Thumbnail Image


umi-umd-2854.pdf (856.58 KB)
No. of downloads: 1217

Publication or External Link






In this thesis, we propose and evaluate several techniques to dynamically increase the memory access locality of scientific and Java server applications running on cache-coherent non-uniform memory access(cc-NUMA) servers. We first introduce a user-level online page migration scheme where applications are profiled using hardware monitors to determine the preferred locations of the memory pages. The pages are then migrated to memory units via system calls. In our approach, both profiling and page migrations are conducted online while the application runs. We also investigate the use of several potential sources of profiles gathered from hardware monitors in dynamic page migration and compare their effectiveness to using profiles from centralized hardware monitors. In particular, we evaluate using profiles from on-chip CPU monitors, valid TLB content and a hypothetical hardware feature.

We also introduce a set of techniques to both measure and optimize the memory access locality in Java server applications running on cc-NUMA servers. In particular, we propose the use of several NUMA-aware Java heap layouts for initial object allocation and use of dynamic object migration during garbage collection to move objects local to the processors accessing them most. To evaluate these techniques, we also introduce a new hybrid simulation approach to simulate memory behavior of parallel applications based on gathering a partial trace of memory accesses from hardware monitors during an actual run of an application and extrapolating it to a representative full trace.

Our dynamic page migration approach achieved reductions up to 90% in the number of non-local accesses, which resulted in up to a 16% performance improvement. Our results demonstrated that the combinations of inexpensive hardware monitors and a simple migration policy can be effectively used to improve the performance of real scientific applications. Our simulation study demonstrated that cache miss profiles gathered from on-chip hardware monitors, which are typically available in current micro-processors, can be effectively used to guide dynamic page migrations in an application. Our NUMA-aware heap layouts reduced the total number of non-local object accesses in SPECjbb2000 up to 41%, which resulted in up to a 40% reduction in the memory wait time of the workload.