Bilkent University
Department of Computer Engineering
S E M I N A R
A Directed-graph Model for Minimising Communication Requirement in Parallel
F. Şükrü Torun
MSc. Student
Computer Engineering Department
Bilkent UniversityMatrix-vector multiplication is an essential operation in many algorithms used in scientific applications. Matrix-vector multiplication requires an extensive amount of computation time. Due to high computational throughput requirement, matrix-vector multiplication is performed in a parallel fashion. Depending on the matrix and its decomposition, parallel matrix vector multiplication necessitates communication among processors in the parallel environment in order to calculate result. Here, we present a directed-graph-based model for minimising communication requirement in parallel matrix-vector multiplication with replication through one dimensional partitioning of the square sparse matrices and we propose an algorithm that shows how the proposed model, which is directed-graph-based model for replicated row parallel matrix vector multiplication, reduces the execution time of multiplication by minimising communication cost with replication.
DATE: 20 December, 2010, Monday @ 15:20
PLACE: EA409