In search of the optimal adaptive load sharing policy for distributed systems
暂无分享,去创建一个
In a distributed computer system environment, it is possible to utilize the capacity of idle workstations without influencing their performance by using an adaptive load sharing (LS) policy.
The objectives of this study are: (1) System performance will be assessed under different system architecture and task nature. The distributed system can be homogeneous or heterogeneous and the tasks can be single class or multiclass. (2) When compared with no LS policy, performance improvement will be determined for the policy with task placement, task migration, and both. (3) Homogeneous and heterogeneous systems with single class and multiclass tasks under different CPU scheduling disciplines and each corresponding optimal policy will be considered.
Unlike most other adaptive LS policy studies, each node in the distributed system is modeled as a central server model. A policy iteration method based on the Markov decision process is proposed to find the optimal adaptive LS policy. This analytical approach can find the benefits (non-benefits) for each associated policy. But it can not find the exact performance measures (e.g., mean response time, average per-process response ratio, and global response ratio).
The simulation model was implemented in the iPSC/2 multicomputer environment. In the homogeneous distributed systems, the simulation results show that the performance improvements between no LS, LS with task placement, and LS with task migration are very small. These results are quite different from the other studies, which show a significant improvement of mean response time by executing LS policy. This is because the social delay cost will mitigate the potential benefits. In the heterogeneous distributed systems, the performance measures improvement is very impressive. Especially, the CPU scheduling is generalized foreground-background (FB). The mean value analysis (MVA) algorithm provides upper and lower bound performance measures for the simulation model.