Document Details

Document Type : Thesis 
Document Title :
A Novel Deep Learning based Iterative Solver for Large Sparse Linear Equation Systems
تقنيات تخزين جديدة وحلول تكرارية لأنظمة المعادلات الخطية المتناثرة
 
Subject : Faculty of Computing and Information Technology 
Document Language : Arabic 
Abstract : Sparse Matrix-Vector multiplication (SpMV) is one of the key operations in linear algebra that lies at the heart of diverse domains such as scientific computing, engineering, economic modeling, and information retrieval, to name a few. They play a significant role in solving linear system of equations using iterative methods. Sparse Kernels are computational operations on matrices whose entries are mostly zero so that computations with and storage of these zero elements may be eliminated. The emergence of parallel architectures, especially GPU, while offering higher computational performance, has led to the redesign of existing algorithms to suit the architecture. The aim of this thesis is to design novel techniques that improve the performance of sparse Jacobi iterative linear solvers on Nvidia based GPUs. A detailed review of the relevant literature is carried out to identify the challenges and the research gaps. The review revealed that the matrix sparsity structures vary widely based on the application domains and this poses major challenges in obtaining consistent high performance from sparse iterative solvers on Nvidia Tesla K20 GPUs. These challenges include coalesced memory access to the sparse matrix and vector and load balancing among threads and warps. We have developed a deep learning tool that uses an extended set of features to dynamically address these challenges and invoke the most suitable storage format for the iterative solution of sparse linear equation systems. The iterative solver tool has been tested on matrices arising from real world problems. Compared to other leading works, our tool demonstrated 25% or higher performance on average in terms of the execution time and GFLOPS. The contributions of this thesis include a state of the art survey on sparse storage schemes, a deep learning based novel methodology based on an extended set of sparse matrix features and a tool for the iterative solution of sparse linear equation systems. 
Supervisor : Prof. Rashid Mehmood 
Thesis Type : Master Thesis 
Publishing Year : 1438 AH
2017 AD 
Co-Supervisor : Dr. Aiaad Al-Beshri 
Added Date : Monday, June 5, 2017 

Researchers

Researcher Name (Arabic)Researcher Name (English)Researcher TypeDr GradeEmail
طه محمد محمدMohammed, Thaha MohammedResearcherMaster 

Files

File NameTypeDescription
 40852.pdf pdf 

Back To Researches Page