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.