Performance of HPF compiler

The figure below shows the speedup ratio for some benchmark programs. We used four benchmark programs, grid, shallow77, x42, and tomcatv, developed by APR, Inc. For each benchmark program, the speedup ratio was calculated by dividing the execution time of the parallel code by that of the serial code. The same source program was used for both codes. The parallel code was generated by our HPF compiler, while the serial code was generated by the single-processor compiler that has the single-processor specific optimization similar to our HPF compiler.

We demonstrate how effective our HPF compiler is in parallelization and optimization of a FORTRAN77 benchmark Shallow Water. A two-dimensional system of shallow water equations is numerically solved in the benchmark.

This animation shows how waves propagate when you throw several stones into a shallow water pool. Pressures at each point in the pool can be calculated in parallel. For parallelization, the pool is decomposed into 3x3 sub-pools by HPF directives in this demonstration. Calculations on each sub-pools are performed in parallel. Click to view a MPEG video where a HPF version outperforms a F77 version. This demonstration is the replay of its execution on RS/6000 SP. The left half of the video shows the benchmark running on nine processors, and the right half shows the benchmark running on a single processor of SP.