Performance Analysis with HPX

With High Performance Computing moving towards Exascale, where parallel applications will be required to run concurrently on millions of cores, every part of the computational model must perform ideally to achieve optimal performance. The task scheduler is one of such entities that could be enhanced to runtime application prerequisites. Not only the overheads associated with task scheduling vary depending on the task size but also are affected by the different schedulers used. One of the best strategies used to ameliorate performance and to decrease the overheads would be to adapt the scheduler to the application’s needs or to vary the task size or a combination of both. In this work, we show how scheduling overheads vary with the different types of schedulers used in the High Performance ParalleX (HPX) runtime system. In addition to this, we show how the different schedulers perform with varying task size. This performance analysis focuses on understanding how the different schedulers perform for a specific HPX benchmark in turn helping us to settle on a scheduler that could give optimal performance. HPX is an ideal experimental platform as it makes use of asynchronous fine-grained task scheduling and incorporates a dynamic performance modeling capability. In addition to this, HPX offers performance counter capabilities with which we can characterize scheduling overheads.