DocumentCode :
3124213
Title :
Fast and scalable parallel algorithms for matrix chain product and matrix powers on distributed memory systems
Author :
Li, Keqin
Author_Institution :
Dept. of Comput. Sci., State Univ. of New York, New Paltz, NY, USA
fYear :
2001
fDate :
36982
Abstract :
Given N matrices A1, A2,…, AN of size N×N, the matrix chain product problem is to compute A1×A2×···AN . Given an N×N matrix A, the matrix powers problem is to calculate the first N powers of A, i.e., A, A2A3,…, AN. We consider distributed memory systems (DMS) with p processors that can support one-to-one communications in O(T(p)) time. Assume that the time complexity of the best known sequential algorithm for matrix multiplication is O(Nα), where α<2.3755. Let p be arbitrarily chosen in the range 1⩽p⩽Nα+1/log N. We show that the two problems can be solved on a p-processor DMS in T chain(N,p)=O(Nα+1/p+T(p)(N 2(1+1/α)/p2/ α(log p/N)1-2α/+log(p log N/Nα) log N)) and Tpower(N,p)=0(Nα+1/p+T(p)(N 2(1+1/α)/p2/ α(log p/log N) 1-2α/+(log N)2)) times, respectively. We also give instantiation of the above results in distributed memory parallel computers and DMS with hypercubic networks, and show that our parallel algorithms are either fully scalable or highly scalable
Keywords :
computational complexity; distributed memory systems; hypercube networks; matrix multiplication; parallel algorithms; distributed memory parallel computers; distributed memory systems; hypercubic networks; matrices; matrix chain product; matrix multiplication; matrix powers; one-to-one communications; scalable parallel algorithms; sequential algorithm; time complexity; Computer networks; Computer science; Concurrent computing; Distributed computing; Equations; Graph theory; High performance computing; Parallel algorithms; Phase change random access memory; Polynomials;
fLanguage :
English
Publisher :
ieee
Conference_Titel :
Parallel and Distributed Processing Symposium., Proceedings 15th International
Conference_Location :
San Francisco, CA
ISSN :
1530-2075
Print_ISBN :
0-7695-0990-8
Type :
conf
DOI :
10.1109/IPDPS.2001.924937
Filename :
924937
Link To Document :
بازگشت