Modeling VM Performance Interference with Fuzzy MIMO Model

Virtual machines (VM) can be a powerful platform for multiplexing resources for applications workloads on demand in datacenters and cloud systems. However, it remains challenging for the resource management in such a virtualized system to deliver performance guarantees because of the contention on nonpartitionable sources such as last-level CPU cache, memory bandwidth, and on-disk buffer, which introduces performance interference between co-hosted VMs and cannot be directly controlled. To address this challenge, this paper proposes to use fuzzy modeling to establish a multi-input-multi-output performance model for co-hosted VMs in order to capture their coupling. Based on the model, the level of contention on the competing resources is quantified by the model parameters, and can be used to guide VM placement and resource allocation decisions. Experimental evaluations on application benchmarks TPC-H and RUBiS demonstrate that this fuzzy modeling approach can detect and quantify the interference from VMs competing for both CPU and I/O; it can better capture the variation of such contention compared to a linear model.