During non-peak hours, data centers can use live migration to consolidate virtual machines (VMs) into a fewer number of hosts. Once some of these VMs become active, they should be quickly migrated to host with enough resources to support the increased resource demands to prevent performance degradation on the source host. Unfortunately, traditional live VM migration technique, such as pre-copy and post-copy, lack the necessary agility to respond to memory hotspots as they always require the transfer of a VM's entire memory state. Prior works have attempted to optimize indirect measures of migration effectiveness such as downtime, total migration time, and network overhead. However, none have treated the performance of VMs impacted by migration as the primary metric of migration effectiveness. We propose Agile migration, a technique to quickly recover the performance of all VMs under resource pressure by eliminating resource pressure faster than traditional live migration techniques. Our approach works by transparently tracking the working set of a VM and transferring the non-working set (cold pages) to a per-VM swap device while the VM is running. We use a watermark-based migration triggering and VM selection mechanism to detect and prevent memory hotspots by pro-actively migrating VMs. During migration, Agile migration uses a hybrid of pre-copy and post-copy migration to transfer only the working set of a VM. Cold pages on the swap device are skipped during Agile migration and can be demand-paged from the per-VM swap device when required. When live migrating under memory pressure, Agile migration demonstrate a reduction in the performance impact on the VMs up to a factor of 2.49, reduction in total migration time up to a factor of 4.7, and a reduction in total data transferred up to a factor of 2.6 compared to pre-copy migration, post-copy migration, and not migrating.
展开▼
