Two Hardware-Based Approaches for Deterministic Multiprocessor Replay

摘要

Many shared-memory multithreaded executions behave nondeterministically when ran on multiprocessor hardware such as emerging multicore systems. Recording nondeter-ministic events in such executions can enable deterministic replay-e.g., for debugging. Most challenging to record are memory races that can potentially occur on almost all memory references. For this reason, researchers have previously proposed hardware to record key memory race interactions among threads.rnThe two research groups coauthoring this paper independently uncovered a dual approach: focus on recording how long threads execute without interacting. From this common insight, the groups developed two significantly different hardware proposals. Wisconsin Rerun makes few changes to standard multicore hardware, while Illinois DeLorean promises much smaller log sizes and higher replay speeds. By presenting both proposals in one paper, we seek to illuminate the promise of the joint insight and inspire future designs.