next up previous contents index
Next: 13.3 Fortran 90 Experiments Up: A Software Tool Previous: 13.2.7 Static Performance Estimator

13.2.8 Conclusion

Our emphasis in this work has been to try to recognize collective communication patterns rather than generate sequences of individual element sends and receives. Algorithm COMM determines this in a very natural way. This is especially important for loosely synchronous problems which represent a large class of scientific computations [Fox:88a]. Several communication utilities have been developed that provide optimal message-passing communication for such problems, provided the communication is of a regular nature and occurs collectively [Fox:88h].

We believe that our approach can be extended to derive partitioning schemes automatically. Data dependence and other information can be used to compute a fairly restricted set of reasonable data partitioning schemes for a selected program segment. The performance estimation module can then be applied in turn to each of the partitionings in the computed set.

The work described in this section was a joint effort between Caltech and Rice University, as part of the Center for Research on Parallel Computation (CRPC) research collaboration [Balasundaram:90a]. The principal researchers were Vasanth Balasundaram and Geoffrey Fox at Caltech, and Ken Kennedy and Ulrich Kremer at Rice. The data partitioning tool described here is being implemented as part of the ParaScope parallel programming environment under development at Rice University [Balasundaram:89c].

Guy Robinson
Wed Mar 1 10:19:35 EST 1995