Experimental Evaluation of Optimal Schedulers Based on Partitioned Proportionate Fairness

The Quasi-Partitioning Scheduling algorithm optimally solves the problem of scheduling a feasible set of independent implicit-deadline sporadic tasks on a symmetric multiprocessor. It iteratively combines bin-packing solutions to determine a feasible task-to-processor allocation, splitting task loads as needed along the way so that the excess computation on one processor is assigned to a paired processor. Though different in formulation, QPS belongs in the same family of schedulers as RUN, which achieve optimality using a relaxed (partitioned) version of proportionate fairness. Unlike RUN, QPS departs from the dual schedule equivalence, thus yielding a simpler implementation with less use of global data structures. One might therefore expect that QPS should outperform RUN in the general case. Surprisingly instead, our implementation of QPS on LITMUS^RT invalidates this conjecture, showing that the QPS offline decisions may have an important influence on run-time performance. In this work, we present an extensive comparison between RUN and QPS, looking at both the offline and the online phases, to highlight their relative strengths and weaknesses.