Software Logging under Speculative Parallelization

Speculative parallelization aggressively runs hard-to-analyze codes in parallel. Speculative tasks gen-erate unsafe state, which is typically buffered in caches. Often, a cache may have to buffer the state of several tasks and, as a result, it may have to hold multiple versions of the same variable. Modifying the cache to hold such multiple versions adds com-plexity and may increase the hit time. It is better to use logging, where the cache only stores the last ver-sions of variables while the log keeps the older ones. Logging also helps to reduce the size of the specula-tive state to be retained in caches. This paper explores efficient software-only log-ging for speculative parallelization. We show that such an approach is very attractive for programs with tasks that create multiple versions of the same variable. Using simulations of a 16-processor CC-NUMA, we show that the execution time of such programs on a system with software logging is on average 36 % shorter than on a system where caches can only hold a single version of any given variable. Furthermore, execution takes only 10 % longer than in a system with hardware support for logging.