A Performance Study of Isotach Version 1.0

Author:Bartholet, Robert, Department of Computer ScienceUniversity of Virginia

The Isotach version 1.0 prototype implements isotach logical time on a cluster of Intel-based personal computers linked with a Myrinet local area network. A performance study was conducted on this unoptimized version of the prototype to determine Isotach's performance characteristics, look for problems in the software and hardware design, and discover ways to optimize the system. Tests were conducted using custom developed programs and the Communications Analysis and Simulation Tool (CAST), a network benchmarking tool developed by the U.S. Navy and used at their request. In order to give us a baseline, we also performed tests on Fast Messages 1.1, a messaging layer known to provide good performance. CAST measurements were not as accurate as our custom developed tests due to software overhead in the tool. We explain the reasons for this overhead, and show that CAST accuracy is degraded on networks with sub-millisecond latencies. Our performance study measured throughput, latency, server performance under contention, effects of isochron size, and effects of token traffic on non-isotach traffic. The all- software Isotach version 1.0 prototype exceeded many of its performance goals. We expect performance to improve even more when custom-built hardware components replace software. Limitations in the design of the Isotach prototype were discovered, and recommendations were made to improve performance in future Isotach systems.

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

Bartholet, Robert. "A Performance Study of Isotach Version 1.0." University of Virginia Dept. of Computer Science Tech Report (1999).

University of Virginia, Department of Computer Science
Published Date: