Procrastination Beats Prevention: Timely Sufficient Persistence for Efficient Crash Resilience

Preserving the integrity of application data across updates in the presence of failure is an essential function of computing systems, and byte-addressable non-volatile memory (NVM) broadens the range of fault-tolerance strategies that implement it. NVM invites database systems to manipulate durable data directly via load and store instructions, but overheads due to the widely used mechanisms that ensure consistent recovery from failures impair performance, e.g., the logging overheads of transactions. We introduce the concept of Timely Sufficient Persistence (TSP) mechanisms, which is relevant to both conventional and emerging computer architectures. For a broad spectrum of faulttolerance requirements, satisfactory TSP mechanisms typically involve lower overheads during failure-free operation than their non-TSP counterparts; hardware and OS support can facilitate TSP mechanisms. We present TSP variants of programs representing two very different classes of sharedmemory multi-threaded software that store application data in persistent heaps: The first employs conventional mutexes for isolation, and TSP substantially reduces the overhead of a fault-tolerance mechanism based on fine-grained logging. The second class of software employs lock-free and wait-free algorithms; remarkably, TSP is very easy to retrofit onto a non-resilient design and enjoys zero runtime overhead. Extensive experiments confirm that TSP yields robust crash resilience with substantially reduced overhead.