Is Watts Per Virtual Machine a Useful Measure of Efficiency?


We have been doing a lot of work trying to come up with some ways that VPs of Infrastructure can evaluate the
efficiency of their virtual environment. What we have found is that simple measures such as VMs/host or even CPU utilization don’t really tell much of a story on their own. These measures combined with others such as $/VM and VM/U are required to tell a holistic efficiency story. But we have always been on the lookout for other ways to measure efficiency. So when the New York Times published an interesting story on data center efficiency, it got us wondering, is watts/VM a meaningful measure for data center operators?
The answer is a resounding yes, but perhaps not as one would imagine. Like all efficiency measures, the relative change over time is more important than the absolute value of watts/VM. With each virtual machine having its own resource consumption characteristic combined with the differences in power consumption by each server, coming up with “the” efficiency number of watts/VM is somewhat meaningless. But tracking a decrease in power consumption per VM over time does have value.
So how does one reduce watts/VM? Much like many efficiency measures, administrators start with a fixed power consumption structure. Using the hardware cluster as the starting point, an admin will place X virtual machines on Y hosts with an overhead on each physical machine to allow for fail over of a single host. The watts consumed per virtual machine is simply the consumed power for the cluster divided by the number of virtual machines. With a fixed number of hardware servers in a cluster, the only way to improve power efficiency is to increase utilization of CPUs. Despite higher power consumption at higher utilization levels, the fixed amount of power required to run a server even at idle creates a situation where higher utilization consumes less power per CPU cycle. Hence the first way to measure power efficiency for virtual machines is to benchmark current consumption by cluster, then work to increase efficiency by increasing utilization in that cluster over time.
The second way to improve watt/VM efficiency is to reset the power consumption model by procuring new hardware (shameless link to new employer). In this case, the goal would be to provide the right number of CPU cycles and memory at the lowest possible power consumption. Theoretically, by making the cluster size large, the failure overhead required per machine is low. In addition, by maximizing utilization with the most power efficient systems, the administrator should arrive at the lowest possible watt/VM number. The key is once again to improve efficiency as part of a hardware refresh and not go backwards.
Watts/VM is a meaningful metric for administrators. While there is no specific absolute number for efficiency, the key is to understand the watts/VM consumed and work to drive it lower through increased CPU utilization and by procuring the most power efficient hardware for a specific virtual load.
To learn more about the efficiency of your virtual environment, download VKernel's free tool vOPS Server Explorer and be on the path to a more efficient infrastructure in 20 minutes.