Experimental Evaluation of QSM, a Simple Shared-Memory Model

Parallel programming models should attempt to satisfy two conflicting goals. On one hand, they should hide architectural details so that algorithm designers can write simple, portable programs. On the other hand, models must expose architectural details so that designers can evaluate and optimize the performance of their algorithms. Using both microbenchmarks and several representative algorithms, we experimentally examine the trade-offs made by a simple shared-memory model, QSM, to address this dilemma. The results indicate that analysis under the QSM model yields quite accurate results for reasonable input sizes and that algorithms developed under QSM achieve performance close to that obtainable through more complex models, such as BSP and LogP.