Towards Efficient Simulation of Task Based Parallel Applications
Alexandru C. Iordan, Magnus Jahre, Lasse Natvig · 2012
For computer architects and software developers, simulation is an indispensable tool for evaluating new designs and ideas. Unfortunately, it is significantly slower to simulate a parallel environment than to work on real hardware. In this work, we take a first step towards efficient simulation of Task Based Parallel (TBP) applications. Our key idea is that the number of completed tasks can be used as a work-related progress metric to sample these applications. Using this metric, we show that the complete execution of TBP programs can be accurately represented by simulating only a few samples. In fact, our TBP applications only need 5 samples on average to get a mean error below 5 %. In our experiments, when sampling at 1000 completed tasks, 5 samples correspond to 0.5 % of the total execution on average.