The Embarrassingly Parallel program comes from the NAS benchmark suite [2]. EP generates pairs of Gaussian random deviates and tabulates the number of pairs in successive square annuli. In the parallel version, the only communication is summing up a ten-integer list at the end of the program. In TreadMarks, updates to the shared list are protected by a lock. In PVM, processor 0 receives the lists from each processor and sums them up.
In our test, we solved the class A problem in the NAS benchmarks, in
which pairs of random numbers are generated.
The results are shown in Figure 1.
The sequential program runs for 2391 seconds.
Both TreadMarks and PVM achieve a speedup of 7.9 using 8 processors,
because compared to the overall execution time, the communication
overhead is negligible.
Rice Systems Group