Space-Bounded Async Scheduling. A UPCxx extension

Abstract

It is estimated that computers and mobile devices use more than 2% of the total energy consumed. That means a lot of energy is going in powering our cpu,display and gpu. In this paper we are trying to optimize the power consumed by cpu in partitioned global address space environment. Recent research suggests that there is scope in improving cpu power usage by having a better scheduler. Simhadri et al.[14] concluded that space bounded scheduler can improve the efficiency by 60% in parallel environment in shared global address space. In this paper we introduce data centric approach to PGAS, particularly UPCxx[17], referred to as DUPC, inspired by space bounded scheduler. The scheduler determines the cache hierarchy which allows it to make intelligent decisions to schedule task given the size of task is known beforehand.Hwloc library is used to detect the cache hierarchy. Once the cache hierarchy and sizes of each cache is known, we can track available cache sizes. With this information in hand, and if task size is known, we are able to provide better cache locality. We use PGAS for high performance computing. We use UPCxx, which exploits power of PGAS.