Performance evaluation of the CPPS distributed-memory parallel production system under various system parameters
✍ Scribed by Jun Miyazaki; Haruo Yokota
- Publisher
- John Wiley and Sons
- Year
- 2000
- Tongue
- English
- Weight
- 347 KB
- Volume
- 31
- Category
- Article
- ISSN
- 0882-1666
No coin nor oath required. For personal study only.
✦ Synopsis
We have proposed a distributed-memory parallel production system, named CPPS. CPPS divides PEs into several virtual clusters, and uses two load balancing strategies for high-speed inference: hash-based for the intercluster case and demand-driven for the intracluster case. A software cache method is also applied inside the cluster for optimization. In this paper, we show a cost model for the CPPS, parametrized by CPU and communication performance. With this cost model, if one parameter is assigned, we can determine the other parameters so that communication overheads can be avoided. We also evaluate the performance of the CPPS when the CPU and communication parameters vary. The results show that the CPPS can work efficiently without saturation as the performance of each CPU increases if that of the communication, in particular the setup time, would be effectively improved.