Optimized Shared Virtual Memory

We propose to develop optimized shared virtual memory (SVM) for message-passing machines. SVM can be implemented in a distributed memory environment much as virtual memory is implemented on uniprocessor machines. The SVM on a parallel machine would permit the user to distribute pages of a virtual memory space across processors, generating interprocessor communications for memory faults. Thus, SVM can hide the daunting task of programming interprocessor communications in a dsitributed memory space from the user, at the price of potentially decreased performance. Unfortunately, current SVM implementations do not allow the user to propagate information to the compiler or operating system which might improve performance. We propose several optimizations to SVM which could make it competitive with hand-coded message passing programs for a wide range of applications, including iterative linear Bolvers, unbtructured computational fluid dynamics codes, and FEM codes for mechanical engineering. Under this grant, we will develop two simulators to determine potential performance improvements for irregular applications running under optimized SVM. We will also address the issues arising from using portable code annotations to generate optimized SVM source code. Our final report will include a functional description and implementation plan for an optimized SVM prototype. Anticipated Benefits/Potential Applications - Optimized Shared Virtual Memory can be used to program irregular HPC applications (e.g. iterative linear solvers, multigrid solvers) as easily as standard shared memory, while keeping performance close to that of explicit message passing code.