Document Details
|
Document Type |
: |
Thesis |
|
Document Title |
: |
JACOBI ITERATIVE METHOD AND SPMV COMPUTATIONS ON GPUS طريقة جاكوبي التكرارية و حسابات SpMV في وحدة معالجة الرسومات |
|
Subject |
: |
Faculty of Computing and Information Technology |
|
Document Language |
: |
Arabic |
|
Abstract |
: |
Graphics processing units (GPUs) have delivered remarkable performance for many high performance computing (HPC) applications through massive parallelism that they provide. Iterative solution of large linear equation systems that forms an essential building block for numerous scientific and engineering applications is one such application. However, obtaining performance for Sparse Matrix-Vector (SpMV) multiplication that is the core kernel of iterative solutions of linear equation systems has remained a challenge due to the varying sparsity patterns of the nonzero entries in the linear equation systems. A number of schemes have been proposed to improve SpMV storage and computation performance. However, these schemes are mainly evaluated and compared in terms of the SpMV throughput in FLOPS, which alone does not provide a deep insight into the SpMV storage and computations.
The aim of this thesis is to analyze, understand, and improve the performance of Jacobi iterative method and SpMV computations on GPU architectures. Towards this aim, we provide a detailed study of four notable schemes (CSR, ELL, HYB, and CSR5). The schemes are evaluated individually and compared using eight different performance metrics including execution time, GFLOPS, achieved occupancy, instructions per warp, warp execution efficiency, global memory throughput, global memory efficiency, and global memory transactions. Subsequently, using the deeper insights into the performance gaps of the current schemes, gained through the detailed performance analysis of the schemes, we propose a novel SpMV computations scheme called the heterogeneous CPU-GPU Hybrid (HCGHYB) scheme. HCGHYB provides improved performance over the HYB scheme, which is a popular choice in many SpMV open source and commercial iterative solvers. Finally, we give directions for further improvements in SpMV computations schemes. |
|
Supervisor |
: |
Dr. Human clinic |
|
Thesis Type |
: |
Master Thesis |
|
Publishing Year |
: |
1440 AH
2019 AD |
|
Added Date |
: |
Monday, August 19, 2019 |
|
Researchers
| ساره حامد الأحمدي | Al-Ahmadi, Sarah Hamed | Researcher | Master | |
|
Back To Researches Page
|