As power becomes an increasingly large component of overall data center costs, we have examined a number of strategies for saving electricity.

Recently, I wrote about the impact of virtualization on electricity consumption. In that article we outlined a strategy for generating ROI in server consolidation - consolidating workloads and switching off the remaining servers until they are needed. This approach can save money by cutting power use, but pretty soon you will need those surplus servers to support your application growth.

More recently we examined a storage technology called MAID (Massive Array of Idle Disks) that intelligently spins-down disks and only reactivates them when their data is needed.

What about combining the two approaches? As workloads in your data center change, can you intelligently power-down and power-up servers on demand?

This power-saving strategy would involve intelligent power management at the data center level by combining these technologies: virtualization with resource pooling such as VMware’s Distributed Resource Scheduler (DRS), Intelligent Platform Management Interface (IPMI), and Wake-On-LAN (WOL).

For the time being, organizations would have to write some custom software to implement this, but the software would mostly consist of glue connecting these technologies together. The potential ROI would more than cover the cost of software development and maintenance. This approach will likely emerge as a turnkey application from a number of vendors if it has not been developed already as a commercial product.

Here’s how it would work. An organization would first deploy server virtualization technology on a pool of physical servers (blade servers or commodity servers, for example). Using virtualization management technology such as DRS, IT managers can increase utilization levels on some servers by dynamically migrating loads from less utilized servers. Once servers are freed up and become idle, a script can trigger a power-down or sleep-mode command through a remote invocation of the IPMI management protocol. This puts the idle server in sleep mode or turns it off completely.

When the workload demand increases, a script initiates Wake-On-LAN Ethernet messages to the switched-off servers, which are booted up in a matter of a few minutes. If rapid resource scaling is required then a small “buffer” of idle but not powered-down servers can be maintained at all times. That way, when there is an increase in demand there are servers available immediately for workload allocation even while the powered-down servers boot up to re-establish the “buffer” (this would be equivalent to the caching mechanism on MAID that accelerates retrieval time - here we would be “caching” CPU resources).