Methods and systems for agile load balancing include detecting an increased load for a first primary virtual machine (VM) on a first node that has a plurality of additional primary VMs running on a processor; deactivating one or more of the additional primary VMs, reducing said one or more deactivated VMs to a secondary state, to free resources at the first node for the first primary VM; and activating secondary VMs, located at one or more additional nodes, that correspond to the one or more deactivated VMs, raising said secondary VMs to a primary state. Activation and deactivation through micro-checkpointing may involve nodes of different CPU architectures during transient periods of peak load.