Recently, Deep Neural Networks (DNNs) have recorded great success in handling medical and other complex classification tasks. However, as the sizes of a DNN model and the available dataset increase, the training process becomes more complex and computationally intensive, which usually takes a longer time to complete. In this work, we have proposed a generic full end-to-end hybrid parallelization approach combining both model and data parallelism for efficiently distributed and scalable training of DNN models. We have also proposed a Genetic Algorithm based heuristic resources allocation mechanism (GABRA) for optimal distribution of partitions on the available GPUs for computing performance optimization. We have applied our proposed approach to a real use case based on 3D Residual Attention Deep Neural Network (3D-ResAttNet) for efficient Alzheimer Disease (AD) diagnosis on multiple GPUs. The experimental evaluation shows that the proposed approach is efficient and scalable, which achieves almost linear speedup with little or no differences in accuracy performance when compared with the existing non-parallel DNN models.
We modified a MPI-friendly density functional theory (DFT) source code within hybrid parallelization including CUDA. Our objective is to find out how simple conversions within the hybrid parallelization with mid-range GPUs affect DFT code not originally suitable to CUDA. We settled several rules of hybrid parallelization for numerical-atomic-orbital (NAO) DFT codes. The test was performed on a magnetite material system with OpenMX code by utilizing a hardware system containing 2 Xeon E5606 CPUs and 2 Quadro 4000 GPUs. 3-way hybrid routines obtained a speedup of 7.55 while 2-way hybrid speedup …
In this paper, we study various parallelization schemes for the Variable Neighborhood Search (VNS) metaheuristic on a CPU-GPU system via OpenMP and OpenACC. A hybrid parallel VNS method is applied to recent benchmark problem instances for the multi-product dynamic lot sizing problem with product returns and recovery, which appears in reverse logistics and is known to be NP-hard. We report our findings regarding these parallelization approaches and present promising computational results.
We present GSPMD, an automatic, compiler-based parallelization system for common machine learning computations. It allows users to write programs in the same way as for a single device, then give hints through a few annotations on how to distribute tensors, based on which GSPMD will parallelize the computation. Its representation of partitioning is simple yet general, allowing it to express different or mixed paradigms of parallelism on a wide variety of models. GSPMD infers the partitioning for every operator based on limited user annotations, making it convenient to scale existing single-dev…
We present our experience with the modernization on the GR-MHD code BHAC, aimed at improving its novel hybrid (MPI+OpenMP) parallelization scheme. In doing so, we showcase the use of performance profiling tools usable on x86 (Intel-based) architectures. Our performance characterization and threading analysis provided guidance in improving the concurrency and thus the efficiency of the OpenMP parallel regions. We assess scaling and communication patterns in order to identify and alleviate MPI bottlenecks, with both runtime switches and precise code interventions. The performance of optimized ve…
Modern OpenMP threading techniques are used to convert the MPI-only Hartree-Fock code in the GAMESS program to a hybrid MPI/OpenMP algorithm. Two separate implementations that differ by the sharing or replication of key data structures among threads are considered, density and Fock matrices. All implementations are benchmarked on a super-computer of 3,000 Intel Xeon Phi processors. With 64 cores per processor, scaling numbers are reported on up to 192,000 cores. The hybrid MPI/OpenMP implementation reduces the memory footprint by approximately 200 times compared to the legacy code. The MPI/Ope…
Deep learning systems have become vital tools across many fields, but the increasing model sizes mean that training must be accelerated to maintain such systems' utility. Current systems like Tensorflow and MXNet focus on one specific parallelization strategy, data parallelism, which requires large training batch sizes in order to scale. We cast the problem of finding the best parallelization strategy as the problem of finding the best tiling to partition tensors with the least overall communication. We propose an algorithm that can find the optimal tiling. Our resulting parallelization soluti…
5G Radio access network disaggregation and softwarization pose challenges in terms of computational performance to the processing units. At the physical layer level, the baseband processing computational effort is typically offloaded to specialized hardware accelerators. However, the trend toward software-defined radio-access networks demands flexible, programmable architectures. In this paper, we explore the software design, parallelization and optimization of the key kernels of the lower physical layer (PHY) for physical uplink shared channel (PUSCH) reception on MemPool and TeraPool, two ma…
Comprehensive simulations of the entire biochemistry of cells have great potential to help physicians treat disease and help engineers design biological machines. But such simulations must model networks of millions of molecular species and reactions. The Stochastic Simulation Algorithm (SSA) is widely used for simulating biochemistry, especially systems with species populations small enough that discreteness and stochasticity play important roles. However, existing serial SSA methods are prohibitively slow for comprehensive networks, and existing parallel SSA methods, which use periodic synch…
Over the last two decades, scientific workflow management systems (SWfMS) have emerged as a means to facilitate the design, execution, and monitoring of reusable scientific data processing pipelines. At the same time, the amounts of data generated in various areas of science outpaced enhancements in computational power and storage capabilities. This is especially true for the life sciences, where new technologies increased the sequencing throughput from kilobytes to terabytes per day. This trend requires current SWfMS to adapt: Native support for parallel workflow execution must be provided to…
Similarity search is critical for many database applications, including the increasingly popular online services for Content-Based Multimedia Retrieval (CBMR). These services, which include image search engines, must handle an overwhelming volume of data, while keeping low response times. Thus, scalability is imperative for similarity search in Web-scale applications, but most existing methods are sequential and target shared-memory machines. Here we address these issues with a distributed, efficient, and scalable index based on Locality-Sensitive Hashing (LSH). LSH is one of the most efficien…
In this work we use the GPU porting task for the operative Japanese weather prediction model "ASUCA" as an opportunity to examine productivity issues with OpenACC when applied to structured grid problems. We then propose "Hybrid Fortran", an approach that combines the advantages of directive based methods (no rewrite of existing code necessary) with that of stencil DSLs (memory layout is abstracted). This gives the ability to define multiple parallelizations with different granularities in the same code. Without compromising on performance, this approach enables a major reduction in the code c…
We propose a new hybrid topology optimization algorithm based on multigrid approach that combines the parallelization strategy of CPU using OpenMP and heavily multithreading capabilities of modern Graphics Processing Units (GPU). In addition to that significant computational efficiency in memory requirement has been achieved using homogenization strategy. The algorithm has been integrated with versitile computing platform of MATLAB for ease of use and customization. The bottlenecking repetitive solution of the state equation has been solved using an optimized geometric multigrid approach along…
University of California Los Angeles（加州大学洛杉矶分校）；NVIDIA（英伟达）
Deploying deep learning (DL) models across multiple compute devices to train large and complex models continues to grow in importance because of the demand for faster and more frequent training. Data parallelism (DP) is the most widely used parallelization strategy, but as the number of devices in data parallel training grows, so does the communication overhead between devices. Additionally, a larger aggregate batch size per step leads to statistical efficiency loss, i.e., a larger number of epochs are required to converge to a desired accuracy. These factors affect overall training time and b…
The implementation of a vast majority of machine learning (ML) algorithms boils down to solving a numerical optimization problem. In this context, Stochastic Gradient Descent (SGD) methods have long proven to provide good results, both in terms of convergence and accuracy. Recently, several parallelization approaches have been proposed in order to scale SGD to solve very large ML problems. At their core, most of these approaches are following a map-reduce scheme. This paper presents a novel parallel updating algorithm for SGD, which utilizes the asynchronous single-sided communication paradigm…
The main computing tasks of a finite element code(FE) for solving partial differential equations (PDE's) are the algebraic system assembly and the iterative solver. This work focuses on the first task, in the context of a hybrid MPI+X paradigm. Although we will describe algorithms in the FE context, a similar strategy can be straightforwardly applied to other discretization methods, like the finite volume method. The matrix assembly consists of a loop over the elements of the MPI partition to compute element matrices and right-hand sides and their assemblies in the local system to each MPI par…
Edge-centric distributed computations have appeared as a recent technique to improve the shortcomings of think-like-a-vertex algorithms on large scale-free networks. In order to increase parallelism on this model, edge partitioning - partitioning edges into roughly equally sized blocks - has emerged as an alternative to traditional (node-based) graph partitioning. In this work, we give a distributed memory parallel algorithm to compute high-quality edge partitions in a scalable way. Our algorithm scales to networks with billions of edges, and runs efficiently on thousands of PEs. Our technique…
Center of Computation & Technology（计算与技术中心）；Louisiana State University（路易斯安那州立大学）；Microsoft Research Montréal（微软研究院蒙特利尔分院）
Although recent scaling up approaches to training deep neural networks have proven to be effective, the computational intensity of large and complex models, as well as the availability of large-scale datasets, require deep learning frameworks to utilize scaling out techniques. Parallelization approaches and distribution requirements are not considered in the preliminary designs of most available distributed deep learning frameworks, and most of them still are not able to perform effective and efficient fine-grained inter-node communication. We present Phylanx that has the potential to alleviat…
This paper describes a new fast and implicitly parallel approach to neighbour-finding in multi-resolution Smoothed Particle Hydrodynamics (SPH) simulations. This new approach is based on hierarchical cell decompositions and sorted interactions, within a task-based formulation. It is shown to be faster than traditional tree-based codes, and to scale better than domain decomposition-based approaches on hybrid shared/distributed-memory parallel architectures, e.g. clusters of multi-cores, achieving a $40\times$ speedup over the Gadget-2 simulation code.
In this article, a new generic higher-order finite-element framework for massively parallel simulations is presented. The modular software architecture is carefully designed to exploit the resources of modern and future supercomputers. Combining an unstructured topology with structured grid refinement facilitates high geometric adaptability and matrix-free multigrid implementations with excellent performance. Different abstraction levels and fully distributed data structures additionally ensure high flexibility, extensibility, and scalability. The software concepts support sophisticated load b…
Scalable Runtime Architecture for Data-driven, Hybrid HPC and ML Workflow Applications
Rutgers, the State University of New Jersey（罗格斯大学（新泽西州立大学））；Brookhaven National Laboratory（布鲁克海文国家实验室）；Princeton Plasma Physics Laboratory（普林斯顿等离子体物理实验室）；Princeton University（普林斯顿大学）
本文扩展 RADICAL-Pilot，提出支持基于服务执行的可扩展运行时架构，以实现跨本地与远程 HPC/云资源的 HPC/ML 无缝耦合和低开销并发 ML 执行。
We have witnessed a boosted demand for graph analytics at Twitter in recent years, and graph analytics has become one of the key parts of Twitter's large-scale data analytics and machine learning for driving engagement, serving the most relevant content, and promoting healthier conversations. However, infrastructure for graph analytics has historically not been an area of investment at Twitter, resulting in a long timeline and huge engineering effort for each project to deal with graphs at the Twitter scale. How do we build a unified graph analytics user experience to fulfill modern data analy…
Processing large complex networks recently attracted considerable interest. Complex graphs are useful in a wide range of applications from technological networks to biological systems like the human brain. Sometimes these networks are composed of billions of entities that give rise to emerging properties and structures. Analyzing these structures aids us in gaining new insights about our surroundings. As huge networks become abundant, there is a need for scalable algorithms to perform analysis. A prominent example is the PageRank algorithm, which is one of the measures used by web search engin…
To reduce training time of large-scale DNNs, scientists have started to explore parallelization strategies like data-parallelism, model-parallelism, and hybrid-parallelism. While data-parallelism has been extensively studied and developed, several problems exist in realizing model-parallelism and hybrid-parallelism efficiently. Four major problems we focus on are: 1) defining a notion of a distributed model across processes, 2) implementing forward/back-propagation across process boundaries that requires explicit communication, 3) obtaining parallel speedup on an inherently sequential task, an…
