Policy Based Job Analysis

Abstract

Many job schedulers for clusters allow system administrators to specify a variety of policy constraints that can be used to implement site objectives. An example of a site objective could be to achieve high system utilization while at the same time giving users fair access to the resources. Policy can include such things as constraints on walltime, number of processors and memory that can be requested, fairshare targets, and privileges to certain users. This paper highlights the importance and uses of job scheduler policy information in job analysis. A suite of job analysis services that make use of job scheduler policy information is described. Using the analysis services, system administrators can determine if the current policy is having the desired effect and adjust policy as appropriate. Users can gain insight into the utilization they are achieving and determine more appropriate job submission strategies that could increase utilization or reduce queuing delays. Examples of the job analysis suite applied to production data from resources part of WestGrid, a high performance computing consortium in Western Canada, are also provided.