Experience with Multi-resource Aware Fair Sharing in Highly Heterogeneous Private Clouds

This work addresses the problem of guaranteeing user-to-user fairness in absence of money. This is a typical problem in private and/or scientific clouds. In such systems, the computational power is not allocated with respect to a user ability to pay. Instead, users of the system are handled according to a resource provider's inner policy. Frequently, this policy somehow incorporates fairness. In this paper we show how fairness can be handled in private cloud systems where both resources and users' requests are highly heterogeneous. The solution is represented by a user-ordering function that defines fair ordering of users' requests with respect to their recent resource utilization, supporting multi-resource aware fair sharing. Unlike other solutions, our approach is less sensitive to the heterogeneity of resources and users' requests. Notably, scheduler decisions do not influence the resulting priority of a user. We discuss our positive experience with a working implementation of this fair sharing mechanism which is used in the Czech national cloud and grid operator Meta Centrum.