Efficient Task Partitioning Algorithms for Distributed Shared Memory Systems

In this paper, we consider the tree task graphs which arise from many important programming paradigms such as divide and conquer, branch and bound etc., and the linear task-graphs that stem from common computation schemes such as pipelining, iterative calculation etc.. The target architecture considered is a distributed shared memory architecture with indirect network or wormhole-routed direct network as its interconnection network (IN). The partitioning problem is formulated for systems where all or most of the communications are interleaved with task executions, in light of latency-hiding techniques employed in most of today's multiprocessors. This is in contrast with previous research in this area where computation and communication are assumed to be completely non-overlapped. We propose a general method for solving the partitioning problem. This method involves three parameters, viz., bandwidth requirement, bottleneck, and load balancing factor that are used to judge the effe...

• Type: Journal
• Year: 2002
• Where: ISCA
• Authors: Sibabrata Ray, Hong Jiang
• Keywords: Hardware, ISCA 2002, Partitioning Problem, Task Graphs, Tree Task Graphs