Many important non-numeric applications are data intensive; hence, they will experience performance degradations as the gap between future processor and memory speeds widen. Unlike scientific workloads, non-numeric applications perform unstructured computations that exhibit low control-flow and address stream predictability, and low instruction and thread-level parallelism. These attributes limit the effectiveness of existing techniques for improving memory performance on non-numeric applications.
This CAREER award addresses the memory gap problem for non-numeric applications through an integrated research and education agenda. In research, several novel latency tolerance techniques will be developed to target unstructured computations. First, techniques for guiding prefetch instrumentation through runtime profile information will be studies to address unpredictable program behavior. Second, pointer prefetching techniques that aggressively schedule serialized memory operations will be developed to expose memory parallelism in pointer-intensive computations. Finally, "code-space" prefetch scheduling techniques will be investigated to enable prefetching along the multiple paths of execution commonly found in unstructured computations. In education, the topics and tools developed through research will be integrated into a curriculum that emphasized a "learn-by-doing" philosophy. In addition, undergraduate research programs and graduate student workshops will be created. Together, these education activities will produce students with a strong background in experimental systems.
