大规模数据并行问题的可扩展性分析

大规模数据并行处理的性能受到处理机数量、I/O速度、通信速度等多方面因素的制约。增加处理机数量或提高处理机的计算速度，可以提高计算机的整体处理速度，但是通信和I/O会成为影响并行效率的主要因素。为了综合分析这些因素对计算性能的影响，用一种比较典型的大规模数据并行的计算模型，具体分析了处理机数量、处理机速度与处理机间的通信延迟、通信速率以及输入输出速度之间的关系。得到了大规模并行机的通信和I/O性能与处理机速度与数量之间存在的关系。指出，增加处理机数量、提高单节点处理速度的同时，必须按照一定的关系相应增加节点间的通信性能和I/O性能。单纯以增加处理机数量、提高单处理机速度提高计算机峰值速度的方法会降低系统的计算效率，不能达到计算速度与计算机处理能力同步增长的目的。

Scalability of Massively Data Parallel Computing Problems

The performance of distributed parallel systems is influenced by many factors, for examples, the total number of nodes in the system, node speed, I/O speed, latency and communication speed of the inter network. Adding more nodes and/or using more powerful nodes can improve the performance, but I/O and communication could suffer from it. To determine the relationship between performance and those factors, the scalability of the massively parallel computers by using a data parallel model is analyzed. The relationship between the number of nodes, the speed of the nodes and the communication speed and latency of the links between nodes, the I/O speed of the system is obtained. The results shows that it is necessary to increase the I/O speed and the speed of the links (decrease the latency) by a certain ratio while increasing the number or the speed of the processors, so as to keep the scalability of the system. Only increasing the number or the speed of the processors will decrease the scalability of the system.