Active memory techniques for ccNUMA multiprocessors

Our recent work on uniprocessor and single-node multiprocessor (SMP) active memory systems uses address remapping techniques in conjunction with extended cache coherence protocols to improve access locality in processor caches. We extend our previous work in this paper and introduce the novel concept of multi-node active memory systems. We present the design of multi-node active memory cache coherence protocols to help reduce remote memory latency and improve scalability of matrix transpose and parallel reduction on distributed shared memory (DSM) multiprocessors. We evaluate our design on seven applications through execution-driven simulation on small and medium-scale multiprocessors. On a 32-processor system, an active-memory optimized matrix transpose attains speedup from 1.53 to 2.01 while parallel reduction achieves speedup from 1.19 to 2.81 over normal parallel executions.