Symbolic Cost Estimation of Parallel Applications

Symbolic cost models are an important performance engineering tool because of their diagnostic value and their very low solution cost when the computation features regularity. However, especially for parallel applications their derivation, including the symbolic simplifications essential for low solution cost, is an effort-intensive and error-prone process. We present a tool that automatically compiles process-oriented performance simulation models into symbolic cost models that are symbolically simplified to achieve extremely low solution cost. As the simulation models are intuitively close to the parallel program and machine under study, derivation effort is significantly reduced. Apart from its use as a stand-alone tool, the compiler is also used within a symbolic cost estimator for data-parallel programs. With minimal program annotation by the user, symbolic cost models are automatically generated in a matter of seconds, while the evaluation time of the models ranges in the milliseconds. Experimental results on four data-parallel programs show that the average prediction error is less than 15 %. Apart from providing program scalability assessment, the models correctly predict the best design alternative in all cases.