Abstract

As multi-core architectures become more pervasive, applications running on commodity systems are being forced to turn to parallelism if they are to increase their performance. These consumer applications must compete with other programs running on the system for access to processing cores, unlike traditional parallel applications running on HPC systems, which typically the only application running on the machine. Therefore, the number of processing cores available to these applications varies from moment to moment. The process of parallelising an application to efficiently exploit a varying number of core-counts is further complicated by architectures which can dynamically varying the number of physically active cores, either for energy efficiency, or due to dynamically reconfiguration in architectures such as TRIPS or E2, which can dynamically re-compose multiple physical cores into a single, more powerful logical core.

In this talk, we will discuss our initial work on a runtime system called Calico. Calico is runtime system designed to provide the flexibility of dynamic work-stealing approaches, such as Cilk, while still providing the efficiency of a more static approach to parallelism, such as that provided by OpenMP. It does this by providing a programming model which enables the application to specify how to efficiently partition work between multiple threads, whilst hiding the work-stealing mechanism within the Calico runtime system. This enables the Calico runtime system to decide when more parallel work should be spawned by, for example, repartitioning the available work as more cores become available.