Using k-exclusion to implement resilient, scalable shared objects (extended abstract)

We present a methodology for the implementation of resilient shared objects that allows the desired level of resiliency to be selected based on performance concerns. This methodology is based on the k-exclusion and renaming problems. To make this methodology practical, we present a number of fast k-exclusion algorithms that employ “local spin” techniques to minimize the impact of the processor-to-memory bottleneck. We also present a new “long-lived” renaming algorithm. Our lcexclusion algorithms are based on commonly-available synchronization primitives, are fast in the absence of contention, and have scalable performance when contention exceeds expected thresholds. By contrast, all prior k-exclusion algorithms either require unrealistic atomic operations or perform badly. Our k-exclusion algorithms are also the first algorithms based on localspin techniques that tolerate process failures.