Abstract

This talk deals with improving the performance of fine grain task parallelism. It is often either cumbersome or impossible to increase the grain size of such programs. Increasing core counts exacerbates the problem; a program that appears coarse-grained on eight cores may well look a lot more fine-grained on sixty four.

We present Wool, a library that schedules tasks using work stealing. In this paradigm, idle processors steal work (tasks) from randomly selected non idle processors. In many cases, especially when tasks are fine grained, most tasks are never stolen. Such tasks incur only a minimal overhead (from a few to a few tens of cycles), while stealing costs hundreds to a couple of thousand cycles.

We discuss the design and implementation of Wool, with an emphasis of the features allowing low overhead implementation.