Real-World Clustering for Task Graphs on Shared Memory Systems

Abstract

Due to the increasing desire for safe and (semi-)automated parallelization of software, the scheduling of automatically generated task graphs becomes increasingly important. Previous static scheduling algorithms assume negligible run-time overhead of spawning and joining tasks. We show that this overhead is significant for small- to medium-sized tasks which can often be found in automatically generated task graphs and in existing parallel applications.

By comparing real-world execution times of a schedule to the predicted static schedule lengths we show that the static schedule lengths are uncorrelated to the measured execution times and underestimate the execution times of task graphs by factors up to a thousand if the task graph contains small tasks. The static schedules are realistic only in the limiting case when all tasks are vastly larger than the scheduling overhead. Thus, for non-large tasks the real-world speedup achieved with these algorithms may be arbitrarily bad, maybe using many cores to realize a speedup even smaller than one, irrespective of any theoretical guarantees given for these algorithms. This is especially harmful on battery driven devices that would shut down unused cores.

We derive a model to predict parallel task execution times on symmetric schedulers, i.e. where the run-time scheduling overhead is homogeneous. The soundness of the model is verified by comparing static and real-world overhead of different run-time schedulers. Finally, we present the first clustering algorithm which guarantees a real-world speedup by clustering all parallel tasks in the task graph that cannot be efficiently executed in parallel. Our algorithm considers both, the specific target hardware and scheduler implementation and is cubic in the size of the task graph.

Our results are confirmed by applying our algorithm to a large set of randomly generated benchmark task graphs.