Cache affinity and resequencing in a shared-memory multiprocessing system

Abstract

Most shared-memory multiprocessor systems have a per-processor memory cache that can improve average memory access time. A task executing on a processor with a cache, accumulates a considerable amount of context at that processor in the form of cache entries. Therefore, a task that restarts after I/O, will probably execute faster on the processor it last ran on, since some of its data may still be present in the cache. A task that is shifted to another processor will have little or no initial cache context, so it must renew its context through cache misses. This not only increases task execution time, but also degrades overall system performance due to an increase in bus traffic. In this paper, we examine processor cache affinity in systems that require resequencing to improve system and program performance. We show that some cache affinity benefits are not exploited due to resequencing delays. Also, simulation is used to study the closed queueing network model of a shared-memory multiprocessing system while it is executing under a variety of workloads.