Download Now Free registration required
This paper evaluates new techniques to improve performance and efficiency of Chip MultiProcessors (CMP) for workloads consisting of multiple multithreaded applications. Multithreaded applications contain serial phases (single thread) and parallel phases (many threads). While scheduling threads, current techniques do not differentiate between these two phases, resulting in sub-optimal usage of the multiprocessor resources. In this paper, the authors propose a new thread scheduling mechanism which takes into account the different requirements of each phase, granting higher priority to applications during their critical-serial phases. The advantages of the proposed scheduling mechanism, shown by analytical and experimental evaluation, are threefold. First, system throughput and power efficiency are improved by making better use of the available multiprocessor computing power.
- Format: PDF
- Size: 723.3 KB