Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “Heterogeneous memory systems”

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 19 records
All sourcesNASA NTRSDOE OSTI
DOE OSTI2025

Flexible and Effective Object Tiering for Heterogeneous Memory Systems

Computing platforms that package multiple types of memory, each with their own performance characteristics, are quickly becoming mainstream. To operate efficiently, heterogeneous memory architectures require new data management solutions that are able to match the needs of each application with an appropriate type of memory. As the primary generators of memory usage, applications create a great deal of information that can be useful for guiding memory management, but the community still lacks tools to collect, organize, and leverage this information effectively. To address this gap, this work introduces a novel software framework that collects and analyzes object-level information to guide memory tiering. The framework includes tools to monitor the capacity and usage of individual data objects, routines that aggregate and convert this information into tier recommendations for the host platform, and mechanisms to enforce these recommendations according to user-selected policies. Moreover, the developed tools and techniques are fully automatic, work on standard Linux systems, and do not require modification or recompilation of existing software. Using this framework, this study evaluates and compares the impact of a variety of design choices for memory tiering, including different policies for prioritizing objects for the fast memory tier as well as the frequency and timing of migration events. In conclusion, the results, collected on a modern Intel platform with conventional DDR4 SDRAM as well as Intel Optane NVRAM, show that guiding data tiering with object-level information can enable significant performance and efficiency benefits compared with standard hardware- and software-directed data-tiering strategies for a diverse set of memory-intensive workloads.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2024

CG-Kit: Code Generation Toolkit for performant and maintainable variants of source code applied to Flash-X hydrodynamics simulations

CG-Kit is a new Code Generation tool-Kit that we have developed as a part of the solution for portability and maintainability for multiphysics computing applications. The development of CG-Kit is rooted in the urgent need created by the shifting landscape of high-performance computing platforms and the algorithmic complexities of a particular large-scale multiphysics application: Flash-X. To efficiently use computing resources on a heterogeneous node, an application must have a map of computation to resources and a mechanism to move the data and computation to the resources according to the map. Most existing performance portability solutions are focussed on abstracting the expression of computations so that a unified source code can be specialized to run on different resources. However, such an approach is insufficient for a code like Flash-X, which has a multitude of code components that can be assembled in various permutations and combinations to form different instances of applications. Similar challenges apply to any code that has composability, where a single specified way of apportioning work among devices may not be optimal. Additionally, use cases arise where the optimal control flow of computation may differ for different devices while the underlying numerics remain identical. This combination leads to unique challenges including handling an existing large code base in Fortran and/or C/C++, subdivision of code into a great variety of units supporting a wide range of physics and numerical methods, different parallelization techniques for distributed and shared memory systems and accelerator devices, and heterogeneity of computing platforms requiring coexisting variants of parallel algorithms. All of these challenges demand that scientific software developers apply existing knowledge about domain applications, algorithms, and computing platforms to determine custom abstractions and granularity for code generation. There is a critical lack of tools to tackle those problems. CG-Kit is designed to fill this gap by providing a user with the ability to express their desired control flow and computation-to-resource map in the form a pseudocode-like recipe. It consists of standalone tools that can be combined into highly specific and, we argue, highly effective portability and maintainability toolchains. Here we present the design of our new tools: parametrized source trees, control flow graphs, and recipes. The tools are implemented in Python. They are agnostic to the programming language of the source code targeted for code generation. In conclusion, we demonstrate the capabilities of the toolkit with two examples, first, multithreaded variants of the basic AXPY operation, and second, variants of parallel algorithms within a hydrodynamics solver, called Spark, from Flash-X that operates on block-structured adaptive meshes.

Algorithmic portability↗
DOE OSTI2025

IRIS-MEMFLOW: Data Flow-Enabled Portable Memory Orchestration in IRIS Runtime for Diverse Heterogeneity

Task-based programming models and execution paradigms provide a means to decompose a computation by expressing it as a graph in which each node represents a specific computation operating on memory objects and the edges define the dependencies in the execution flow. In this execution model, independent nodes in the graph can be executed concurrently in different computing devices, making it suitable for heterogeneous systems in which computing devices with different architectures coexist. However, careful memory orchestration across heterogeneous devices is needed because copies of the same memory object may reside in multiple devices during execution. Manually ensuring such an orchestration is quite challenging. Not only must an application developer guard against race conditions, but they must also optimize data movement between the host and devices because unnecessary data movement significantly impacts performance. To mitigate these challenges, we enhance the IRIS heterogeneous runtime and introduce IRIS-MEMFLOW–a data flow–enabled portable memory abstraction for seamlessly orchestrating memory in diverse heterogeneous computing environments. By using data-flow analysis, IRIS-MEMFLOW guards against race conditions while multiple heterogeneous devices access memory objects. IRIS-MEMFLOW also optimizes data movement between the host and devices without manual intervention. As a result, IRIS provides improved programming productivity, performance, and portability for multidevice heterogeneous executions in high-performance computing and cloud systems that run diverse architectures from different vendors. The efficacy of IRIS-MEMFLOW is evaluated through experiments that show its capability in terms of programming productivity, multidevice heterogeneity, portability, and low overhead versus the state of the art.

Monil, M. A. H. [ORNL] (ORCID:0000000334194037)↗
DOE OSTI2026

Risk-Aware Measurement Synchronization and Recovery for DSSE With Heterogeneous Data Sources

Power distribution systems are increasingly integrating heterogeneous sensors with varying data reporting rates and types, which pose challenges to achieving observability at the desired temporal resolution of distribution system state estimation (DSSE). Multisensor failures caused by extreme events exacerbate these issues, introducing substantial uncertainties into DSSE. This article proposes a novel solution to these challenges by ensuring high-resolution system observability despite heterogeneous data sources and multisensor failures. First, a deep learning architecture combining long short-term memory (LSTM) and graph convolutional network (GCN) is employed to synchronize meters with different reporting rates, aiming to achieve system observability. A random-walk-model-based approach is introduced to generate pseudo-measurements while properly characterizing their uncertainties under multisensor failures. Finally, a disaster-risk-informed observability metric (RiOM) is defined to quantify the uncertainty associated with state estimation results. The proposed framework offers deeper insights into the system observability on the fly compared with conventional analysis. The effectiveness of the framework is demonstrated on an IEEE standard test case and a large-scale real-world distribution feeder in mid-Minnesota in the U.S.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2024

ECP libraries and tools: An overview

The Exascale Computing Project (ECP) Software Technology and Co-Design teams addressed the growing complexities in high-performance computing (HPC) by developing scalable software libraries and tools that leverage exascale system capabilities. As we enter the exascale era, the need for reusable, optimized software solutions that can handle the unique challenges posed by these systems becomes increasingly important. The primary challenges the ECP teams faced were to create software libraries and tools that are performant on exascale architectures and portable and usable across diverse hardware platforms. Efforts addressed issues related to concurrent execution, memory management, and the integration of heterogeneous computing resources, such as GPUs from multiple vendors. The ECP’s strategy involved a structured development process encompassing the creation, optimization, and deployment of software in collaboration with industry, academia, and national laboratories. The project was organized into several technical areas: co-design of domain-specific suites with target applications, programming models and runtimes, development tools, mathematical libraries, data and visualization tools, and software ecosystem and delivery mechanisms. ECP has successfully developed a large portfolio of software libraries and tools that demonstrate significant improvements in performance and scalability on exascale systems. These products have been integrated into the Department of Energy’s computing facilities, supporting various scientific applications and ensuring robust performance across different hardware setups. ECP advancements in software development for exascale computing highlight the importance of a collaborative and adaptive approach to handling next-generation HPC systems complexities. The lessons learned emphasize the need for continuous engagement with end-users and vendors, and the importance of maintaining a balance between innovation and practical implementation. Future efforts will focus on ensuring scalability, keeping pace with rapid hardware advancements, and further enhancing the interoperability and usability of the software ecosystem. In conclusion, subsequent articles in this special issue provide in-depth discussions and case studies into specific library and tool efforts.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2025

Dispatch Manager for NEML2 Constitutive Model Calculations Embedded in MOOSE

This report describes the extended capabilities of the NEML2 constitutive modeling library, including a flexible and efficient work dispatching system designed to leverage both CPU and GPU resources. This enhancement addresses one of the primary computational challenges in large-scale simulations: the ability to distribute and execute batches of material model evaluations across heterogeneous computing devices. The new dispatch system introduces a modular set of dispatcher and scheduler classes that coordinate the flow of data and execution between devices. The dispatcher is responsible for efficiently packaging work, managing device-specific memory operations, and synchronizing results. This modularity allows for extensibility, making it straightforward to integrate additional computing backends in the future. From an implementation standpoint, the dispatcher system interfaces seamlessly with NEML2's existing models. They handle device-aware tensor operations, optimize memory transfers, and support asynchronous execution when applicable. This design ensures that batches of material points can be evaluated concurrently, substantially improving throughput compared to previous single-device or serial implementations. These improvements not only enhance the raw performance of NEML2 but also improve its usability in multiscale and high-fidelity simulations, where the simultaneous evaluation of large material point batches is critical. Benchmarks included in the report demonstrate the system’s scalability, highlighting its effectiveness when leveraging modern GPU architectures.

22 GENERAL STUDIES OF NUCLEAR REACTORS↗
DOE OSTI2024

Performance Improvements of the Griffin Solvers in FY24

The Griffin code is a MOOSE-based reactor physics application jointly developed by Idaho National Laboratory and Argonne National Laboratory under the Department of Energy Office of Nuclear Energy Nuclear Energy Advanced Modeling and Simulation Program. This fiscal year, we have made significant efforts to improve the performance of transport solver options and cross-section generation for the efficient use of Griffin in advanced reactor applications. For the HFEM-PN solver, the residual evaluations of HFEM kernels were optimized by utilizing the pre- computed averaged cross sections for individual elements. Numerical integration involving the evaluation of basis functions at quadrature points was bypassed by facilitating precomputed element mass matrices for response matrices. Red-black iterations were improved by introducing a new generalized minimum residual based solver. The memory usage of response matrix storage was significantly reduced by applying basis function rotations on interfaces and calculating volumetric odd-parity moments on the fly. Additionally, the adjoint flux and transient calculation capabilities of the HFEM-PN solver were successfully implemented and verified using the TWIGL benchmark problem. For the DFEM-SN solver, memory footprint and computation time were significantly reduced by not treating angular flux vectors as the MOOSE nonlinear system vectors. Specifically for IQS, scalar adjoint weighting was introduced to further eliminate angular adjoint flux storage in the MOOSE auxiliary system. It was demonstrated through the three-dimensional Advanced Burner Test Reactor core problem that the memory usage for transient calculations with the IQS method was reduced by over 7.5× compared to before the optimizations. For the self-shielding application programming interface, a new double-heterogeneity treatment method, named the Bell Function-Based Analytic Two-Region Slowing Down Method, was developed to efficiently flux-volume homogenize TRISO particles with the matrix. Additionally, optimizations were made to hyper- fine group (HFG) slowing down calculations by pretabulating collision probability coefficients and grouping isotopes, significantly reducing the computational time for calculating scattering sources per HFG. Lastly, the pin power reconstruction module was extended to account for temporal behavior in a microreactor analysis problem, specifically for a control drum transient. Verification tests for each of these improvements demonstrated significant performance enhancements and memory reduction.

22 - GENERAL STUDIES OF NUCLEAR REACTORS↗
DOE OSTI2026

Scalable Quantum Monte Carlo Method for Polariton Chemistry via Mixed Block Sparsity and Tensor Hypercontraction Method

We present a reduced-scaling auxiliary-field quantum Monte Carlo (AFQMC) framework designed for large molecular systems and ensembles, with or without coupling to optical cavities. Our approach leverages the natural block sparsity of the Cholesky decomposition (CD) of electron repulsion integrals in molecular ensembles and employs tensor hypercontraction (THC) to efficiently compress low-rank Cholesky blocks. By representing the Cholesky vectors in a mixed format, keeping high-rank blocks in block-sparse form and compressing low-rank blocks with THC, we reduce the scaling of exchange-energy evaluation from quartic to robust cubic in the number of molecular orbitals N, while lowering memory from cubic toward quadratic. Benchmark analyses on one-, two-, and three-dimensional molecular ensembles (up to ∼1,200 orbitals) show that (a) the number of nonzeros in Cholesky tensors grows linearly with system size across dimensions; (b) the average numerical rank increases sublinearly and does not saturate at these sizes; and (c) rank heterogeneity─some blocks nearly full rank and many low rank, naturally motivates the proposed mixed block sparsity and THC scheme for efficient calculation of exchange energy. In conclusion, we demonstrate that the mixed scheme yields cubic wall-time scaling with favorable prefactors and preserves AFQMC accuracy.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
DOE OSTI2024

IRIS-GNN: Leveraging Graph Neural Networks for Scheduling on Truly Heterogeneous Runtime Systems

The diversity of accelerators in computer systems poses significant challenges for software developers, such as managing vendor-specific compiler toolchains, code fragmentation requiring different kernel implementations, and performance portability issues. To address these, the Intelligent Runtime System (IRIS) was developed. IRIS works across various systems, from smartphones to supercomputers, enabling automatic performance scaling based on available accelerators. It introduces abstract tasks for seamless execution transitions between accelerators while ensuring memory consistency and task dependencies. Although IRIS simplifies system details, optimal dynamic scheduling still requires user input to understand workload structures. To address this, we introduce a new scheduling policy for IRIS, termed IRIS-GNN, which is the first IRIS hybrid policy that operates in conjunction with the dynamic policies. This policy employs a Graph-Neural Network (GNN) to conduct Graph Classification of any task graphs submitted to IRIS. This GNN analyzes the structure and attributes of the task graph, categorizing it as either locality, concurrency, or mixed. This classification subsequently guides the selection of the dynamic policy used by IRIS. We provide a comparison of the performance of IRIS-GNN against the complete spectrum of IRIS’s dynamic policies, assess the overhead introduced by the GNN within this scheduling framework, and ultimately explore its practical application in real-world scenarios.

Johnston, Beau↗
DOE OSTI2025

BULKI-Store v0.3.2

BULKI-Store is a distributed object storage system optimized for high-performance computing environments. Built with a Rust core and Python bindings, it efficiently manages scientific and machine learning datasets across HPC clusters. The system employs a client-server architecture with MPI integration, enabling seamless scaling on supercomputers like Perlmutter. BULKI-Store's object-oriented approach provides intuitive data organization with rich metadata support, contrasting with traditional file-based solutions. Key optimizations include selective checkpoint loading, unified checkpoint files, and object chunking for large data transfers. For machine learning workloads, BULKI-Store offers advantages through fine-grained access patterns, dynamic data sharing between training instances, and reduced memory pressure. Memory management features include strategic Python GC calls, minimized data copies, and batch processing capabilities. The system leverages Rayon's thread pool for asynchronous data prefetching and supports multiple CPU architectures (ARM64, x86, AMD, RISC-V). By combining performance optimizations with developer-friendly APIs, BULKI-Store addresses the complex data management challenges of modern HPC applications while maintaining compatibility across heterogeneous computing environments.

Zhang, Wei [Lawrence Berkeley National Laboratory ↗
DOE OSTI2024

IRIS: A Performance-Portable Framework for Cross-Platform Heterogeneous Computing

From edge to exascale, computer architectures are becoming more heterogeneous and complex. The systems typically have fat nodes, with multicore CPUs and multiple hardware accelerators such as GPUs, FPGAs, and DSPs. This complexity is causing a crisis in programming systems and performance portability. Several programming systems are working to address these challenges, but the increasing architectural diversity is forcing software stacks and applications to be specialized for each architecture. As we show, all of these approaches critically depend on their software framework for discovery, execution, scheduling, and data orchestration. To address this challenge, we believe that a more agile and proactive software framework is essential to increase performance portability and improve user productivity. To this end, we have designed and implemented IRIS: a performance-portable framework for cross-platform heterogeneous computing. IRIS can discover available resources, manage multiple diverse programming platforms (e.g., CUDA, Hexagon, HIP, Level Zero, OpenCL, OpenMP) simultaneously in the same execution, respect data dependencies, orchestrate data movement proactively, and provide for user-configurable scheduling. To simplify data movement, IRIS introduces a shared virtual device memory with relaxed consistency among different heterogeneous devices. IRIS also adds an automatic kernel workload partitioning technique using the polyhedral model so that it can resize kernels for a wide range of devices. Our evaluation on three architectures, ranging from Qualcomm Snapdragon to a Summit supercomputer node, shows that IRIS improves portability across a wide range of diverse heterogeneous architectures with negligible overhead.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2025

Benchmarking Operators in Deep Neural Networks for Improving Performance Portability of SYCL

SYCL is a portable programming model for heterogeneous computing, so it is important to obtain reasonable performance portability of SYCL. Towards the goal of better understanding and improving performance portability of SYCL for machine learning workloads, we have been developing benchmarks for basic operators in deep neural networks (DNNs). These operators could be offloaded to heterogeneous computing devices such as graphics processing units (GPUs) to speed up computation. In this paper, we introduce the benchmarks, evaluate the performance of the operators on GPU-based systems, and describe the causes of the performance gap between the SYCL and Compute Unified Device Architecture (CUDA) kernels. We find that the causes are related to the utilization of the texture cache for read-only data, optimization of the memory accesses with strength reduction, use of local memory, and register usage per thread. We hope that the efforts of developing benchmarks for studying performance portability will stimulate discussion and interactions within the community.

Jin, Zheming [ORNL] (ORCID:000000027197780X)↗
DOE OSTI2024

Evaluating Operators in Deep Neural Networks for Improving Performance Portability of SYCL

SYCL is a portable programming model for heterogeneous computing, so it is important to obtain reasonable performance portability of SYCL. Towards the goal of better understanding and improving performance portability of SYCL for machine learning workloads, we have been developing benchmarks for basic operators in deep neural networks (DNNs). These operators could be offloaded to heterogeneous computing devices such as graphics processing units (GPUs) to speed up computation. In this work, we introduce the benchmarks, evaluate the performance of the operators on GPU-based systems, and describe the causes of the performance gap between the SYCL and Compute Unified Device Architecture (CUDA) kernels. We find that the causes are related to the utilization of the texture cache for read-only data, optimization of the memory accesses with strength reduction, shared local memory accesses, and register usage per thread. We hope that the efforts of developing benchmarks for studying performance portability will stimulate discussion and interactions within the community.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2024

P38 heterogeneous multi-tiled system with support for message queues (MoSAIC) v0.1

The proposed system is written in the hardware description language (HDL) verilog targeting an FPGA board. It is intended as a testbed to explore architecture tradeoffs in multi-tiled heterogeneous architectures. Although we target FPGAs, the system can be implemented as a monolithic SoC or a package comprised of many chiplets that are interconnected in the same package using a NoC. The proposed NoC is lightweight and follows an axi-lite interface. The endpoints of the NoC are a heterogeneous mix of "tiles" as endpoints that are general purpose processors, fixed function accelerators, and programmable accelerators. We assume that the network interfaces for the NoC endpoints are all addressable in a global name-space in that they represent an address range (for memory addresses) or a range of unique identifiers that are associated with each individual tile. This makes the functionality abstract from the standpoint of the NoC design details. Message queues offer a direct inter-processor interface between peer general purpose cores and diverse accelerators that comprise an SoC. Although they share the same NoC infrastructure for inter-tile communication within an SoC or SiP, the hardware message queues bypass the memory hierarchy and thus do not pollute the memory state or invoke the cache coherence mechanism.

Gonzalez, LouisaPatricia↗
DOE OSTI2025

CRCNS22 Learning Rules in the Hippocampus and their Mapping to Neuromorphic Systems (Final Technical Report)

Large scale biologically-realistic computational models are key to investigating the interplay between structure and function in nervous systems, thus paving the way to new clinical methods and neuro-inspired computing solutions. This project focuses on the hippocampus, in particular the CA3-CA1 regions, due to their role in associative learning and memory, pattern separation and completion, and spatial navigation. Investigations into the neuronal organization and learning rule(s) of this circuit can shed light into how declarative memories are formed, stored, recalled and forgotten and inform computational, experimental and clinical neuroscience work. Our project aims at developing a novel data-driven methodology supported by a broad heterogeneous base of neuroscience experimental knowledge and inspired from advances in computer science and engineering. Specifically, this work will benchmark existing and new learning rules within a full-scale spiking neural network simulation of the CA3-CA1 region. The model will be based on an open-source repository, called the Hippocampome, which contains neuronal morphologies, firing patterns, synapse probabilities, and most other required parameters for all known neuron types in the rodent hippocampal formation. The model will be first trained in a supervised fashion for associative memory tasks using backpropagation through time traditionally used in computer science, enhanced with a new technique called the surrogate gradient method. This optimization method will be used to obtain a global loss minimization, but it is not biologically inspired as it assumes the use of data not locally available to the synapses. However, we propose its use as a benchmarking tool, to compare the training performance of local biologically plausible and hardware-mappable learning rules at scale. New rules or combinations will be proposed and tested as needed, based on the obtained results. Progress in this area will also drive the development of novel hardware-mappable algorithms for continual lifelong learning and categorization of new events from few presented examples. This project goes beyond the existing state-of-the-art by looking at large scale realistic neuronal circuits as networks trainable via global optimization methods such as surrogate gradient descent. The objective function of the brain that supports learning is largely unknown, but it is likely that it operates through local learning rules. Studying network trajectories around local minima as proposed in this work represents a useful strategy for understanding whether a network is training by using a specific (set of) learning rule(s). Starting from a completely untrained network is a challenging test since it is difficult to determine how the learning rule affects the trajectory of the network. This interdisciplinary project will help understand what rule governs learning in these regions or if multiple learning rules are involved. The work will develop a robust methodology to measure if the network is converging to the target solution, oscillating around it, or diverging away.

59 BASIC BIOLOGICAL SCIENCES↗
DOE OSTI2026

Block encoding of the three-dimensional heterogeneous Poisson equation with application to fracture flow

Quantum linear system (QLS) algorithms offer the potential to solve large-scale linear systems exponentially faster than classical methods. However, applying QLS algorithms to real-world problems remains challenging due to issues such as state preparation, data loading, and efficient information extraction. In this work, we study the feasibility of applying QLS algorithms to solve discretized three-dimensional (3D) heterogeneous Poisson equations, with specific examples relating to groundwater flow through geologic fracture networks. We explicitly construct a block encoding for the 3D heterogeneous Poisson matrix by leveraging the sparse local structure of the discretized operator. While classical solvers benefit from preconditioning, we show that block encoding the system matrix and preconditioner separately does not improve the effective condition number that dominates the QLS run-time. This differs from classical approaches where the preconditioner and the system matrix can often be implemented independently. Nevertheless, due to the structure of the problem in three dimensions, the quantum algorithm achieves a run-time of 𝑂⁡(𝑁 2/3 polylog 𝑁 ⋅log (1/𝜖)), outperforming the best classical methods (with run times of 𝑂⁡(𝑁⁢log 𝑁 ⋅log (1/𝜖))) and offering exponential memory savings. These results highlight both the promise and limitations of QLS algorithms for practical scientific computing, and point to effective condition-number reduction as a key barrier in achieving quantum advantages.

58 GEOSCIENCES↗
DOE OSTI2026

Thin-film lithium niobate on diamond as a platform for efficient spin-phonon coupling

Negatively charged silicon vacancy (SiV − ) centers in diamond are leading candidates for solid-state quantum memory that can be controlled using electromagnetic or acoustic waves. The latter are particularly promising due to the strong strain response of SiV − , enabling large spin-phonon interaction strengths. Indeed, coherent spin control via surface acoustic waves (SAWs) has been demonstrated and is essential for developing on-chip phononic quantum networks. However, the absence of piezoelectricity in diamond requires an interface with a piezoelectric material for efficient transduction and delivery of acoustic waves to spins in diamonds. Here, we demonstrate a heterogeneously integrated phononic platform that combines thin-film lithium niobate with diamond to enable acoustic control of single SiV − spins. Additionally, leveraging large SAW-induced strain at the location of SiV − , we achieve coherent acoustic control of an electron spin with more than twofold improvement in Rabi frequency compared to state-of-the-art devices based on aluminum nitride on diamond. Furthermore, this work represents a crucial step towards realizing phonon-based quantum information processing systems.

Optically detected magnetic resonance↗
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)↗