing MPI library for our systems, was found to gather small writes from a large number of ranks to perform larger writes by a small subset of collective buffering ranks. The number of collective buffering ranks invoked by MPT depends on the Lustre stripe count and the number of nodes used for the run. A demonstration of varying the stripe count to achieve double-digit speedup of one program's I/O was presented. Another program, which concurrently opens private files by all ranks and could potentially create a heavy load on the Lustre servers, was identified. The ability to systematically characterize I/O for a large number of programs running on a supercomputer, seek I/O optimization opportunity and identify programs that could cause a high load and instability on the filesystems is important for pursuing exascale in a real production environment.
Data-intensive science frontiers and challenges are emerg ing as computer technology has evolved substantially. Large-scale simu lations demand significant I/O workload, and as a result the I/O performance often becomes a bottleneck preventing high performance in scientific applications. In this paper we introduce a variety of I/O optimization techniques developed and imple mented when scaling a seismic application to petascale. These techniques include file system striping, data aggreg ation, reader/writer limit ing and less interleaving of data, collective MPI-IO, and data staging. The optimizations result in nearly perfect scalability of the target applicat ion on some of the most advanced petascale systems. The techniques introduced in this paper are applicab le to other scientific applications facing similar petascale I/O challenges.
As leadership computing facilities grow their storage capacity into the multi- petabyte range, the number of files and directories leap into the scale of billions. A complete profiling of such a parallel file system in a production environment presents a unique challenge. On one hand, the time, resources, and negative performance impact on production users can make regular profiling difficult. On the other hand, the result of such profiling can yield much needed understanding of the file system's general characteristics, as well as provide insight to how users write and access their data on a grand scale. This paper presents a lightweight and scalable profiling solution that can efficiently walk, analyze, and profile multi-petabyte parallel file systems. This tool has been deployed and is in regular use on very large-scale production parallel file systems at both Oak Ridge National Lab's Oak Ridge Leadership Facility (OLCF) and Lawrence Livermore National Lab's Livermore Computing (LC) facilities. We present the results of our initial analysis on the data collected from these two large-scale production systems, organized into three use cases: (1) file system snapshot and composition, (2) striping pattern analysis for Lustre, and (3) simulated storage capacity utilization in preparation for future file systems. Our analysis shows that on the OLCF file system, over 96% of user files exhibit the default stripe width, potentially limiting performance on large files by underutilizing storage servers and disks. Our simulated block analysis quantitatively shows the space overhead when doing a forklift system migration. It also reveals that due to the difference in system compositions (OLCF vs. LC), we can achieve better performance and space trade-offs by employing different native file system block sizes.
The design of HPC file and storage systems has largely been driven by the requirements on capability, reliability, and capacity. However, the convergence of large-scale simulations with big data analytics have put the data, its usability, and management back on the front and center position. One of the most common and time consuming data management tasks is the transfer of very large datasets within and across file systems.In this paper, we are introducing the FCP tool, a file system agnostic copy tool designed at the OLCF for scalable and high-performance data transfers between two file system endpoints. It provides an array of interesting features such as adaptive chunking, checksumming on the fly, checkpoint and resume capabilities to handle failures, and preserving stripe information for Lustre file system among others. It is currently available on the Titan supercomputer at the OLCF. Initial tests have shown that FCP has much better and scalable performance than traditional data copy tools and it was capable of transferring petabyte-scale datasets between two Lustre file systems.
In this paper we study the performance of the Lustre file system using five scientific and engineering applications representative of NASA workload on large-scale supercomputing systems such as NASA s Pleiades. In order to facilitate the collection of Lustre performance metrics, we have developed a software tool that exports a wide variety of client and server-side metrics using SGI's Performance Co-Pilot (PCP), and generates a human readable report on key metrics at the end of a batch job. These performance metrics are (a) amount of data read and written, (b) number of files opened and closed, and (c) remote procedure call (RPC) size distribution (4 KB to 1024 KB, in powers of 2) for I/O operations. RPC size distribution measures the efficiency of the Lustre client and can pinpoint problems such as small write sizes, disk fragmentation, etc. These extracted statistics are useful in determining the I/O pattern of the application and can assist in identifying possible improvements for users applications. Information on the number of file operations enables a scientist to optimize the I/O performance of their applications. Amount of I/O data helps users choose the optimal stripe size and stripe count to enhance I/O performance. In this paper, we demonstrate the usefulness of this tool on Pleiades for five production quality NASA scientific and engineering applications. We compare the latency of read and write operations under Lustre to that with NFS by tracing system calls and signals. We also investigate the read and write policies and study the effect of page cache size on I/O operations. We examine the performance impact of Lustre stripe size and stripe count along with performance evaluation of file per process and single shared file accessed by all the processes for NASA workload using parameterized IOR benchmark.
Data-intensive applications are becoming increasingly common in high-performance computing. Examples include combustion simulation, human genome analysis, and satellite image processing. Efficient access of data sets is critical to the performance of these applications. Because of the size of the data today's economically feasible approach is to store the data files on an array of hard disks or data servers equipped with hard disks and managed by a parallel file system such as PVFS or Lustre wherein the data is striped over a (large) number of disks for high aggregate I/O throughout. With file striping, a request for a segment of logically contiguous file space is decomposed into multiple sub-requests, each to a different server. While the data unit for this striping is usually reasonably large to benefit disk efficiency, the first and/or last sub-requests can be much smaller than the striping unit if the request does not align with the striping pattern, severely compromising hard disk efficiency and thus application performance. We propose to exploit solid state drives (SSD), whose efficiency is much less sensitive to small random accesses, to enable the alignment of requests to disk with the data striping pattern. In this scheme hard disks mainly serve large, aligned, sequential requests, with SSDs serving small or unaligned requests, thus respecting the relative cost, performance, and durability characteristics of the two media, and thereby achieving synergy in performance/cost. We will describe the design of the proposed scheme, its implementation on CCS-7's Darwin cluster, and performance results.
Active Storage is a technology aimed at reducing the bandwidth requirements of current supercomputing systems, and leveraging the processing power of the storage nodes used by some modern file systems. Active Storage achieves both objectives by moving certain processing tasks to the storage nodes, near the data they manage. Our proposal for Active Storage has several key features: user-space implementation which facilitates the port to different file systems, analytical model to anticipate the performance of Active Storage with respect to a traditional system, support for striped files and complex format files such as netCDF, scientific-friendly programming and runtime environment, etc.
Collective I/O is a widely used technique to improve I/O performance in parallel computing. It can be implemented as a client-based or server-based scheme. The client-based implementation is more widely adopted in MPI-IO software such as ROMIO because of its independence from the storage system configuration and its greater portability. However, existing implementations of client-side collective I/O do not take into account the actual pattern offile striping over multiple I/O nodes in the storage system. This can cause a significant number of requests for non-sequential data at I/O nodes, substantially degrading I/O performance. Investigating the surprisingly high I/O throughput achieved when there is an accidental match between a particular request pattern and the data striping pattern on the I/O nodes, we reveal the resonance phenomenon as the cause. Exploiting readily available information on data striping from the metadata server in popular file systems such as PVFS2 and Lustre, we design a new collective I/O implementation technique, resonant I/O, that makes resonance a common case. Resonant I/O rearranges requests from multiple MPI processes to transform non-sequential data accesses on I/O nodes into sequential accesses, significantly improving I/O performance without compromising the independence ofa client-based implementation. We have implemented our design in ROMIO. Our experimental results show that the scheme can increase I/O throughput for some commonly used parallel I/O benchmarks such as mpi-io-test and ior-mpi-io over the existing implementation of ROMIO by up to 157%, with no scenario demonstrating significantly decreased performance.
Supercomputer I/O loads are often dominated by writes. HPC (High Performance Computing) file systems are designed to absorb these bursty outputs at high bandwidth through massive parallelism. However, the delivered write bandwidth often falls well below the peak. This paper characterizes the data absorption behavior of a center-wide shared Lustre parallel file system on the Jaguar supercomputer. We use a statistical methodology to address the challenges of accurately measuring a shared machine under production load and to obtain the distribution of bandwidth across samples of compute nodes, storage targets, and time intervals. We observe and quantify limitations from competing traffic, contention on storage servers and I/O routers, concurrency limitations in the client compute node operating systems, and the impact of variance (stragglers) on coupled output such as striping. We then examine the implications of our results for application performance and the design of I/O middleware systems on shared supercomputers.
Supercomputer I/O loads are often dominated by writes. HPC (High Performance Computing) file systems are designed to absorb these bursty outputs at high bandwidth through massive parallelism. However, the delivered write bandwidth often falls well below the peak. This paper characterizes the data absorption behavior of a center-wide shared Lustre parallel file system on the Jaguar supercomputer. We use a statistical methodology to address the challenges of accurately measuring a shared machine under production load and to obtain the distribution of bandwidth across samples of compute nodes, storage targets, and time intervals. We observe and quantify limitations from competing traffic, contention on storage servers and I/O routers, concurrency limitations in the client compute node operating systems, and the impact of variance (stragglers) on coupled output such as striping. We then examine the implications of our results for application performance and the design of I/O middleware systems on shared supercomputers.
This article studies the I/O write behaviors of the Titan supercomputer and its Lustre parallel file stores under production load. The results can inform the design, deployment, and configuration of file systems along with the design of I/O software in the application, operating system, and adaptive I/O libraries.We propose a statistical benchmarking methodology to measure write performance across I/O configurations, hardware settings, and system conditions. Moreover, we introduce two relative measures to quantify the write-performance behaviors of hardware components under production load. In addition to designing experiments and benchmarking on Titan, we verify the experimental results on one real application and one real application I/O kernel, XGC and HACC IO, respectively. These two are representative and widely used to address the typical I/O behaviors of applications.In summary, we find that Titan’s I/O system is variable across the machine at fine time scales. This variability has two major implications. First, stragglers lessen the benefit of coupled I/O parallelism (striping). Peak median output bandwidths are obtained with parallel writes to many independent files, with no striping or write sharing of files across clients (compute nodes). I/O parallelism is most effective when the application—or its I/O libraries—distributes the I/O load so that each target stores files for multiple clients and each client writes files on multiple targets in a balanced way with minimal contention. Second, our results suggest that the potential benefit of dynamic adaptation is limited. In particular, it is not fruitful to attempt to identify “good locations” in the machine or in the file system: component performance is driven by transient load conditions and past performance is not a useful predictor of future performance. For example, we do not observe diurnal load patterns that are predictable.