Performance Analysis of PC Cluster-based CC-NUMA System using Execution-driven Simulation
Chi-Jeong Ha, Sang-Hwa Jeong, O Su-Cheol · Jeongbo gwahaghoe nonmunji. si'seu'tem mich i'lon · 2001
This paper presents a PC cluster-based CC-NUMA system and evaluates the performance of the system using simulation. The CC-NUMA system if based on a CC-NUMA card which is plugged into PCI slot. The CC-NUMA card contains shared memory, network cache, and network control module. The CC-NUMA system shares memory on PCI bus, The coherence between the shared memory and the network cache is maintained by the IEEE SCI standard. To simulate the CC-NUMA system, Limes, an execution-driven simulator, Is used with a modification. The cache coherence algorithm is based on the SCI's typical set. For comparison purpose, a NUMA system using the Dolphin's PCI-SCI card, which does not utilize network cache, is also simulated. Various experiments are performed to measure the performance of the CC-NUMA system. According to the experiments, the CC-NUMA system shows considerable improvements compared with the NUMA system. The parameters obtained during experiments are reflected by the real implementation of the CC-NUMA system.