Modeling and generating large-scale Google-like workload

摘要

One of the key element needed to test most large-scale scheduling algorithms is a testing infrastructure. Large scale is of upmost importance as failures and complex behaviors are common occurrences only at such scale. In order to test the reaction of a system to failures or extreme behaviors, it is necessary to be able to create large scale environments. Such an infrastructure must be reproducible so that several work are able to compare themselves but also capable of diversity as otherwise it would risk to lead to particular subcases. In this article, we propose a generic adaptable and reusable model of large scale workload. The original schema comes from the Google Cluster Workload Traces which is a perfect representative of large scale production workload. Contrary to most model analysis of such traces, we propose along with our model a reference implementation in order for other studies using our results to produce comparable experiments.