Clusters of inexpensive workstations are an increasingly attractive
alternative to traditional supercomputers for high-performance
applications. However, there are certain new challenges in meeting the
I/O needs of applications running on clusters. Often clusters have
relatively slow interconnects, so that the interconnect may be the
ultimate bottleneck for parallel I/O. Further, clusters are often
heterogeneous, which may wreak havoc with parallel I/O strategies that
were designed to be deployed in a homogeneous environment. In this
talk we describe our experiences with collective I/O requests in two
different cluster environments: a set of 8 HP workstations connected
with FDDI, and a set of 100 NT workstations connected by Myrinet. We
discuss the strategies we developed to provide high-performance I/O in
these environments using the Panda parallel I/O library.
Last modified June 10, 1998
(hiper98@ethz.ch)
!!! Dieses Dokument stammt aus dem ETH Web-Archiv und wird nicht mehr gepflegt !!!
!!! This document is stored in the ETH Web archive and is no longer maintained !!!