Memory latency consideration for load sharing on heterogeneous network of workstations

With the development of cheap personal computer and high-speed network, heterogeneous network of workstation has become the trend in high performance computing. This paper focuses on the load sharing problem for heterogeneous network of workstations. Load sharing means even workloads of all coordinated computers in the heterogeneous system without leaving any computer idle. When some nodes suffer from heavy loading, it is necessary to migrate some processes to the nodes with light loading. However, most load sharing policies focus only on different CPU speed and/or memory capacity without taking the effect of memory access latencies into consideration. In the paper, we propose a new load sharing policy, CPU-Memory-Power-based policy, to improve CPU-Memory-based policy. In addition to CPU speed and memory capacity, this policy also puts emphasis on memory access latency. Experimental results show that this method performs better than the other policies, and that memory access latency is actually an important consideration in the design of load sharing policies on heterogeneous network of workstation.