Optimizing COOP Languages: Study of a Protein Dynamics Program

Fine-grained concurrent object-oriented programming (COOP) models which provide a shared names-pace, object-level concurrency and implicit dynamic thread creation can simplify the programming of irregular parallel applications on distributed memory machines. Unfortunately, COOP models are often perceived as ineecient and thus few complete applications have been implemented in COOP languages. In this paper , we study the implementation techniques required to obtain eecient parallel execution of ne-grained COOP languages using a complete, medium-sized protein molecular dynamics program, IC-CEDAR. We found that even given high data locality and achieving good sequential eeciency, an implementation that relies only on thread-oriented compiler and runtime optimizations and software multithreading fails to achieve parallel eeciency. We show that two major sources of overhead | lack of processor-level data reuse and ne-grained threads for remote object accesses | contribute to this ineeciency. Using code transformations that can be automated in a compiler, we demonstrate two processor-oriented optimizations for ne-grained COOP languages that overcome the ineeciency and enable IC-CEDAR to achieve parallel performance comparable to a highly-tuned SPMD program.