Improving distributed workload performance by sharing both CPU and memory resources

Abstract

We develop and examine job migration policies by considering effective usage of global memory in addition to CPU load sharing in distributed systems. When a node is identified for lacking sufficient memory space to serve jobs, one or more jobs of the node will be migrated to remote nodes with low memory allocations. If the memory space is sufficiently large the jobs will be scheduled by a CPU-based load sharing policy. Following the principle of sharing both CPU and memory resources, we present several load sharing alternatives. Out objective is to reduce the number of page faults caused by unbalanced memory allocations for jobs among distributed nodes, so that overall performance of a distributed system can be significantly improved. We have conducted trace-driven simulations to compare CPU-based load sharing policies with our policies. We show that our load sharing policies not only improve performance of memory bound jobs, but also maintain the same load sharing quality as the CPU-based policies for CPU-bound jobs. Regarding remote execution and preemptive migration strategies, our experiments indicate that a strategy selection in load sharing is dependent on the amount of memory demand of jobs-remote execution is more effective for memory-bound jobs, and preemptive migration is more effective for CPU-bound jobs. Our CPU memory-based policy using either high performance or high throughput approach and using the remote execution strategy performs the best for both CPU-bound and memory-bound jobs.