Performance Modeling

Accurately estimating the performance of target applications on massively parallel machines is useful to application programmers in adapting their codes to the new architectures. Such a performance estimator is also an essential tool for designers of petaflops machines who, in order to make good design choices, need to evaluate alternate architectural features in the context of specific benchmarks.

It is clearly impractical, if not impossible, to simulate a million processor machine on a single processor. Instead, we aim at the challenges involved in carrying out such simulations on a conventional parallel machine with over 1,000 processors, attaining the desired timing accuracy using multi-level simulation techniques.

We have developed BigSim[6] for simulating petaflops class machines such as Blue Gene/L. In the rest of this section, we will first present the BigSim simulator, followed by BigNetSim. BigNetSim is a work in progress that extends BigSim with network simulation capability.