Document Details
|
Document Type |
: |
Thesis |
|
Document Title |
: |
MACHINE LEARNING BASED PERFORMANCE OPTIMIZATION OF SPMV KERNELS FOR DISTRIBUTED MEMORY EXASCALE SYSTEM تعلم الآلة القائم على الأداء الامثل لنواة (SpMV) للذاكرة الموزعة في نظم إكساسكال |
|
Subject |
: |
Faculty of Computing and Information Technology |
|
Document Language |
: |
Arabic |
|
Abstract |
: |
Numerous scientific, engineering, economic and social applications require the solution of sparse linear equation systems, and these applications are increasingly being used to develop timely intelligence for the design, operations, and management of smart cities and societies. Sparse matrix-vector product (SpMV) is the most important and time-consuming kernel for the iterative solution of sparse linear equation systems. High performance computing (HPC) typically exploits parallel computing features of the underlying software and hardware infrastructure to solve large problems faster. HPC has been applied to SpMV and linear algebra, and other problems for several decades. Big data and data-driven approaches have been used relatively recently in scientific computing to address HPC related challenges, and this has given rise to the convergence of HPC and big data. Moreover, artificial intelligence (AI) is increasingly being used to improve big data, HPC, scientific computing, and other problem domains. This trend has given rise to the convergence of big data, HPC and AI. This thesis aims to contribute to this convergence, as a step towards and beyond exascale computing, and applies it (the convergence of the three areas) to the area of SpMV computations.
Several factors affect the performance of SpMV computations. These include matrix characteristics, storage formats, software implementations, and hardware platforms. Performance optimization of an application on multicore architectures is challenging due to the heterogeneity and diversity of architectures. Modern machines have a range of shared and distributed memory, and hybrid architectures, with several hierarchies involving non-uniform communication latencies. The sparsity pattern of the matrix affects the performance of SpMV computations, particularly in case of distributed memory implementations, resulting in load imbalance that causes both computation and communication overheads. The manual process of trial and error experimentation to find the best number of processes to compute SpMV is time-consuming. This thesis proposes ZAKI and ZAKI+, two data-driven and machine-learning approaches to predict the optimal number of process configurations and optimal parallelization mapping strategies for SpMV computations of an arbitrary sparse matrix on a distributed memory machine. Four machine learning algorithms are used for predictions -- Decision Trees, Random Forest, Xtreme and Gradient Boosting -- and have been evaluated using 1838 real-world sparse matrices associated with 45 application domains.
Data locality is a major challenge in delivering exascale computing due to the high energy cost of data movement. This thesis also contributes the first literature review on data locality for converged (HPC, big data, and AI) exascale systems. An architecture based on design patterns for convergence of HPC and big data is also provided. Moreover, a preliminary investigation and future directions are provided on the applicability of our proposed tool to energy efficiency optimization. This thesis has contributed a total of seven SCOPUS and/or ISI-indexed papers. To the best of our knowledge, this thesis is the first attempt, where we have exploited the structure of the matrices to predict the optimal number of processes and best mapping strategy for a given matrix in distributed memory environment by using different base and ensemble machine learning methods. |
|
Supervisor |
: |
Dr. Iyad Katib |
|
Thesis Type |
: |
Doctorate Thesis |
|
Publishing Year |
: |
1440 AH
2019 AD |
|
Added Date |
: |
Monday, May 20, 2019 |
|
Researchers
| سردار عثمان | Usman, Sardar | Researcher | Doctorate | |
|
Back To Researches Page
|