Runtime Framework for Parallel and Adaptive Applications

This paper presents the design and implementation of the new version of our software, the Parallel Runtime Environment for Multicomputer Applications. This framework provides large-scale applications with implicit load balancing, scheduling and latency hiding through a simple but powerful interface. In the new design, the framework has been augmented with multithreading, separating communication and execution into different threads to provide asynchronous message reception and instant computation execution at the arrival of new work requests. Furthermore, it allows the application to run multiple computations in parallel, while monitoring the load of the system and performing migrations when desired. Scheduling and load balancing are also enhanced by introducing custom intra-node schedulers and the ability to perform multiple migrations that initiate from the same process further reducing the time spent in this procedure. The motivation for the development of the runtime system is to provide dynamic runtime support for parallel mesh refinement applications.Testing the system on such an adaptive application indicates an overall performance improvement of up to 40 percent with an overhead of less than one percent, in cases of up to 270 computing nodes, compared to no load balancing, by retaining a better work-load distribution among the execution units.