Experiences with Legion on the Centurion Cluster

Report
Authors:Lindahl, Greg, Department of Computer ScienceUniversity of Virginia Chapin, Steve, Department of Computer ScienceUniversity of Virginia Beekwilder, Norman, Department of Computer ScienceUniversity of Virginia Grimshaw, Andrew, Department of Computer ScienceUniversity of Virginia
Abstract:

<p>The Legion Project spends most of its time building metacomputing environments and applications, but we are also actively building a clustered system, which we have named Centurion. Centurion is currently a cluster of 64 Alpha PCs, networked with low-latency gigabit networking hardware from Myricom, joined by a dozen x86-architecture PCs. We plan on doubling the size of the cluster in the near future. This shared-nothing, heterogeneous environment is an ideal fit for the Legion metacomputing system, which provides services such as naming, scheduling, heterogeneous communications, and parallel I/O. Centurion is roughly the size of "small" HPC systems in use today at supercomputing centers, giving us the opportunity to compare price and performance with traditional systems, and investigate the hardware and software challenges of building large clusters with commodity hardware.</p> <p>This paper will cover our experiences with Centurion and Legion. We begin with an overview of the Legion metacomputing system. We will then discuss the costs and hidden costs of assembling a cluster. We then describe the performance of Centurion, using microbenchmarks of the communications system, standard parallel application benchmarks, and user applications. User applications running on the system come from a variety of scientific disciplines, and range from traditional MPI codes taken straight from supercomputers to more novel applications using "bag of tasks" and macro-dataflow formalisms. Within this range of applications, we will discuss our successes and failures with hiding network latency, load balancing, and most importantly, usability of the system by researchers without extensive training in parallel computing.</p>

Rights:
All rights reserved (no additional license for public reuse)
Language:
English
Source Citation:

Lindahl, Greg, Steve Chapin, Norman Beekwilder, and Andrew Grimshaw. "Experiences with Legion on the Centurion Cluster." University of Virginia Dept. of Computer Science Tech Report (1998).

Publisher:
University of Virginia, Department of Computer Science
Published Date:
1998