Abstract

We demonstrate an improved consensus-driven utility accrual scheduling algorithm (DUA-CLA) for distributable threads which execute under run-time uncertainties in execution time, arrival models, and node crash failures. The DUA-CLA algorithm’s message complexity (O(fn)), lower time complexity bound (O(D + fd + nk)), and failure-free execution time (O(D + nk)) are established, where D is the worst-case communication delay, d is the failure detection bound, n is the number of nodes, and f is the number of failures. DUA-CLA is shown to have the “lazy-abort ” property — abortion of currently-infeasible tasks is deferred until there is no possibility of completing the task on time. Further, it exhibits “schedule-safety ” — segments (and therefore, threads) proposed as feasible for execution by a node which fails during the consensus decision will be removed from the consensus set and will not cause an otherwise-feasible segment to be excluded. These properties mark improvements over earlier strategies in common- and worst-case performance. Finally, we present our implementation of DUA-CLA in a DRTSJ-compliant Java virtual machine, and quantitative results are presented validating fundamental properties of the algorithm. 1

Citations