To improve the performance and energy consumption of chip multiprocessor (CMP) system, memory request serving latencies should be minimized. These latencies can be minimized by scheduling appropriate memory command at appropriate time. This paper proposes a scheduler that reduces latency related to serving memory read requests by delaying switching into write drain mode when memory traffic is not heavy and write queue is not full. Memory reads are more important to handle than memory writes for system's performance. Further precharge and activate operations are performed using constant stride prefetcher. In idle memory cycles the scheduler issues row precharge commands using cache prefetching technique based on Global History Buffer. Authors in [1] have used stride detector and Global History Buffer based speculative precharges and activates, but they treat memory reads and memory writes equally. Whereas, proposed scheduler in this paper prioritizes reads over writes for better system performance. Our evaluations show that proposed scheduling policy significantly outperforms previous schedulers [1], [2] in varied multicore environments in terms of performance as well as energy consumption. Across a wide range of workloads based on PARSEC benchmark suite, proposed policy improves systems performance by 2.51%, on 2-core, 0.012% on 4-core environment in comparison to scheduler proposed in [1].
Improving the Performance of Chip Multiprocessor by Delayed Write Drain and Prefetcher based Memory Scheduler
01.03.2018
6036454 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
Accelerating Graph Analytics Using Attention-Based Data Prefetcher
Springer Verlag | 2024
|Link scheduler for network-on-chip
British Library Online Contents | 2008
|Multiprocessor Shared Memory (MSM) Development
NTIS | 1983
|High-performance, Dependable Multiprocessor
IEEE | 2006
|