A Memory Consumption Benchmark for MPI Implementations
This report describes the design, implementation, and usage of mpi memu, a memory consumption benchmark for implementations of the Message Passing Interface (MPI).
SEARCH · Search NASA
Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.
Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.
This report describes the design, implementation, and usage of mpi memu, a memory consumption benchmark for implementations of the Message Passing Interface (MPI).
In dislocation dynamics (DD) simulations, the most computationally intensive step is the evaluation of the elastic interaction forces among dislocation ensembles. Because the pair-wise interaction between dislocations is long-range, this force calculation step can be significantly accelerated by the fast multipole method (FMM). In this study, we implemented and compared four different methods in isotropic and anisotropic elastic media: one based on the Taylor series expansion (Taylor FMM), one based on the spherical harmonics expansion (Spherical FMM), one kernel-independent method based on the Chebyshev interpolation (Chebyshev FMM), and a new kernel-independent method that we call the Lagrange FMM. The Taylor FMM is an existing method, used in ParaDiS, one of the most popular DD simulation softwares. The Spherical FMM employs a more compact multipole representation than the Taylor FMM does and is thus more efficient. However, both the Taylor FMM and the Spherical FMM are difficult to derive in anisotropic elastic media because the interaction force is complex and has no closed analytical formula. The Chebyshev FMM requires only being able to evaluate the interaction between dislocations and thus can be applied easily in anisotropic elastic media. But it has a relatively large memory footprint, which limits its usage. The Lagrange FMM was designed to be a memory-efficient black-box method. Lastly, various numerical experiments are presented to demonstrate the convergence and the scalability of the four methods.
Future extreme-scale high-performance computing systems will be required to work under frequent component failures. The MPI Forum s User Level Failure Mitigation proposal has introduced an operation, MPI Comm shrink, to synchronize the alive processes on the list of failed processes, so that applications can continue to execute even in the presence of failures by adopting algorithm-based fault tolerance techniques. This MPI Comm shrink operation requires a failure detection and consensus algorithm. This paper presents three novel failure detection and consensus algorithms using Gossiping. The proposed algorithms were implemented and tested using the Extreme-scale Simulator. The results show that in all algorithms the number of Gossip cycles to achieve global consensus scales logarithmically with system size. The second algorithm also shows better scalability in terms of memory and network bandwidth usage and a perfect synchronization in achieving global consensus. The third approach is a three-phase distributed failure detection and consensus algorithm and provides consistency guarantees even in very large and extreme-scale systems while at the same time being memory and bandwidth efficient.
The newest NERSC supercomputer Cori is a Cray XC40 system consisting of 2,388 Intel Xeon Haswell nodes and 9,688 Intel Xeon-Phi “Knights Landing” (KNL) nodes. Compared to the Xeon-based clusters NERSC users are familiar with, optimal performance on Cori requires consideration of KNL mode settings; process, thread, and memory affinity; fine-grain parallelization; vectorization; and use of the high-bandwidth MCDRAM memory. This paper describes our efforts preparing NERSC users for KNL through the NERSC Exascale Science Application Program, Web documentation, and user training. We discuss how we configured the Cori system for usability and productivity, addressing programming concerns, batch system configurations, and default KNL cluster and memory modes. Here, system usage data, job completion analysis, programming and running jobs issues, and a few successful user stories on KNL are presented.
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.
Future extreme-scale high-performance computing systems will be required to work under frequent component failures. The MPI Forum's User Level Failure Mitigation proposal has introduced an operation, MPI_Comm_shrink, to synchronize the alive processes on the list of failed processes, so that applications can continue to execute even in the presence of failures by adopting algorithm-based fault tolerance techniques. This MPI_Comm_shrink operation requires a fault tolerant failure detection and consensus algorithm. This paper presents and compares two novel failure detection and consensus algorithms. The proposed algorithms are based on Gossip protocols and are inherently fault-tolerant and scalable. The proposed algorithms were implemented and tested using the Extreme-scale Simulator. The results show that in both algorithms the number of Gossip cycles to achieve global consensus scales logarithmically with system size. The second algorithm also shows better scalability in terms of memory and network bandwidth usage and a perfect synchronization in achieving global consensus.
This paper presents a sparsified Fourier neural operator for coupled time-dependent partial differential equations (ST-FNO) as an efficient machine learning surrogate for fluid and particle-based fusion codes such as NIMROD (Non-Ideal Magnetohydrodynamics with Rotation - Open Discussion) and GTC (Gyrokinetic Toroidal Code). ST-FNO leverages the structures in the governing equations and utilizes neural operators to represent Green's function-like numerical operators in the corresponding numerical solvers. Once trained, ST-FNO can rapidly and accurately predict dynamics in fusion devices compared with first-principle numerical algorithms. In general, ST-FNO represents an efficient and accurate machine learning surrogate for numerical simulators for multi-variable nonlinear time-dependent partial differential equations, with the proposed architectures and loss functions. The efficacy of ST-FNO has been demonstrated using quiescent H-mode simulation data from NIMROD and kink-mode simulation data from GTC. The ST-FNO H-mode results show orders of magnitude reduction in memory and central processing unit usage in comparison with the numerical solvers in NIMROD when computing fields over a selected poloidal plane. The ST-FNO kink-mode results achieve a factor of 2 reduction in the number of parameters compared to baseline FNO models without accuracy loss.
GeantV is a complex system based on the interaction of different modules needed for detector simulation, which include transport of particles in fields, physics models simulating their interactions with matter and a geometrical modeler library for describing the detector and locating the particles and computing the path length to the current volume boundary. The GeantV project is recasting the classical simulation approach to get maximum benefit from SIMD/MIMD computational architectures and highly massive parallel systems. This involves finding the appropriate balance between several aspects influencing computational performance (floating-point performance, usage of off-chip memory bandwidth, specification of cache hierarchy, etc.) and handling a large number of program parameters that have to be optimized to achieve the best simulation throughput. This optimization task can be treated as a black-box optimization problem, which requires searching the optimum set of parameters using only point-wise function evaluations. Here, the goal of this study is to provide a mechanism for optimizing complex systems (high energy physics particle transport simulations) with the help of genetic algorithms and evolution strategies as tuning procedures for massive parallel simulations. One of the described approaches is based on introducing a specific multivariate analysis operator that could be used in case of resource expensive or time consuming evaluations of fitness functions, in order to speed-up the convergence of the black-box optimization problem.
Support vector machine (SVM) technology is a growing class of computational algorithms for solving classification problems using supervised learning. There are many freely available SVM codes with implementations of varying granularity in the core optimization task. The performance of SVM implementations is related to four key elements; (1) the size and dimension of the data set on which they are operating, (2) the granularity of their core SVM optimization implementation, (3) the kernel transformation applied to the data, and the (4) underlying hardware on which the implementation is running. To assess the performance of different SVM implementations, several freely available codes representing the spectrum of optimization granularity were built on a variety of hardware—two linux clusters, and a shared memory machine. Binary classifiers were trained on datasets of varying size containing anywhere from a few hundred to over 50,000 vectors. Performance of the methods was measured in terms of wall-clock time for training, statistical quality of the resulting classifier, hardware performance (memory footprint, memory bandwidth, floating point unit usage, I/O bandwidth, and instruction stalls), robustness and portability.
A scalable density functional electronic code with Gaussian basis set, called UTEP-NRLMOL, is developed to perform simulations of molecular systems in the presence of the environment with particular attention to the memory requirements. In the electronic structure calculations, the memory and computation time are proportional to the number of atoms. Memory requirements for density functional calculations scale as N*N, where N is the number of atoms. While the recent advances in HPC offer platforms with large numbers of cores, the limited amount of memory available on a given node and poor scalability of the electronic structure codes hinder their efficient usage of these platforms. We have introduced new scaling and parallelization paradigms using MPI-3 shared-memory functionality combined with usage of sparse algebra and storage of matrices in sparse format. This extends the range of applicability of the UTEP-NRLMOL code to large systems over 10,000 atoms, or using up to 67,000 basis functions, and making use of HPC architectures using over 6,000 processors utilizing all available cores. We have also interfaced code with effective fragment potential and polarizable continuum model libraries. The code was used in simulations of several applications which are published in reputed scientific journals.
RMEM - a prototyping platform to explore the usage of RMA-based strategies in GPU-to-GPU communication. The implementation relies on libfabric (BSD-2 or GNUv2) to explore different synchronization approach and their associated performance. It is compatible with different network technologies such as InfiniBand, SlingShot-11, and TCP. It provides to the user host-to-host and device-to-device communication. The latter can be initiated either from the host, or from the device.
A system may provide virtual energy audits of one or more target buildings. The system may retrieve weather data and energy usage data specific to a given target building from a weather server and a utility server, respectively. The system may store predefined building characteristics corresponding to the given target building in local memory. Based on the weather data, energy usage data, and/or predefined building characteristics, the system may generate one or more building markers that characterize the energy usage and efficiency of the given target building. Building efficiency diagnostics and energy conservation prognostics may be generated based on the building markers and may be sent by the system to be displayed via a user interface of a client device. The energy conservation prognostics may include one or more energy conservation measure recommendations and corresponding predicted cost/energy savings.
Resource utilization statistics of submitted jobs on a supercomputer can help us understand how users from various scientific domains use HPC platforms and better design a job scheduler. We explore to generate insight regarding workload distribution and usage pattern domains from job scheduler trace, GPU failure information, and project-specific information collected from Titan supercomputer. Furthermore, we want to know how the scheduler performance varies over time and how the users' scheduling behavior changes following a system failure. These observations have the potential to provide valuable insight, which is helpful to prepare for system failures. These practices will help us develop and apply novel machine learning algorithms in understanding system behavior, requirement, and better scheduling of HPC systems. There are two datasets, RUR and GPU. RUR: This dataset is the job scheduler traces collected from the Titan supercomputerfrom 01/01/2015 to 07/31/2019 (2015.csv - 2019.csv). These were collected usingResource Utilization Report (RUR), a Cray-developed resource-usage data collectionand reporting system. It contains the usage information of its critical resources (CPU,Memory, GPU, and I/O) of each running job on Titan during that period [2]. ProjectAreas: Every job is associated with a project ID. TheProjectAreas.csvdatasetprovides a mapping of the project ID to its domain science. GPU: There have been some hardware-related issues in the GPUs in Titan that caused some GPUs to fail, sometimes irrecoverably during some job runs. This dataset provides information regarding these failures during the execution of the submitted jobs. GPUs on Titan are uniquely identified by a serial number (SN), and they are installed in a location. A GPU can be installed in a location, then removed from that location following a failure, and then re-installed in a different location after fixing the problem. If the failure can't be recovered, the GPU might be removed entirely from Titan. There are two prominent types of failures that resulted in the removal of GPUs from Titan: Double Bit Error (DBE) and Out of the Bus (OTB). The dataset (gc_full.csv) has the following fields: 1. SN : Serial number of a GPU 2. location : The location where it is installed 3. insert : The time when it was inserted into that location 4. remove : The time when it was removed from that location 5. duration : Amount of time the GPU spent in this location 6. out : If the device was taken out entirely w/o a re-installment into a new location. 7. event : If the GPU was taken out entirely, the reason for its removal. To learn more about this dataset, please refer to the git repositoryhttps://github.com/olcf/TitanGPULifeand the related publication [1]. References [1] George Ostrouchov, Don Maxwell, Rizwan A Ashraf, Christian Engelmann, MallikarjunShankar, and James H Rogers. Gpu lifetimes on titan supercomputer: Survival analysisand reliability. InSC20: International Conference for High Performance Computing,Networking, Storage and Analysis, pages 1-14. IEEE, 2020. [2] Feiyi Wang, Sarp Oral, Satyabrata Sen, and Neena Imam. Learning from five-yearresource-utilization data of titan system. In2019 IEEE International Conference onCluster Computing (CLUSTER), pages 1-6. IEEE, 2019.
Resource utilization statistics of submitted jobs on a supercomputer can help us understand how users from various scientific domains use HPC platforms and better design a job scheduler. We explore to generate insight regarding workload distribution and usage pattern domains from job scheduler trace, GPU failure information, and project-specific information collected from Titan supercomputer. Furthermore, we want to know how the scheduler performance varies over time and how the users' scheduling behavior changes following a system failure. These observations have the potential to provide valuable insight, which is helpful to prepare for system failures. These practices will help us develop and apply novel machine learning algorithms in understanding system behavior, requirement, and better scheduling of HPC systems. There are two datasets, RUR and GPU: RUR dataset is the job scheduler traces collected from the Titan supercomputer from 01/01/2015 to 07/31/2019 (2015.csv - 2019.csv). These were collected using resource Utilization Report (RUR), a Cray-developed resource-usage data collection and reporting system. It contains the usage information of its critical resources (CPU, Memory, GPU, and I/O) of each running job on Titan during that period (https://ieeexplore.ieee.org/abstract/document/8891001). It includes ProjectAreas as additional information, every job is associated with a project ID. TheProjectAreas.csv dataset provides a mapping of the project ID to its domain science. GPU dataset has information regarding GPU failure on Titan. There have been some hardware-related issues in the GPUs in Titan that caused some GPUs to fail, sometimes irrecoverably during some job runs. This dataset provides information regarding these failures during the execution of the submitted jobs. GPUs on Titan are uniquely identified by a serial number (SN), and they are installed in a location. A GPU can be installed in a location, then removed from that location following a failure, and then re-installed in a different location after fixing the problem. If the failure can't be recovered, the GPU might be removed entirely from Titan. There are two prominent types of failures that resulted in the removal of GPUs from Titan: Double Bit Error (DBE) and Out of the Bus (OTB). The dataset (gc_full.csv) has seven attributes, we provided a short description of these attributes in the ReadMe file. To learn more about this dataset, please refer to the git repository https://github.com/olcf/TitanGPULife and the related publication (https://ieeexplore.ieee.org/abstract/document/9355319).
Systems, apparatuses, and methods for managing buffers in a neural network implementation with heterogeneous memory are disclosed. A system includes a neural network coupled to a first memory and a second memory. The first memory is a relatively low-capacity, high-bandwidth memory while the second memory is a relatively high-capacity, low-bandwidth memory. During a forward propagation pass of the neural network, a run-time manager monitors the usage of the buffers for the various layers of the neural network. During a backward propagation pass of the neural network, the run-time manager determines how to move the buffers between the first and second memories based on the monitored buffer usage during the forward propagation pass. As a result, the run-time manager is able to reduce memory access latency for the layers of the neural network during the backward propagation pass.
Energy consumption is one of the top challenges for achieving the next generation of supercomputing. Codesign of hardware and software is critical for improving energy efficiency (EE) for future large-scale systems. Many architectural power-saving techniques have been developed, and most hardware components are approaching physical limits. Accordingly, parallel computing software, including both applications and systems, should exploit power-saving hardware innovations and manage efficient energy use. In addition, new power-aware parallel computing methods are essential to decrease energy usage further. This article surveys software-based methods that aim to improve EE for parallel computing. It reviews the methods that exploit the characteristics of parallel scientific applications, including load imbalance and mixed precision of floating-point (FP) calculations, to improve EE. In addition, this article summarizes widely used methods to improve power usage at different granularities, such as the whole system and per application. In particular, it describes the most important techniques to measure and to achieve energy-efficient usage of various parallel computing facilities, including processors, memories, and networks. Overall, this article reviews the state-of-the-art of energy-efficient methods for parallel computing to motivate researchers to achieve optimal parallel computing under a power budget constraint.
Despite state-of-the-art technologies like artificial intelligence, human judgment is critically essential in cooperative systems, such as the multi-agent system (MAS), which collect information among agents based on multiple-cue judgment. Human agents can prevent impaired situational awareness of automated agents by confirming situations under environmental uncertainty. System error caused by uncertainty can result in an unreliable system environment, and this environment affects the human agent, resulting in non-optimal decision-making in MAS. Thus, it is necessary to know how human behavior is changed to capture system reliability under uncertainty. Another issue affecting MAS is time delay, which can delay agent information transfer, resulting in low performance and instability. However, it is difficult to find studies on the influence of time delay on human agents. This study is about understanding the human decision-making process under a specific system reliability environment by uncertainty with time delay. We used concepts of expected and unexpected uncertainty to implement reliability of the system usage environment with three types of time delay conditions: no time delay, regular time delay, and irregular time delay conditions. We used electroencephalogram (EEG) for human cognitive neural mechanisms in multiple-cue judgment systems to understand human decision-making. In the reliability of system usage environment, the unreliable system environment significantly creates less memory load by less utilization of system rules for decision-making. In terms of time delay, delayed information delivery does not significantly affect memory load for decision-making.
In the age of massive parallelism, the focus of performance analysis has switched from the processor and related structures to the memory and I/O resources. Adapting to this new reality, a performance analysis tool has to provide a way to analyze resource usage to pinpoint existing and potential problems in a given application. This paper provides an overview of the Memory Observant Data Analysis (MODA) tool, a memory-centric tool first implemented on the Cray XMT supercomputer. Throughout the paper, MODA's capabilities have been showcased with experiments done on matrix multiply and Graph-500 application codes.