Conclusions

• Granularity analysis/control is a particularly promising technique because it can make it feasible to exploit low-cost parallel architectures, such as workstations on a local area network.

• The assessment (in collaboration with ECRC) of the granularity analysis/control module within the ECLPS environment has shown very promising results, significantly increasing speedups, both for and- and or-parallelism.

• It is not essential to be absolutely precise in inferring the best grain size for a problem: there is a reasonable amount of leeway in how precise this information has to be.

• This suggests that granularity control can usefully be performed automatically by a compiler.

• The techniques are not only useful in the improvement of independent and-parallelism and or-parallelism , but also for ``data-parallelism.''