Date: Tue, 10 Dec 1996 03:21:29 GMT Server: NCSA/1.4.2 Content-type: text/html Presentation of Simulation Results

Presentation of Simulation Results


Because the format of presentation of simulation (or experimental) results in network routing studies varies greatly, we present these pages as a suggestion of standard presentation formats. The formats were inspired by discussion at PCRCW '94 . There were two main issues discussed:


Graph Format

Two formats were suggested for graphing results. Paired throughput vs. applied load and latency vs. applied load comprise Chaos Normal Form. A single latency vs. achieved throughput graph makes Burton Normal Form. Use of only latency (including source queueing) vs. applied load is discouraged because it is impossible to gain any data about performance above saturation using such graphs.


Units for Bandwidth/Load

The literature describing results typically includes graphs of latency or throughput vs. applied load. Many units have been used to express this, including flits injected/cycle, flits/ns, bits/cycle, normalized bandwidth, and many others. Because the many forms require conversions to compare with other results, we propose using a single form: normalized bandwidth.

Normalized bandwidth simply expresses the load/throughput as a fraction of the bisection-bandwidth limited maximum bandwidth of the network for uniform random traffic. Essentially, this constraint, corresponding to normalized throughput = 1.0, is derived by considering that 50% of uniform random traffic crosses the bisection of the network. Thus, if a network has bisection bandwidth B bits/sec, each node in an N-node network can inject 2B/N bits/sec at the maximum load. An optimal routing algorithm could handle such a load without before saturating.

Non-uniform or non-random traffic patterns may saturate at different loads than uniform random traffic. However, please normalize to the uniform random traffic limit.


Back to the Chaotic Routing Home Page

kwb@cs.washington.edu