COMMUNICATION-SENSITIVE LOAD BALANCING AND ITS PERFORMANCE UNDER DISTRIBUTED WORKLOAD ENVIRONMENTS
November 1, 2PM
Dynamic load balancing algorithms that use process migration facilities can potentially improve overall system performance of workstation clusters by utilizing idle resources. However, separating tightly-coupled processes from groups of communicating processes by migration can cause a drop in system performance. This drop in performance is attributed to the increased communication cost between communicating processes that are separated. Current load balancers are not aware of the communication pattern between processes. We have implemented a communication-sensitive load balancer on DUNES---a user level distributed operating system---that uses the communication pattern between processes when making load balancing decisions. This load balancer, by keeping tightly-coupled processes together while migrating loosely-coupled processes to utilize idle resources, performs substantially better than a load balancer that uses CPU load information.
In this talk I will describe our communication-sensitive load balancing algorithm and compare its performance with a load balancer that is not aware of communication pattern between processes when subjected to different workload environments. This is a joint work with Prof. Kihong Park.
The Network Systems Colloquium is sponsored by the Network Systems Lab at Purdue University. For further information, please contact Kihong Park (firstname.lastname@example.org or 765-494-7821).