Tracking working set sizes of virtual machines using miss ratio curves

Abstract

Working sets are sets of pages that have been most recently accessed by virtual machines (VMs). They are often used within the memory scheduler of a hypervisor to estimate the memory demands of VMs running on the hypervisor. In order to manage the memory resources of the hypervisor efficiently, it is essential that these working set sizes be estimated accurately at any given point in time. Currently, a statistical sampling strategy is used within VMware ESX hypervisors to estimate the working set sizes of VMs. Using this technique, a small number of random pages is selected to form a sample set. Access to these sampled pages is then tracked and the percentage of sampled pages that are accessed is used to estimate the working set size of a VM. This technique, though simple, does not provide a very accurate estimation of the working set size. A more promising tool that can be used to accurately estimate the working set size of a VM is a miss ratio curve (MRC). An MRC is a curve that plots the predicted miss ratio of a VM against the total available memory given to the VM. Even though MRCs can estimate working set sizes of VMs with much better accuracy, they are still not widely used in practice because building these curves incurs too much overhead, thus affecting the overall system performance. However, a recent study has found a way to reduce the cost of building these curves, making them a promising tool that can be used to estimate working set sizes. In this thesis, I propose that MRCs be used as an alternative to the statistical sampling strategy currently employed within VMware ESX. I will demonstrate how to apply the state of the art technique found in the recent study to construct accurate MRCs without incurring too much overhead, and use these curves to track working set sizes of VMs. I will also show that these curves can estimate working set sizes of VMs with much better accuracy than the statistical sampling strategy.