An Experiment with VM/XA Share

Abstract

The share parameter defines for VM/XA the relative importance of a userid's demand for CPU time. The normal or default share is 100. If a userid has a share of 40, one might expect it would receive 40% of the CPU time of a normal userid. Also, one might expect a fixed ratio of delivery of CPU time under various load conditions. For experiments documented in the paper, performed on the CERN and Cornell systems, these relationships do not hold. These experiments helped me to better understand what I call Effective Share in a production system. Even though Effective Share differs from expectations, it is a powerful tool for managing VM/XA workload. The audience for this paper is systems managers who want to understand more about the dynamics of share. They can use this document as a source to define a share policy for their workload management tools.