Determination of application times-to-solution for large-scale clustered computers continues to be a difficult problem in high-end computing, which will only become more challenging as multi-core consumer machines become more prevalent in the market. Both researchers and consumers of these multi-core systems desire reasonable estimates of how long their programs will take to run (time-to-solution, or TTS), and how many resources will be consumed in the execution. Currently there are few methods of determining these values, and those that do exist are either overly simplistic in their assumptions or require great amounts of effort to parameterize and understand. One previously untried method is queuing network modeling (QNM), which is easy to parameterize and solve, and produces results that typically fall within 10 to 30% of the actual TTS for our test cases. Using characteristics of the computer network (bandwidth, latency) and communication patterns (number of messages, message length, time spent in communication), the QNM model of the NAS-PB CG application was applied to MCR and ALC, supercomputers at LLNL, and the Keck Cluster at USF, with average errors of 2.41%, 3.61%, and -10.73%, respectively, compared to the actual TTS observed. While additional work is necessary to improve the predictive capabilities of …

None