Dynamic Data Structures for Taskgraph Scheduling Policies with Applications in OpenCL Accelerators

OpenCL is an emerging open framework for parallel programming in heterogenous systems. Devices compliant with OpenCL need to schedule the execution of submitted jobs with no (or only very imprecise) estimates of execution times, but respecting dependencies among them, which are given in the form of directed acyclic graph. This problem is known as stochastic taskgraph scheduling, stochastic scheduling with precedencies, or stochastic scheduling with data dependencies. We study the complexity of implementing static out-of-order policies for taksgraph scheduling, which approach optimality in the long run, under certain assumptions. We present a simple data structure allowing for the “what next” query of such scheduling policies to be answered in time O(1), while vertices can be added in time O(1).