SCHEDULING OF REGULAR AND IRREGULARCOMPUTATIONS : A System and Applications

We study the performance of task graph scheduling for both regular and irregular parallel computation. We present a scheduling tool named PYRROS, which takes as an input a task graph and a compile time estimation of computation and communication weights. It then produces a schedule and a parallel code based on this schedule. We analyze the eeect of compile time errors to run time scheduling performance. For coarse grain task graphs small errors at compile time have a minimal impact in run time performance. We use PYRROS to produce parallel code for Gauss Jordan(GJ) and Gaussian Elim-ination(GE), sparse matrix computation and Fast Multipole method for irregular n-body simulations. The performance for GJ and GE on dense matrix computation is comparable to optimized hand written codes. For irregular problems such as n-body, scheduling overhead could become signiicant, but because of the iterative nature of such problems the overhead can be amortized over many iterations. And the overall performance with automatic scheduling is better than optimized manually derived schedules with no overhead.