In Algorithm Design for Performance Aware VM Consolidation we learn some shocking facts (gambling in Casablanca?):
- Average server utilization in many data centers is low, estimated between 5% and 15%. This is wasteful because an idle server often consumes more than 50% of peak power.
- Surely that's just for old style datacenters? Nope. In Google data centers, workloads that are consolidated use only 50% of the processor cores. Every other processor core is left unused simply to ensure that performance does not degrade.
It's a VM wasteland. The goal is to reduce waste by packing VMs onto machines without hurting performance or wasting resources. The idea is to select VMs that interfere the least with each other and places them together on the same server.
It's a NP-Complete problem, but this paper describes a practical method that performs provably close to the optimal. Interestingly they can optimize for performance or power efficiency, so you can use different algorithms for different workloads.
The result when optimizing for performance are utilizations between 75% and 80% compared to 50% from the naıve method. This gives a 22% reduction in TCO, a significant savings at scale:
Experimental evaluations showed that the proposed system performed well on realistic VM performance degradations, yielding over 30% savings in energy and up to 52% reduction in degradation.
It's hard to get greater utilization from VMs because systems are structured in such large chunks. If we could decompose systems into smaller more modular pieces they could be much more efficiently scheduled over a server farm.