Efficient Parallel Scheduling of Malleable Tasks

Abstract

We give an $O(n + \min\{n,m\} \log{m})$ work algorithm for scheduling $n$ tasks with flexible amount of parallelism on $m$ processors, provided the speedup functions of the tasks are concave. We give efficient parallelizations of the algorithm that run in polylogarithmic time. Previous algorithms were sequential and required quadratic work. This is in some sense a best-possible result since the problem is NP-hard for more general speedup functions.