PTLib (Parallel Tools Library) Software Catalog

MagPIe

Abstract

This library is based on MPICH, a widely used MPI implementation, but the collective communication primitives use new algorithms, optimized for wide area systems. The basic as­ sumption is that the wide area system is hierarchically structured. We expect a computational grid to consist of many parallel com­ puters connected by wide area networks. We use the term cluster to denote a single parallel machine (which can be a network of workstations, an MPP, or an SMP). All nodes within a cluster are connected by a high­speed interconnect that is orders of magnitude faster than the wide area interconnect between clusters. Collective communication algorithms are usually designed for local area networks, which have a low latency. Wide area systems however, have a high latency (and a lower bandwidth) and the per­ formance of collective communication operations is dominated by the traffic over the wide area links. Thus, our algorithms are de­ signed to reduce traffic over the slow links, resulting in a markedly different communication structure. Our algorithms are wide area optimal in that an operation incurs only a single wide area latency, and every data item is sent at most once across each wide area link.

DateOfInformation

Wed Mar 17 14:52:59 1999

Domain

Distributed Processing Tools

Webpage

http://www.cs.vu.nl/albatross/

ContactIs

Aske Plaat
Thilo Kielmann