Live migration of Virtual machines (VMs) refers to transfer of a running VM's state from one physical machine to another. The main benefit of live VM migration is to re-purpose physical servers by migrating some or all their workloads. For instance, VM migration enables consolidation, i.e. increasing VM density by packing more VMs into fewer physical hosts. Consolidation can be used to either increase server utilization, and hence revenue for the service provider, or it can used to save energy by shutting down idle servers, and hence reduce operating costs. VM migration can also enable dynamic scale-out to handle sudden surge in demand by migrating previously consolidated VMs to new servers that have more resources. Another use of VM migration is in responding to imminent platform failures by quickly migrating VMs to new servers, or by maintaining hot-standbys.

Faster the VM migration, more quickly the servers can be re-purposed in the above scenarios. Unfortunately, traditional live VM migration techniques have significant overheads which hamper their widespread adoption in datacenters. The chief among these overheads is that migrating a VM (or worse many VMs together) is a highly resource intensive operation. VM migration involves the transfer of large amounts of data over the datacenter network in the order of tens or hundreds of Gigabytes. This increases network, CPU, and memory usage at both migration endpoints and in the network infrastructure. Excessive migration overhead can ultimately result in degraded performance of VMs, violation of Service Level Agreements (SLA), and increased time to re-purpose physical servers.
 
This talk will focus on new techniques for fast and agile live VM migration. Our specific contributions are as follows: (1) Hybrid pre/post-copy live migration to reduce the total migration time and downtime when migrating a single VM (2) Live Gang Migration that uses local and distributed deduplication to speed up the simultaneous migration of multiple VMs, (3) Scatter-Gather live migration to reduce the VM eviction time and enable faster de-provisioning of server resources, and (4) Agile migration of scattered VMs for fast response during scale-out operations in datacenters. These individual techniques constitute various tools in a datacenter administrators toolbox that can be used to achieve desired performance objective under different circumstances.