Performance tuning scientific codes for dataflow execution

Performance tuning programs for dataflow execution involves tradeoffs and optimizations which may be significantly different than for execution on conventional machines. We examine some tuning techniques for scientific 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 dataflow machine. Dataflow implementations of loops, in particular, require careful examination by the compiler or programmer to attain high performance because of overheads due to fine-grained synchronization, control, and operand distribution.