Consensus-driven distributable thread scheduling in networked embedded systems

Abstract

We demonstrate a consensus utility accrual scheduling algorithm for distributable threads with 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. The "lazy-abort" property is shown - abortion of currently-infeasible tasks is deferred until timely task completion is impossible. DUA-CLA also exhibits "schedule-safety" - threads proposed as feasible for execution by a node which fails during the decision process will not cause an otherwise-feasible thread to be excluded. These properties mark improvements over earlier strategies in common- and worst-case performance. Quantitative results obtained from our Distributed Real-Time Java implementation validate properties of the algorithm. © IFIP International Federation for Information Processing 2007.