As physical machine memory sizes and CPU counts rise, so does the demand for large virtual workloads as well. One particular challenge of scaling large VMs is migrating them to new hosts for maintenance. This talk will look at the performance of the demand paging mechanism userfaultfd and how well it performs under stress, in particular how it scales with large number of vCPUs. Then, the talk will examine the guest memory and CPU performance during a demand paging migration of a VM on a 4-socket host with 160 vCPUs, and additionally compare userfaultfd to an alternate approach.
