With the commencement of the exascale computing era, we realize that the majority of the leadership supercomputers are heterogeneous and massively parallel even on a single node with multiple co-processors such as GPUs and multiple cores on each node. For example, ORNLs Summit accumulates six NVIDIA Tesla V100s and 42 core IBM Power9s on each node. Synchronizing across all these compute resources in a single node or even across multiple nodes is prohibitively expensive. Hence it is necessary to develop and study asynchronous algorithms that circumvent this issue of bulk-synchronous computing for massive parallelism. In this study, we examine the asynchronous version of the abstract Restricted Additive Schwarz method as a solver where we do not explicitly synchronize, but allow for communication of the data between the sub-domains to be completely asynchronous thereby removing the bulk synchronous nature of the algorithm.
We accomplish this by using the onesided RMA functions of the MPI standard. We study the benefits of using such an asynchronous solver over its synchronous counterpart on both multi-core architectures and on multiple GPUs. We also study the communication patterns and local solvers and their effect on the global solver. Finally, we show that this concept can render attractive runtime benefits over the synchronous counterparts.
For many, Graphics Processing Units (GPUs) provides a source of reliable computing power. Recently, Nvidia introduced its 9th generation HPC-grade GPUs, the Ampere 100, claiming significant performance improvements over previous generations, particularly for AI-workloads, as well as introducing new architectural features such as asynchronous data movement. But how well does the A100 perform on non-AI benchmarks, and can we expect the A100 to deliver the application improvements we have grown used to with previous GPU generations? In this paper, we benchmark the A100 GPU and compare it to four …
DisagFusion: Asynchronous Pipeline Parallelism and Elastic Scheduling for Disaggregated Diffusion Serving
Renmin University of China（中国人民大学）；SenseTime（商汤科技）；Peking University（北京大学）；Hong Kong University of Science and Technology（香港科技大学）；Beihang University（北京航空航天大学）
In this paper, we present GossipGraD - a gossip communication protocol based Stochastic Gradient Descent (SGD) algorithm for scaling Deep Learning (DL) algorithms on large-scale systems. The salient features of GossipGraD are: 1) reduction in overall communication complexity from Θ(log(p)) for p compute nodes in well-studied SGD to O(1), 2) model diffusion such that compute nodes exchange their updates (gradients) indirectly after every log(p) steps, 3) rotation of communication partners for facilitating direct diffusion of gradients, 4) asynchronous distributed shuffle of samples during the f…
Robustness of deep learning classification to adversarial input on GPUs: asynchronous parallel accumulation is a source of vulnerability
Maxeler Technologies, a Groq Company（Maxeler Technologies，Groq公司）；ETH Zurich / CSCS（苏黎世联邦理工学院 / CSCS）；Georgia Institute of Technology（佐治亚理工学院）；Oak Ridge National Laboratory（橡树岭国家实验室）
iPIC3D is a widely used massively parallel Particle-in-Cell code for the simulation of space plasmas. However, its current implementation does not support execution on multiple GPUs. In this paper, we describe the porting of iPIC3D particle mover to GPUs and the optimization steps to increase the performance and parallel scaling on multiple GPUs. We analyze the strong scaling of the mover on two GPU clusters and evaluate its performance and acceleration. The optimized GPU version which uses pinned memory and asynchronous data prefetching outperform their corresponding CPU versions by 5-10x on …
University of Guelph（圭尔夫大学）；SHARCNET, Compute Canada（加拿大计算公司SHARCNET）
We develop a scalable and extendable training framework that can utilize GPUs across nodes in a cluster and accelerate the training of deep learning models based on data parallelism. Both synchronous and asynchronous training are implemented in our framework, where parameter exchange among GPUs is based on CUDA-aware MPI. In this report, we analyze the convergence and capability of the framework to reduce training time when scaling the synchronous training of AlexNet and GoogLeNet from 2 GPUs to 8 GPUs. In addition, we explore novel ways to reduce the communication overhead caused by exchangin…
GPUs are readily available in cloud computing and personal devices, but their use for data processing acceleration has been slowed down by their limited integration with common programming languages such as Python or Java. Moreover, using GPUs to their full capabilities requires expert knowledge of asynchronous programming. In this work, we present a novel GPU run time scheduler for multi-task GPU computations that transparently provides asynchronous execution, space-sharing, and transfer-computation overlap without requiring in advance any information about the program dependency structure. W…
This paper is focused on improving multi-GPU performance of a research CFD code on structured grids. MPI and OpenACC directives are used to scale the code up to 16 GPUs. This paper shows that using 16 P100 GPUs and 16 V100 GPUs can be 30$\times$ and 70$\times$ faster than 16 Xeon CPU E5-2680v4 cores for three different test cases, respectively. A series of performance issues related to the scaling for the multi-block CFD code are addressed by applying various optimizations. Performance optimizations such as the pack/unpack message method, removing temporary arrays as arguments to procedure cal…
Modern machine learning frameworks can train neural networks using multiple nodes in parallel, each computing parameter updates with stochastic gradient descent (SGD) and sharing them asynchronously through a central parameter server. Due to communication overhead and bottlenecks, the total throughput of SGD updates in a cluster scales sublinearly, saturating as the number of nodes increases. In this paper, we present a solution to predicting training throughput from profiling traces collected from a single-node configuration. Our approach is able to model the interaction of multiple nodes and…
Evaluating OpenMP Offloading for Intra-node Multi-GPU Programming across NVIDIA, AMD, and Intel Architectures: A 3D Heat Transfer Case Study
Rudolfovo-Science and Technology Centre Novo mesto（鲁多尔沃夫诺沃梅斯托科学与技术中心）；Faculty of Mechanical Engineering, University of Ljubljana（卢布尔雅那大学机械工程学院）；Department of Computer Science, FSTM, University of Luxembourg（卢森堡大学科学技术学院计算机科学系）
