Search NASA⌕ Search

SEARCH · Search NASA

Results for “cache locality”

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 55 records · Page 3
All sourcesNASA NTRSDOE OSTI
DOE OSTI2017

Locality-Aware CTA Clustering For Modern GPUs

In this paper, we proposed a novel clustering technique for tapping into the performance potential of a largely ignored type of locality: inter-CTA locality. We first demonstrated the capability of the existing GPU hardware to exploit such locality, both spatially and temporally, on L1 or L1/Tex unified cache. To verify the potential of this locality, we quantified its existence in a broad spectrum of applications and discussed its sources of origin. Based on these insights, we proposed the concept of CTA-Clustering and its associated software techniques. Finally, We evaluated these techniques on all modern generations of NVIDIA GPU architectures. The experimental results showed that our proposed clustering techniques could significantly improve on-chip cache performance.

Data Locality↗
NASA NTRS2001

Using Minimum-Surface Bodies for Iteration Space Partitioning

A number of known techniques for improving cache performance in scientific computations involve the reordering of the iteration space. Some of these reorderings can be considered as coverings of the iteration space with the sets having good surface-to-volume ratio. Use of such sets reduces the number of cache misses in computations of local operators having the iteration space as a domain. We study coverings of iteration spaces represented by structured and unstructured grids. For structured grids we introduce a covering based on successive minima tiles of the interference lattice of the grid. We show that the covering has good surface-to-volume ratio and present a computer experiment showing actual reduction of the cache misses achieved by using these tiles. For unstructured grids no cache efficient covering can be guaranteed. We present a triangulation of a 3-dimensional cube such that any local operator on the corresponding grid has significantly larger number of cache misses than a similar operator on a structured grid.

Frumlin, Michael↗
DOE OSTI2019

Analyzing the suitability of contemporary 3D-stacked PIM architectures for HPC scientific applications

Scaling off-chip bandwidth is challenging due to fundamental limitations, such as a fixed pin count and plateauing signaling rates. Recently, vendors have turned to 2.5D and 3D stacking to closely integrate system components. Interestingly, these technologies can integrate a logic layer under multiple memory dies, enabling computing capability inside a memory stack. This trend in stacking is making PIM architectures commercially viable. In this work, we investigate the suitability of offloading kernels in scientific applications onto 3D stacked PIM architectures. We evaluate several hardware constraints resulted from the stacked structure. We perform extensive simulation experiments and in-depth analysis to quantify the impact of application locality in TLBs, data caches, and memory stacks. Our results also identify design optimization areas in software and hardware for HPC scientific applications.

Peng, Ivy Bo↗
NASA NTRS2002

Minimizing Cache Misses Using Minimum-Surface Bodies

A number of known techniques for improving cache performance in scientific computations involve the reordering of the iteration space. Some of these reorderings can be considered as coverings of the iteration space with the sets having good surface-to-volume ratio. Use of such sets reduces the number of cache misses in computations of local operators having the iteration space as a domain. First, we derive lower bounds which any algorithm must suffer while computing a local operator on a grid. Then we explore coverings of iteration spaces represented by structured and unstructured grids which allow us to approach these lower bounds. For structured grids we introduce a covering by successive minima tiles of the interference lattice of the grid. We show that the covering has low surface-to-volume ratio and present a computer experiment showing actual reduction of the cache misses achieved by using these tiles. For planar unstructured grids we show existence of a covering which reduces the number of cache misses to the level of structured grids. On the other hand, we present a triangulation of a 3-dimensional cube such that any local operator on the corresponding grid has significantly larger number of cache misses than a similar operator on a structured grid.

Frumkin, Michael↗
NASA NTRS1993

Avoiding and tolerating latency in large-scale next-generation shared-memory multiprocessors

A scalable solution to the memory-latency problem is necessary to prevent the large latencies of synchronization and memory operations inherent in large-scale shared-memory multiprocessors from reducing high performance. We distinguish latency avoidance and latency tolerance. Latency is avoided when data is brought to nearby locales for future reference. Latency is tolerated when references are overlapped with other computation. Latency-avoiding locales include: processor registers, data caches used temporally, and nearby memory modules. Tolerating communication latency requires parallelism, allowing the overlap of communication and computation. Latency-tolerating techniques include: vector pipelining, data caches used spatially, prefetching in various forms, and multithreading in various forms. Relaxing the consistency model permits increased use of avoidance and tolerance techniques. Each model is a mapping from the program text to sets of partial orders on program operations; it is a convention about which temporal precedences among program operations are necessary. Information about temporal locality and parallelism constrains the use of avoidance and tolerance techniques. Suitable architectural primitives and compiler technology are required to exploit the increased freedom to reorder and overlap operations in relaxed models.

Probst, David K.↗
DOE OSTI2026

FitCache: A Transparent Drop-In Framework for Multi-Tier Caching to Accelerate Distributed Deep Learning Workloads

Training in Deep learning (DL) remains highly compute- and data-intensive, with I/O becoming a critical bottleneck as models and datasets scale. Recent studies report that data loading can dominate training time, especially on large-scale HPC systems with shared parallel file systems (PFS). Existing caching approaches either rely on single-tier designs or require intrusive modifications to training pipelines, limiting their portability and effectiveness. In this work, we present FitCache, a transparent drop-in framework for multi-tier caching to accelerate distributed DL training by coordinating fast local memory (e.g., DRAM, Persistent Memory (PMem)) and NVMe as hierarchical caches atop PFS. Our design adapts to hardware diversity, i.e., if NVMe is missing, memory transparently acts as a caching tier, ensuring stable performance. FitCache transparently intercepts I/O requests and issues concurrent fetches across all tiers, returning data from the fastest responder without centralized metadata or static redirection paths. FitCache adapts to dynamic workloads and heterogeneous clusters while maintaining POSIX compatibility. Experiments on Frontier (2048 GPUs) and smaller research clusters show that FitCache reduces training time by up to 40% and per-batch I/O latency by up to 71.6% compared to Lustre Orion PFS, offering a drop-in solution for scalable DL training.

Hu, Guangxing [ORNL] (ORCID:0009000283203614)↗
NASA NTRS1997

Efficacy of Code Optimization on Cache-based Processors

The current common wisdom in the U.S. is that the powerful, cost-effective supercomputers of tomorrow will be based on commodity (RISC) micro-processors with cache memories. Already, most distributed systems in the world use such hardware as building blocks. This shift away from vector supercomputers and towards cache-based systems has brought about a change in programming paradigm, even when ignoring issues of parallelism. Vector machines require inner-loop independence and regular, non-pathological memory strides (usually this means: non-power-of-two strides) to allow efficient vectorization of array operations. Cache-based systems require spatial and temporal locality of data, so that data once read from main memory and stored in high-speed cache memory is used optimally before being written back to main memory. This means that the most cache-friendly array operations are those that feature zero or unit stride, so that each unit of data read from main memory (a cache line) contains information for the next iteration in the loop. Moreover, loops ought to be 'fat', meaning that as many operations as possible are performed on cache data-provided instruction caches do not overflow and enough registers are available. If unit stride is not possible, for example because of some data dependency, then care must be taken to avoid pathological strides, just ads on vector computers. For cache-based systems the issues are more complex, due to the effects of associativity and of non-unit block (cache line) size. But there is more to the story. Most modern micro-processors are superscalar, which means that they can issue several (arithmetic) instructions per clock cycle, provided that there are enough independent instructions in the loop body. This is another argument for providing fat loop bodies. With these restrictions, it appears fairly straightforward to produce code that will run efficiently on any cache-based system. It can be argued that although some of the important computational algorithms employed at NASA Ames require different programming styles on vector machines and cache-based machines, respectively, neither architecture class appeared to be favored by particular algorithms in principle. Practice tells us that the situation is more complicated. This report presents observations and some analysis of performance tuning for cache-based systems. We point out several counterintuitive results that serve as a cautionary reminder that memory accesses are not the only factors that determine performance, and that within the class of cache-based systems, significant differences exist.

VanderWijngaart, Rob F.↗
DOE OSTI2020

Efficient I/O for Neural Network Training with Compressed Data

FanStore is a shared object store that enables efficient and scalable neural network training on supercomputers. By providing a global cache layer on node-local burst buffers using a compressed representation, it significantly enhances the processing capability of deep learning (DL) applications on existing hardware. In addition, FanStore allows POSIX-compliant file access to the compressed data in user space. We investigate the tradeoff between runtime overhead and data compression ratio using real-world datasets and applications, and propose a compressor selection algorithm to maximize storage capacity given performance constraints. We consider both asynchronous (i.e., with prefetching) and synchronous I/O strategies, and propose mechanisms for selecting compressors for both approaches. Using FanStore, the same storage hardware can host 2-13x more data for example applications without significant runtime overhead. Empirically, our experiments show that FanStore scales to 512 compute nodes with near linear performance scalability.

Zhang, Zhao↗
DOE OSTI2024

Shifting Between Compute and Memory Bounds: A Compression-Enabled Roofline Model

In the evolving landscape of high-performance computing, especially to fight the end of Moore’s Law and Dennard’s Scaling, the ability to shift between compute-bound and memory-bound states is critical for enhancing adaptability and flexibility to diverse system and domain-specific architectures. Such capability is vital for optimizing performance across distinguished hardware configurations, such as accelerators, memory hierarchies, and cache systems. Despite that ad hoc optimization techniques, such as compressed/approximate computation, have been enabled for compute-/data-intensive computing for improved performance in distinct hardware settings, there lacks an understanding of 1) the rational behind performance improvement; 2) capability of different optimizations; 3) what optimization to respond to specific computational and memory demands. This work proposes a compression-enabled roofline model to facilitate this adaptability with data compression techniques to balance and transform between computational and memory demands. This model enables applications to adjust in response to the specific strengths and limitations of the underlying hardware and system to optimize resource utilization. The effectiveness of this approach is demonstrated with matrix multiplication kernels on different input sizes, with turning on/off various compression techniques, including 1) low-precision floating point; 2) sparse matrix formulation; and 3) compressed arrays with ZFP. By reducing memory transfer volumes and cache misses and increasing data locality and computational intensity through compression, the specific roofline model can transform between compute and memory bounds to align more efficiently with system capabilities. This advancement not only improves overall performance but also maximizes adaptability in diverse computing environments.

Naraparaju, Ramasoumya [University of Washington]↗
DOE OSTI2021

HAM: Hotspot-Aware Manager for Improving Communications with 3D-Stacked Memory

merging High-Performance Computing (HPC) workloads, such as graph analytics, machine learning, and big data science, are data-intensive. Data-intensive workloads usually present fine-grained memory accesses with limited or no data locality, and thus incur frequent cache misses and low utilization of memory bandwidth. 3D-stacked memory devices such as Hybrid Memory Cube (HMC) and High Bandwidth Memory (HBM) can provide significantly higher bandwidth than conventional memory modules. However, the traditional interfaces and optimization methods for JEDEC DDR devices do not allow to fully exploit the potential performance of 3D-stacked memory with the massive amount of irregular memory accesses of data-intensive applications. In this paper, we propose a novel Hotspot-Aware Manager (HAM) infrastructure for 3D-stacked memory devices capable of optimizing memory access streams via request aggregation, hotspot detection, and in-memory prefetching. %and an associated hotspot-aware page policy. We present the HAM design and implementation, and simulate it on a system using RISC-V embedded cores with attached HMC devices. We extensively evaluate HAM with over 12 benchmarks and applications representing diverse irregular memory access patterns. The results show that, on average, HAM reduces redundant requests by 37.51\% and increases the prefetch buffer hit rate by 4.2 times, compared to a baseline streaming prefetcher. On the selected benchmark set, HAM provides performance gains of 21.81\% in average (up to 34.28\%) and power savings of 35.07\% over a standard 3D-stacked memory.

Wang, Xi↗
DOE OSTI2020

Prioritizing local and remote memory access in a non-uniform memory access architecture

A miss in a cache by a thread in a wavefront is detected. The wavefront includes a plurality of threads that are executing a memory access request concurrently on a corresponding plurality of processor cores. A priority is assigned to the thread based on whether the memory access request is addressed to a local memory or a remote memory. The memory access request for the thread is performed based on the priority. In some cases, the cache is selectively bypassed depending on whether the memory access request is addressed to the local or remote memory. A cache block is requested in response to the miss. The cache block is biased towards a least recently used position in response to requesting the cache block from the local memory and towards a most recently used position in response to requesting the cache block from the remote memory.

Boyer, Michael W.↗
DOE OSTI2019

MAC: Memory Access Coalescer for 3D-Stacked Memory

Emerging data-intensive applications, such as graph analytics and data mining, exhibit irregular memory access patterns. Research has shown that with these memory-bound applications, traditional cache-based processor architectures, which exploit locality and regular patterns to mitigate the memory-wall issue, are inefficient. Meantime, novel 3D-stacked memory devices, such as Hybrid Mem- ory Cube (HMC) and High Bandwidth Memory (HBM), promise significant increases in bandwidth that appear extremely appealing for memory-bound applications. However, conventional memory interfaces designed for cache-based architectures and JEDEC DDR devices fit poorly with the 3D-stacked memory, which leads to significant under-utilization of the promised high bandwidth. As a response to these issues, in this paper we propose MAC (Memory Access Coalescer), a coalescing unit for the 3D-stacked memory. We discuss the design and implementation of MAC, in the context of a custom designed cache-less architecture targeted at data-intensive, irregular applications. Through a custom simulation infrastructure based on the RISC-V toolchain, we show that MAC achieves a coalescing efficiency of 52.85% on average. It improves the performance of the memory system by 60.73% on average for a large set of irregular workloads.

Wang, Xi↗
DOE OSTI2012

Local rollback for fault-tolerance in parallel computing systems

A control logic device performs a local rollback in a parallel super computing system. The super computing system includes at least one cache memory device. The control logic device determines a local rollback interval. The control logic device runs at least one instruction in the local rollback interval. The control logic device evaluates whether an unrecoverable condition occurs while running the at least one instruction during the local rollback interval. The control logic device checks whether an error occurs during the local rollback. The control logic device restarts the local rollback interval if the error occurs and the unrecoverable condition does not occur during the local rollback interval.

97 MATHEMATICS AND COMPUTING↗
NASA NTRS2008

The NPOESS Preparatory Project (NPP) Science Data Segment (SDS) Data Depository and Distribution Element (SD3E) System Architecture

The National Polar-orbiting Operational Environmental Satellite System (NPOESS), the U.S. Government's future low-Earth orbiting satellite system, will monitor global weather and environmental conditions. Serving as a risk reduction for NPOESS, the NPOESS Preparatory Project (NPP) will provide remotely sensed atmospheric, land, ocean, ozone, and sounder data that will serve the meteorological and global climate change scientific communities. The National Aeronautics and Space Administration (NASA) NPP Science Data Segment's (SDS) primary role is to independently assess the quality of the NPP science and environmental data records for their ability to support climate research. The SDS is composed of nine elements; an input element that receives data from the operational agencies and acts as a buffer, a calibration analysis element, five elements devoted to measurement based quality assessment, an element used to test algorithmic improvements, and an element that provides overall science direction. Each element requires a set of sensor specific science data products for their evaluation. There are four NPP sensors that will be flown on the NPP observatory. They are the Visible Infrared Imagining Radiometer Suite (VIIRS), the Advanced Technology Microwave Sounder (ATMS), the Cross-Track Infrared Sounder (CrIS), and the Ozone Mapper/Profiler Suite (OMPS). It is estimated that these four sensors combined will make daily data requests for approximately six terabytes of NPP science products from the operational data providers. As a result, issues associated with duplicate data requests, data transfers of large volumes of diverse products, and data transfer failures raised concerns with respect to the network traffic and bandwidth consumption. Therefore, a central data broker system for receiving and buffering data requests and data products for the SDS was developed. The data element for this system is called the SDS Data Depository and Distribution Element (SD3E). It supports science mission data assessment by assuring the timely and validated acquisition and subsequent transfer of the NPP Science Mission data to the SDS Elements and NPP Science Team. The six science elements that interface with the SD3E span across the NASA Goddard Space Flight Center (GSFC), the NASA Jet Propulsion Laboratory (JPL), and the University of Wisconsin. As the primary communication vehicle for the science elements and science team, the SD3E has an interface to the operational data providers: National Environment Satellite, Data, and Information Service (NESDIS) Interface Data Processing System (IDPS) and the National Oceanic Atmospheric Administration's (NOAA) Comprehensive Large Array-data Stewardship system (CLASS) Archive Data System (ADS), that are responsible for product generation and archive and distribution respectively. The SD3E is designed to be a semi-customizable and semi-automated system. This system is designed to provide flexibility and ease of use for the science users in accessing the latest data products by creating a rolling data cache that temporarily stores the products locally before transferring the data to the SDS Measurement based elements for the land, ocean, atmosphere, sounder, and ozone. This paper describes the design and architecture of one of the nine SDS elements, the SD3E, and how this system has provided a mechanism for efficient data exchange, how it has helped in alleviating some of the network traffic and usage, and how it has contributed to reducing operational costs.

Ho, Evelyn L.↗
DOE OSTI2016

Tag-Split Cache for Efficient GPGPU Cache Utilization

Modern GPUs employ cache to improve memory system efficiency. However, large amount of cache space is underutilized due to irregular memory accesses and poor spatial locality which exhibited commonly in GPU applications. Our experiments show that using smaller cache lines could improve cache space utilization, but it also frequently suffers from significant performance loss by introducing large amount of extra cache requests. In this work, we propose a novel cache design named tag-split cache (TSC) that enables fine-grained cache storage to address the problem of cache space underutilization while keeping memory request number unchanged. TSC divides tag into two parts to reduce storage overhead, and it supports multiple cache line replacement in one cycle.

performance and energy optimization↗
NASA NTRS2000

Interference Lattice-based Loop Nest Tilings for Stencil Computations

A common method for improving performance of stencil operations on structured multi-dimensional discretization grids is loop tiling. Tile shapes and sizes are usually determined heuristically, based on the size of the primary data cache. We provide a lower bound on the numbers of cache misses that must be incurred by any tiling, and a close achievable bound using a particular tiling based on the grid interference lattice. The latter tiling is used to derive highly efficient loop orderings. The total number of cache misses of a code is the sum of (necessary) cold misses and misses caused by elements being dropped from the cache between successive loads (replacement misses). Maximizing temporal locality is equivalent to minimizing replacement misses. Temporal locality of loop nests implementing stencil operations is optimized by tilings that avoid data conflicts. We divide the loop nest iteration space into conflict-free tiles, derived from the cache miss equation. The tiling involves the definition of the grid interference lattice an equivalence class of grid points whose images in main memory map to the same location in the cache-and the construction of a special basis for the lattice. Conflicts only occur on the boundaries of the tiles, unless the tiles are too thin. We show that the surface area of the tiles is bounded for grids of any dimensionality, and for caches of any associativity, provided the eccentricity of the fundamental parallelepiped (the tile spanned by the basis) of the lattice is bounded. Eccentricity is determined by two factors, aspect ratio and skewness. The aspect ratio of the parallelepiped can be bounded by appropriate array padding. The skewness can be bounded by the choice of a proper basis. Combining these two strategies ensures that pathologically thin tiles are avoided. They do not, however, minimize replacement misses per se. The reason is that tile visitation order influences the number of data conflicts on the tile boundaries. If two adjacent tiles are visited successively, there will be no replacement misses on the shared boundary. The iteration space may be covered with pencils larger than the size of the cache while avoiding data conflicts if the pencils are traversed by a scanning-face method. Replacement misses are incurred only on the boundaries of the pencils, and the number of misses is minimized by maximizing the volume of the scanning face, not the volume of the tile. We present an algorithm for constructing the most efficient scanning face for a given grid and stencil operator. In two dimensions it is based on a continued fraction algorithm. In three dimensions it follows Voronoi's successive minima algorithm. We show experimental results of using the scanning face, and compare with canonical loop orderings.

VanderWijngaart, Rob F.↗
DOE OSTI2024

Accelerating Application Bulk Synchronous Writes in HPC Environments

High-bandwidth storage tiers are becoming more common for their capability to absorb high-rate, bursty I/Os. Notably, the designs of these fast storage tiers differ from system to system. The variation of these layers and non-uniform methods of access can pose chal- lenges for applications seeking to run at multiple HPC facilities. Therefore, in this work, we present Spectral, a rapid-output ab- straction library to accelerate application, bulk-synchronous writes on HPC systems. We design Spectral to enable applications to use high-bandwidth storage, such as node-local storage and dis- tributed, write-caches (e.g., burst buffers) transparently without requiring modifications to the application or file system source code. The key idea is to allow applications to spend most of the time performing productive work and to not require any source code changes for maximum portability on different HPC archi- tectures. Spectral internally re-routes write-only files through available, high-performance I/O resources before ultimately mi- grating them to the shared global parallel file system. For instance, on Summit, Spectral transparently places application outputs on node-local storage and then utilizes asynchronous migration to the center-wide GPFS file system. We evaluate Spectral on the Summit HPC system (1024 nodes) using the IOR benchmark and real scientific applications. Spectral shows linear performance scaling, improving application write performance by over an order of magnitude when compared to GPFS.

Khan, Awais↗
DOE OSTI2019

Extreme-bandwidth scalable performance-per-watt GPU architecture

A technique for accessing memory in an accelerated processing device coupled to stacked memory dies is provided herein. The technique includes receiving a memory access request from an execution unit and identifying whether the memory access request corresponds to memory cells of the stacked dies that are considered local to the execution unit or non-local. For local accesses, the access is made “directly”, that is, without using a bus. A control die coordinates operations for such local accesses, activating particular through-silicon-vias associated with the memory cells that include the data for the access. Non-local accesses are made via a distributed cache fabric and an interconnect bus in the control die. Various other features and details are provided below.

97 MATHEMATICS AND COMPUTING↗