SEARCH · Engineering Papers
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.
Modern HPC systems are collecting large amounts of I/O performance data. The massive volume and heterogeneity of this data, however, have made timely performance of in-depth integrated analysis difficult. To overcome this difficulty and to allow users to identify the root causes of poor application I/O performance, we developed IOMiner, an I/O log analytics framework.
Using parallel file systems efficiently is a tricky problem due to inter-dependencies among multiple layers of I/O software, including high-level I/O libraries (HDF5, netCDF, etc.), MPI-IO, POSIX, and file systems (GPFS, Lustre, etc.). Profiling tools such as Darshan collect traces to help understand the I/O performance behavior. However, there are significant gaps in analyzing the collected traces and then applying tuning options offered by various layers of I/O software. Seeking to connect the dots between I/O bottleneck detection and tuning, we propose DXT Explorer, an interactive log analysis tool. In this paper, we present a case study using our interactive log analysis tool to identify and apply various I/O optimizations. We report an evaluation of performance improvement achieved for four I/O kernels extracted from science applications.
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, it remains unclear about the reliability challenges faced by the HPC applications under the SSD-related failures, in particular, failures resulting in data corruptions. The goal of this paper is to understand the impact of SSD-related data corruptions 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 also enables the investigation on the error resilience characteristics of the scientific file format for the first time. We demonstrate the use of FFIS with three representative real HPC applications, show 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.
Explore the source record for details and available documents.
NERSC's newest system, Perlmutter, features a 35 PB all-flash Lustre file system built on HPE Cray ClusterStor E1000. Here, we present its architecture, early performance figures, and performance considerations unique to this architecture. We demonstrate the performance of E1000 OSSes through low-level Lustre tests that achieve over 90% of the theoretical bandwidth of the SSDs at the OST and LNet levels. We also show end-to-end performance for both traditional dimensions of I/O performance (peak bulk-synchronous bandwidth) and nonoptimal workloads endemic to production computing (small, incoherent I/Os at random offsets) and compare them to NERSC's previous system, Cori, to illustrate that Perlmutter achieves the performance of a burst buffer and the resilience of a scratch file system. Finally, we discuss performance considerations unique to all-flash Lustre and present ways in which users and HPC facilities can adjust their I/O patterns and operations to make optimal use of such architectures.
NERSC's newest system, Perlmutter, features a 35 PB all-flash Lustre file system built on HPE Cray ClusterStor E1000. We present its architecture, early performance figures, and performance considerations unique to this architecture. We demonstrate the performance of E1000 OSSes through low-level Lustre tests that achieve over 90% of the theoretical bandwidth of the SSDs at the OST and LNet levels. We also show end-to-end performance for both traditional dimensions of I/O performance (peak bulk-synchronous bandwidth) and non-optimal workloads endemic to production computing (small, incoherent I/Os at random offsets) and compare them to NERSC's previous system, Cori, to illustrate that Perlmutter achieves the performance of a burst buffer and the resilience of a scratch file system. Finally, we discuss performance considerations unique to all-flash Lustre and present ways in which users and HPC facilities can adjust their I/O patterns and operations to make optimal use of such architectures.
Privateer is a general-purpose data store that optimizes the tradeoff between storage space utilization and I/O performance. Privateer uses memory-mapped I/O with private mapping and an optimized writeback mechanism to maximize write parallelism and eliminate redundant writes; it also uses contentaddressable storage to optimize storage space via de-duplication.
The combination of ever-growing scientific datasets and distributed workflow complexity creates I/O performance bottlenecks due to data volume, velocity, and variety. Although the increasing use of descriptive data formats (e.g., HDF5, netCDF) helps organize these datasets, it also creates obscure bottlenecks due to the need to translate high level operations into file addresses and then into low-level I/O operations. To address this challenge, we introduce DaYu, a method and toolset for analyzing (a) semantic relationships between logical datasets and file addresses, (b) how dataset operations translate into I/O, and (c) the combination across entire workflows. DaYu's analysis and visualization enables identification of critical bottlenecks and reasoning about remediation. We describe our methodology and propose optimization guidelines. Evaluation on scientific workflows demonstrates up to 3.7x performance improvements in I/O time for obscure bottlenecks. The time and storage overhead for DaYu's time-ordered data is typically under 0.2% of runtime and 0.25% of data volume, respectively.
Most high-fidelity physics simulation codes, such as Flash-X, need to save intermediate results (checkpoint files) to restart or gain insights into the evolution of the simulation. These simulation codes save such intermediate files synchronously, where computation is stalled while the data is written to storage. Depending on the problem size and computational requirements, this file write time can be a substantial portion of the total simulation time. In order to hide the I/O latency of checkpointing, asynchronous I/O methods have been introduced. These methods use background threads for performing I/O while the main threads continue with the simulation. The usage of background threads can compete for resources on the node as well as with communication. In this paper, we evaluate the overheads and the overall benefit of asynchronous I/O in HDF5 to simulations. Results from real-world high-fidelity simulations on the Summit supercomputer show that I/O operation is overlapped with application communication or computation or both, effectively hiding some or all of the I/O latency. Our evaluation shows that while using asynchronous I/O adds overhead to the application, the I/O time reduction is more significant, resulting in overall up to 1.5X performance speedup.
Two-phase I/O is a well-known strategy for implementing collective MPI-IO functions. It redistributes I/O requests among the calling processes into a form that minimizes the file access costs. As modern parallel computers continue to grow into the exascale era, the communication cost of such request redistribution can quickly overwhelm collective I/O performance. In this study, this effect has been observed from parallel jobs that run on multiple compute nodes with a high count of MPI processes on each node. Furthermore, to reduce the communication cost, we present a new design for collective I/O by adding an extra communication layer that performs request aggregation among processes within the same compute nodes. This approach can significantly reduce inter-node communication contention when redistributing the I/O requests. We evaluate the performance and compare it with the original two-phase I/O on Cray XC40 parallel computers (Theta and Cori) with Intel KNL and Haswell processors. Using I/O patterns from two large-scale production applications and an I/O benchmark, we show our proposed method effectively reduces the communication cost and hence maintains the scalability for a large number of processes.
This article proposes an object layout regenerator called Regen which regenerates and removes the object layout dynamically to improve the read performance of applications. Regen first detects frequent access patterns from the I/O requests of the applications. Second, Regen reorganizes the objects and regenerates or preallocates new object layouts according to the identified access patterns. Finally, Regen removes or reuses the obsolete or regenerated object layouts as necessary. As a result, Regen accelerates access to objects by providing a flexible object layout. We implement Regen as a framework on top of Proactive Data Container (PDC) and evaluate it on Cori supercomputer, a production-scale HPC system, by using realistic HPC I/O benchmarks. The experimental results show that Regen improves the I/O performance by up to 16.92 × compared with an existing system.
With the advancement of exascale computing, the amount of scientific data is increasing day by day. Efficient data access is necessary for scientific discoveries. Unfortunately, the I/O performance is not improved, like the CPU and network speed. So, I/O operations take longer time than data generation or analysis. Asynchronous I/O has been proposed to extenuate the I/O bottleneck by overlapping I/O and computation time. However, multiple small write operations can diminish the benefits of asynchronous I/O, as the I/O time becomes significantly longer than the compute time, with little time to overlap with. To overcome these issues, we present an optimization technique to merge small contiguous write operations. We integrated our solution into the HDF5 asynchronous I/O VOL connector and demonstrated the effectiveness of merging HDF5 write operations automatically and transparently without requiring any code change from the application.
Scientific data generated at experimental and observational facilities are increasingly being processed on large-scale compute systems. Most of the experimental data analysis workflows are not designed or implemented to run on large scale environments and take full advantage of HPC compute and storage resources. These applications are unlike the traditional tightly-coupled scientific applications and hence face significant performance and scalability challenges as the volume of data increases exponentially. In this paper, we conduct a performance and scalability analysis for experimental analysis applications and workflows operating on data from light sources. Our analysis detects and quantifies I/O performance, scalability and runtime bottlenecks for three data analysis applications that run on NERSC resources. Based on our analysis we propose and implement a set of optimizations that lead to reducing the amount of time spent on I/O operations by almost 90%.
In this paper, we propose OctoKV, an innovative network-based key-value storage system. OctoKV addresses the repetitive address translation overhead associated with traditional key-value stores running on file systems on the client side. To mitigate this overhead, we implemented the key-value store on the server side using NVMe-oF and a user-level NVMe driver. In particular, we employed fine-grained resource monitoring and load balancing based on heuristics to optimize I/O performance. OctoKV is deployed on a Linux cluster with Intel SPDK. The extensive evaluation shows that OctoKV achieves lower I/O response times in comparison to traditional approaches where key-value stores run on the client side. Also, the proposed load balancing strategies efficiently enhance I/O response times by equally distributing the workload from overloaded cores to other cores.
Exascale HPC systems are being designed with accelerators, such as GPUs, to accelerate parts of applications. In machine learning workloads as well as large-scale simulations that use GPUs as accelerators, the CPU (or host) memory is currently used as a buffer for data transfers between GPU (or device) memory and the file system. If the CPU does not need to operate on the data, then this is sub-optimal because it wastes host memory by reserving space for duplicated data. Furthermore, this “bounce buffer” approach wastes CPU cycles spent on transferring data. A new technique, NVIDIA GPUDirect Storage (GDS), can eliminate the need to use the host memory as a bounce buffer. Thereby, it becomes possible to transfer data directly between the device memory and the file system. This direct data path shortens latency by omitting the extra copy and enables higher-bandwidth. To take full advantage of GDS in existing applications, it is necessary to provide support with existing I/O libraries, such as HDF5 and MPI-IO, which are heavily used in applications. In this paper, we describe our effort of integrating GDS with HDF5, the top I/O library at NERSC and at DOE leadership computing facilities. We design and implement this integration using a HDF5 Virtual File Driver (VFD). The GDS VFD provides a file system abstraction to the application that allows HDF5 applications to perform I/O without the need to move data between CPUs and GPUs explicitly. We compare performance of the HDF5 GDS VFD with explicit data movement approaches and demonstrate superior performance with the GDS method.
Scientific communities are increasingly adopting deep learning (DL) models in their applications to accelerate scientific discovery processes. However, with rapid growth in the computing capabilities of HPC supercomputers, large-scale DL applications have to spend a significant portion of training time performing I/O to a parallel storage system. Previous research works have investigated optimization techniques such as prefetching and caching. Unfortunately, there exist non-trivial challenges to adopting the existing solutions on HPC supercomputers for large-scale DL training applications, which include non-performance and/or failures at extreme scale, lack of portability and generality in design, complex deployment methodology, and being limited to a specific application or dataset. To address these challenges, we propose High-Velocity AI Cache (HVAC), a distributed read-cache layer that targets and fully exploits the node-local storage or near node-local storage technology. HVAC seamlessly accelerates read I/O by aggregating node-local or near node-local storage, avoiding metadata lookups and file locking while preserving portability in the application code. We deploy and evaluate HVAC on 1,024 nodes (with over 6000 NVIDIA V100 GPUS) of the Summit supercomputer. In particular, we evaluate the scalability, efficiency, accuracy, and load distribution of HVAC compared to GPFS and XFS-on-NVMe. With four different DL applications, we observe an average 25 % performance improvement atop GPFS and 9% drop against XFS-on-NVMe, which scale linearly and are considered the performance upper bound. We envision HVAC as an important caching library for upcoming HPC supercomputers such as Frontier.
This Cooperative Research and Development Agreement (CRADA) between Oak Ridge National Laboratory (ORNL) and General Electric (GE) investigated methods for reducing the size of large computational fluid dynamics (CFD) simulation datasets using scientific data compression techniques. The work focused on adapting the MultiGrid Adaptive Reduction of Data (MGARD) compression framework and integrating it with high-performance I/O and visualization tools used in CFD workflows. MGARD uses hierarchical multilevel decomposition to enable error-controlled compression of floating-point scientific data while preserving quantities of interest. During the project, MGARD compression was integrated with the ADIOS I/O framework and visualization tools such as ParaView to enable efficient storage, transfer, and analysis of simulation data. The collaboration also explored approaches for improving compression performance for CFD data defined on unstructured meshes. Results demonstrate that scientific data compression can significantly reduce storage requirements and improve data management for large-scale CFD simulations.
Large-scale data analytics workflows ingest massive input data into various data structures, including graphs and key-value datastores. These data structures undergo multiple transformations and computations and are typically reused in incremental and iterative analytics workflows. Persisting in-memory views of these data structures enables reusing them beyond the scope of a single program run while avoiding repetitive raw data ingestion overheads. Memory-mapped I/O enables persisting in-memory data structures without data serialization and deserialization overheads. However, memory-mapped I/O lacks the key feature of persisting consistent snapshots of these data structures for incremental ingestion and processing. The obstacles to efficient virtual memory snapshots using memory-mapped I/O include background writebacks outside the application’s control, and the significantly high storage footprint of such snapshots. To address these limitations, we present Privateer, a memory and storage management tool that enables storage-efficient virtual memory snapshotting while also optimizing snapshot I/O performance. Here, we integrated Privateer into Metall, a state-of-the-art persistent memory allocator for C++, and the Lightning Memory-Mapped Database (LMDB), a widely-used key-value datastore in data analytics and machine learning. Privateer optimized application performance by 1.22× when storing data structure snapshots to node-local storage, and up to 16.7× when storing snapshots to a parallel file system. Privateer also optimizes storage efficiency of incremental data structure snapshots by up to 11× using data deduplication and compression.