Clouseau: Probabilistic Dynamic Verification of Multithreaded Memory Systems

Dynamic verification enables a system to improve its availability by checking that its execution is correct as it is running. While high performance and low power are desirable, correctness— despite hardware faults and subtle design bugs—is most important. For multithreaded systems, memory system correctness is defined by the memory consistency model. Thus, dynamically verifying memory consistency would ensure that the entire memory system is operating correctly. We present the first implementable design for probabilistic dynamic verification of sequential consistency (pDVSC) in multithreaded systems. The system dynamically creates a total order of memory operations (loads and stores) and verifies that this total order obeys SC. In the theoretical world of systems without resource constraints, DVSC would have to consider the entire total order, but we show how to leverage resource constraints to verify only a sliding window of the total order. While we cannot bound the size of this window and still eliminate all false verifications (false positives or negatives), we can implement probabilistic verification and make the probability of false verification arbitrarily small. We use full-system simulation of a multithreaded system running commercial workloads to evaluate our first implementation of pDVSC, called Clouseau. Clouseau’s implementation costs are kept reasonable via extensive compression and caching of the data that is used for dynamic verification. Clouseau, combined with backward error recovery, improves availability by recovering from injected errors. Clouseau adds only negligible performance overhead. While Clouseau adds to system design complexity, we believe this is a small price to pay for improving system availability.