Performance tuning scientific codes for dataflow execution

Abstract

Pe~ormance tuning programs for datafow execution involves tradeoffs and optimizations which may be significantly different than for execution on conventional machines. We examine some tuning techniques for scientijic programs with regular control but irregular geometry. We use as an example the core of an ocean modeling code developed in the implicitly parallel language Id for the Monsoon datafow machine. DatafZow implementations of loops, in particulal; require careful examination by the compiler or programmer to attain high pe$ormance because of overheads due to$ne-grained synchronization, control, and operand distribution.