Search NASA⌕ Search

SEARCH · Search NASA

Results for “memory usage”

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

Speeding up and reducing memory usage for scientific machine learning via mixed precision

Scientific machine learning (SciML) has emerged as a versatile approach to address complex computational science and engineering problems. Within this field, physics-informed neural networks (PINNs) and deep operator networks (DeepONets) stand out as the leading techniques for solving partial differential equations by incorporating both physical equations and experimental data. However, training PINNs and DeepONets require significant computational resources, including long computational times and large amounts of memory. In search of computational efficiency, training neural networks using half precision (float16) rather than the conventional single (float32) or double (float64) precision has gained substantial interest, given the inherent benefits of reduced computational time and memory consumed. However, we find that float16 cannot be applied to SciML methods, because of gradient divergence at the start of training, weight updates going to zero, and the inability to converge to a local minima. To overcome these limitations, we explore mixed precision, which is an approach that combines the float16 and float32 numerical formats to reduce memory usage and increase computational speed. Our experiments showcase that mixed precision training not only substantially decreases training times and memory demands but also maintains model accuracy. Here, we also reinforce our empirical observations with a theoretical analysis. The research has broad implications for SciML in various computational applications.

97 MATHEMATICS AND COMPUTING↗

Analysis of Vector Particle-In-Cell (VPIC) memory usage optimizations on cutting-edge computer architectures

Vector Particle-In-Cell (VPIC) is one of the fastest plasma simulation codes in the world, with particle numbers ranging from one trillion on the first petascale system, Roadrunner, to ten trillion particles on the more recent Blue Waters supercomputer. As supercomputers continue to grow rapidly in size, so too does the gap between computing capability and memory capability. Current memory systems limit VPIC simulations greatly as the maximum number of particles that can be simulated directly depends on the available memory. In this study, we present a suite of VPIC memory optimizations (i.e., particle weight, half-precision, and fixed-point optimizations) that enable a significant increase in the number of particles in VPIC simulations. Here, we assess the optimizations’ impact on memory and runtime performance for a suite of cutting-edge computer architectures such has the NVIDIA V100 GPU, the IBM Power9, and the Fujitsu A64FX architectures. Our optimizations enable a 31.25% reduction in memory usage and up to 40% increase in the number of particles. This paper extends our work on developing particle storage format optimizations Tan et al.

97 MATHEMATICS AND COMPUTING↗

Memory Optimization for Phase-field Simulations

Phase-field simulations are computationally and memory intensive applications. Many of the phase-field simulations being conducted in support of NEAMS were not capable of running on “normal clusters” with 2-4GB of RAM per core, and instead required specialized “big-memory” clusters with 64GB per core. To address this issue, the MOOSE team developed a new Python-based utility called MemoryLogger, and applied it to locate, diagnose, and eradicate memory bottlenecks within the MOOSE framework. MemoryLogger allows for a better understanding of the memory usage of an application being run in parallel across a cluster. Memory usage information is captured for every individual process in a parallel job, and communicated to the head node of the cluster. Console text output from the application itself is automatically matched with this memory usage information to produce a detailed picture of memory usage over time, making it straightforward to identify the subroutines which contribute most to the application’s peak memory usage. The information produced by the MemoryLogger quickly and effectively narrows the search for memory optimizations to the most data-intensive parts of the simulation.

99 GENERAL AND MISCELLANEOUS↗

On the memory attribution problem: A solution and case study using MPI

As parallel applications running on large–scale computing systems become increasingly memory constrained, the ability to attribute memory usage to the various components of the application is becoming increasingly important. We present the design and implementation of memnesia, a novel memory usage profiler for parallel and distributed message–passing applications. Our approach captures both application– and message–passing library–specific memory usage statistics from unmodified binaries dynamically linked to a message–passing communication library. Using microbenchmarks and proxy applications, we evaluated our profiler across three Message Passing Interface (MPI) implementations and two hardware platforms. Furthermore, the results show that our approach and the corresponding implementation can accurately quantify memory resource usage as a function of time, scale, communication workload, and software or hardware system architecture, clearly distinguishing between application and MPI library memory usage at a per–process level. With this new capability, we show that job size, communication workload, and hardware/software architecture influence peak runtime memory usage. In practice, this tool provides a potentially valuable source of information for application developers seeking to measure and optimize memory usage.

97 MATHEMATICS AND COMPUTING↗

CCC7 Cielo Report

ASC Level 2 Milestone FY 2013 continuation. L2 revealed memory pressures from using in situ analysis; Developed tools to determine memory usage; Reduced memory footprint by more than 50%.

97 MATHEMATICS AND COMPUTING↗

Ares I-X Ground Diagnostic Prototype

Automating prelaunch diagnostics for launch vehicles offers three potential benefits. First, it potentially improves safety by detecting faults that might otherwise have been missed so that they can be corrected before launch. Second, it potentially reduces launch delays by more quickly diagnosing the cause of anomalies that occur during prelaunch processing. Reducing launch delays will be critical to the success of NASA's planned future missions that require in-orbit rendezvous. Third, it potentially reduces costs by reducing both launch delays and the number of people needed to monitor the prelaunch process. NASA is currently developing the Ares I launch vehicle to bring the Orion capsule and its crew of four astronauts to low-earth orbit on their way to the moon. Ares I-X will be the first unmanned test flight of Ares I. It is scheduled to launch on October 27, 2009. The Ares I-X Ground Diagnostic Prototype is a prototype ground diagnostic system that will provide anomaly detection, fault detection, fault isolation, and diagnostics for the Ares I-X first-stage thrust vector control (TVC) and for the associated ground hydraulics while it is in the Vehicle Assembly Building (VAB) at John F. Kennedy Space Center (KSC) and on the launch pad. It will serve as a prototype for a future operational ground diagnostic system for Ares I. The prototype combines three existing diagnostic tools. The first tool, TEAMS (Testability Engineering and Maintenance System), is a model-based tool that is commercially produced by Qualtech Systems, Inc. It uses a qualitative model of failure propagation to perform fault isolation and diagnostics. We adapted an existing TEAMS model of the TVC to use for diagnostics and developed a TEAMS model of the ground hydraulics. The second tool, Spacecraft Health Inference Engine (SHINE), is a rule-based expert system developed at the NASA Jet Propulsion Laboratory. We developed SHINE rules for fault detection and mode identification. The prototype uses the outputs of SHINE as inputs to TEAMS. The third tool, the Inductive Monitoring System (IMS), is an anomaly detection tool developed at NASA Ames Research Center and is currently used to monitor the International Space Station Control Moment Gyroscopes. IMS automatically "learns" a model of historical nominal data in the form of a set of clusters and signals an alarm when new data fails to match this model. IMS offers the potential to detect faults that have not been modeled. The three tools have been integrated and deployed to Hangar AE at KSC where they interface with live data from the Ares I-X vehicle and from the ground hydraulics. The outputs of the tools are displayed on a console in Hangar AE, one of the locations from which the Ares I-X launch will be monitored. The full paper will describe how the prototype performed before the launch. It will include an analysis of the prototype's accuracy, including false-positive rates, false-negative rates, and receiver operating characteristics (ROC) curves. It will also include a description of the prototype's computational requirements, including CPU usage, main memory usage, and disk usage. If the prototype detects any faults during the prelaunch period then the paper will include a description of those faults. Similarly, if the prototype has any false alarms then the paper will describe them and will attempt to explain their causes.

Schwabacher, Mark↗

Teuchos C++ memory management classes, idioms, and related topics, the complete reference : a comprehensive strategy for safe and efficient memory management in C++ for high performance computing.

The ubiquitous use of raw pointers in higher-level code is the primary cause of all memory usage problems and memory leaks in C++ programs. This paper describes what might be considered a radical approach to the problem which is to encapsulate the use of all raw pointers and all raw calls to new and delete in higher-level C++ code. Instead, a set of cooperating template classes developed in the Trilinos package Teuchos are used to encapsulate every use of raw C++ pointers in every use case where it appears in high-level code. Included in the set of memory management classes is the typical reference-counted smart pointer class similar to boost::shared ptr (and therefore C++0x std::shared ptr). However, what is missing in boost and the new standard library are non-reference counted classes for remaining use cases where raw C++ pointers would need to be used. These classes have a debug build mode where nearly all programmer errors are caught and gracefully reported at runtime. The default optimized build mode strips all runtime checks and allows the code to perform as efficiently as raw C++ pointers with reasonable usage. Also included is a novel approach for dealing with the circular references problem that imparts little extra overhead and is almost completely invisible to most of the code (unlike the boost and therefore C++0x approach). Rather than being a radical approach, encapsulating all raw C++ pointers is simply the logical progression of a trend in the C++ development and standards community that started with std::auto ptr and is continued (but not finished) with std::shared ptr in C++0x. Using the Teuchos reference-counted memory management classes allows one to remove unnecessary constraints in the use of objects by removing arbitrary lifetime ordering constraints which are a type of unnecessary coupling [23]. The code one writes with these classes will be more likely to be correct on first writing, will be less likely to contain silent (but deadly) memory usage errors, and will be much more robust to later refactoring and maintenance. The level of debug-mode runtime checking provided by the Teuchos memory management classes is stronger in many respects than what is provided by memory checking tools like Valgrind and Purify while being much less expensive. However, tools like Valgrind and Purify perform a number of types of checks (like usage of uninitialized memory) that makes these tools very valuable and therefore complement the Teuchos memory management debug-mode runtime checking. The Teuchos memory management classes and idioms largely address the technical issues in resolving the fragile built-in C++ memory management model (with the exception of circular references which has no easy solution but can be managed as discussed). All that remains is to teach these classes and idioms and expand their usage in C++ codes. The long-term viability of C++ as a usable and productive language depends on it. Otherwise, if C++ is no safer than C, then is the greater complexity of C++ worth what one gets as extra features? Given that C is smaller and easier to learn than C++ and since most programmers don't know object-orientation (or templates or X, Y, and Z features of C++) all that well anyway, then what really are most programmers getting extra out of C++ that would outweigh the extra complexity of C++ over C? C++ zealots will argue this point but the reality is that C++ popularity has peaked and is becoming less popular while the popularity of C has remained fairly stable over the last decade22. Idioms like are advocated in this paper can help to avert this trend but it will require wide community buy-in and a change in the way C++ is taught in order to have the greatest impact. To make these programs more secure, compiler vendors or static analysis tools (e.g. klocwork23) could implement a preprocessor-like language similar to OpenMP24 that would allow the programmer to declare (in comments) that certain blocks of code should be ''pointer-free'' or allow smaller blocks to be 'pointers allowed'. This would significantly improve the robustness of code that uses the memory management classes described here.

99 GENERAL AND MISCELLANEOUS↗

SuperNeurons: Dynamic GPU Memory Management for Training Deep Neural Networks

Going deeper and wider in neural architectures improves their accuracy, while the limited GPU DRAM places an undesired restriction on the network design domain. Deep Learning (DL) practitioners either need to change to less de- sired network architectures, or nontrivially dissect a network across multiGPUs. These distract DL practitioners from concentrating on their original machine learning tasks. We present SuperNeurons: a dynamic GPU memory scheduling runtime to enable the network training far beyond the GPU DRAM capacity. SuperNeurons features 3 memory optimizations, Liveness Analysis, Unified Tensor Pool, and Cost-Aware Recomputation; together they effectively reduce the network-wide peak memory usage down to the maximal memory usage among layers. We also address the performance issues in these memory-saving techniques. Given the limited GPU DRAM, SuperNeurons not only provisions the necessary memory for the training, but also dynamically allocates the memory for convolution workspaces to achieve the high performance. Evaluations against Caffe, Torch, MXNet and TensorFlow have demonstrated that SuperNeurons trains at least 3.2432 deeper network than current ones with the leading performance. Particularly, SuperNeurons can train ResNet2500 that has 104 basic network layers on a 12GB K40c.

Wang, Linnan↗

Scaling molecular dynamics beyond 100,000 processor cores for large-scale biophysical simulations

The growing interest in the complexity of biological interactions is continuously driving the need to increase system size in biophysical simulations, requiring not only powerful and advanced hardware but adaptable software that can accommodate a large number of atoms interacting through complex force fields. To address this, we developed and implemented strategies in the GENESIS molecular dynamics package designed for large numbers of processors. Long–range electrostatic interactions were parallelized by minimizing the number of processes involved in communication. Here, a novel algorithm was implemented for nonbonded interactions to increase single instruction multiple data (SIMD) performance, reducing memory usage for ultra large systems. Memory usage for neighbor searches in real–space nonbonded interactions was reduced by approximately 80%, leading to significant speedup. Using experimental data describing physical 3D chromatin interactions, we constructed the first atomistic model of an entire gene locus (GATA4). Taken together, these developments enabled the first billion–atom simulation of an intact biomolecular complex, achieving scaling to 65,000 processes (130,000 processor cores) with 1 ns/day performance.

3D modeling↗

Graph-based Reversible Evaluation and Tangents Library

GRETL is a C++ library for evaluation, re-evaluation and algorithmic differentiation of functional operations on an arbitrary computational graph with limited memory usage. Similar to popular machine learning frameworks in Python, like PyTorch and JAX, it tracks and stores both operations and output data as functions are evaluated. Once this composition of functions is built up, the entire chain of operations can be back propagated to compute sensitivities of the final result with respect to any number of inputs. In contrast to most machine learning applications, memory usage becomes the bottleneck for back propagation in many physics applications, especially for time-dependent PDEs. Dynamic check pointing becomes essential. An important distinguishing feature of GRETL is its ability to limit the maximum memory usage by automatically dynamic checkpointing the data output for each graph operation (see Wang, Moin, Iaccarino, 2009). During backpropagation, parts of the graph that are no longer in memory are automatically re-evaluated from upstream checkpointed states as needed for derivative sensitivity calculations (or more precisely, for vector-Jacobian products). GRETL is particularly beneficial for applications, such as coupled multi-physics, where deriving adjoint-based sensitivities and managing checkpoint memory across modules becomes onerous. Cases which can be readily handled by the GRETL library include: different time-integration algorithms per physics (e.g., coupled predictor-corrector algorithms, IMEX, etc.), sub-cycling, asynchronous integrators, state dependent timestep sizes, iterative solvers and coupling algorithms, controller algorithms, and more.

Tupek, MichaelR [Lawrence Livermore National Labor↗

The Memory Scaling of Reverse-Mode Differentiation in Particle Accelerator Simulations with Space Charge

The recent development of differentiable simulation codes for particle accelerators has enabled gradient-based workflows that promise finer control and more realistic modeling of accelerator facilities. However, when using reverse-mode automatic differentiation, the memory usage continuously increases during the simulation, and can potentially exceed the available hardware memory - especially when costly space charge computation is included. To study the memory requirements for differentiable simulations, we have implemented space charge in Cheetah, a PyTorch-based beam tracking code that supports reverse-mode differentiation. We find that the memory usage for reverse-mode differentiation grows linearly with the number of macroparticles and cells, and that it is proportional to the number of space charge kicks involved in the simulation. This general scaling can be used to evaluate whether a given differentiable simulation is feasible given hardware memory constraints.

Dhamrait, Arjun↗

Learning from Five-year Resource-Utilization Data of Titan System

Titan was the flagship supercomputer at the Oak Ridge Leadership Computing Facility (OLCF). It was deployed in late 2012, became the fastest supercomputer in the world and was retired on August 2, 2019. With Titan's mission complete, this paper provides a first-order examination of the usage of its critical resources (CPU, Memory, GPU, and I/O) over a five-year production period (2015-2019). In particular, we show quantitatively that the majority of CPU time was spent on the large-scale jobs, which is consistent with the policy of driving ground-breaking science through leadership computing. We also corroborate the general observation of the low CPU-memory usage with 95% jobs utilizing only 15% or less available memory. Additionally, we correlate the increase of total job submissions and the decrease of GPU-enabled jobs during 2016 with the GPU reliability issue which impacted the large-scale runs. We further show the surprising read/write ratio over the five-year period, which contradicts the general mindset of the large-scale simulation machines being “write-heavy”. This understanding will have potential impact on how we design our next-generation large-scale storage systems. We believe that our analyses and findings are going to be of great interest to the high-performance computing (HPC) community at large.

Wang, Feiyi↗

An efficient hybrid orbital representation for quantum Monte Carlo calculations

The scale and complexity of the quantum system to which real-space quantum Monte Carlo (QMC) can be applied in part depends on the representation and memory usage of the trial wavefunction. B-splines, the computationally most efficient basis set, can have memory requirements exceeding the capacity of a single computational node. This situation has traditionally forced a difficult choice of either using slow internode communication or a potentially less accurate but smaller basis set such as Gaussians. In this paper, we introduce a hybrid representation of the single particle orbitals that combine a localized atomic basis set around atomic cores and B-splines in the interstitial regions to reduce the memory usage while retaining the high speed of evaluation and either retaining or increasing overall accuracy. We present a benchmark calculation for NiO demonstrating a superior accuracy while using only one eighth of the memory required for conventional B-splines. Finally, the hybrid orbital representation therefore expands the overall range of systems that can be practically studied with QMC.

74 ATOMIC AND MOLECULAR PHYSICS↗