The University of Tulsa
Memory is commonly a shared resource for a modern Chip Multi-Processor (CMP) system. Concurrently running threads have different memory access behaviors and compete for memory resources. A memory scheduling algorithms should be designed to arbitrate memory requests from different threads, provide high system throughput as well as fairness. This paper proposes a memory scheduling algorithm, Priority-Based Fair Scheduling (PBFS), which classifies threads memory access behavior by dynamically updated priorities. Latency-sensitive threads have top-priority to guarantee system throughput, and starvation of memory-sensitive threads can be avoided simultaneously.