An Application-Level View on Data Center Energy Efficiency
The broad acceptance of Power Usage Effectiveness (PUE) as a measure for data center energy efficiency is to be welcomed. After all, if we don’t measure it, we won’t improve it. However, the limits of PUE are widely recognized. Amongst the most glaring anomalies is the fact that improving the efficiency of your IT equipment, without an equivalent improvement to your cooling and power infrastructure, will cause your PUE rating to increase (as it is the ratio of IT equipment power used to the power draw of the infrastructure).
Underlying this distortion is the lack of any accepted measurement of the productivity of the data center, which would measure business value delivered against resources consumed. The ultimate goal for those working on data center efficiency metrics is to agree on an approach that can include IT workload measures as part of the standard metric for data center productivity. The EPA, the EU, the Japanese government, and The Green Grid are among the organizations looking to develop such metrics.
As in many areas of energy efficiency, making the true energy cost of any operation visible and assignable is key step to improving the way we use resources. What this means for data center management is that we will need to become more savvy about the work being done in the data center and its relationship to energy costs.
A simple solution is to turn off servers that are unproductive. The Green Grid has estimated on average around 15 percent of servers in a data center are comatose, that is to say still drawing power but not running any useful application or service. The next step is to look at the usefulness of the processes being run across all servers. This means developing a more granular understanding of what applications are being run and the resources they are using.
We have recently looked at two different approaches to application workload efficiency. The first is from 1E, the power management software vendor. 1E’s server power management solution has two main elements. The first is the provision of detailed analysis of the application-level utilization of IT servers. This enables IT managers to understand in detail the utilization of their IT assets. NightWatchman reporting can therefore help in the identification and elimination of comatose servers, as well as the manual fine-tuning of server performance. The second element is the NightWatchman Drowsy Server capability, which utilizes built-in processor power management capabilities to allow voltage levels to be modified on demand. 1E combines this capability with defined business rules to decide which workloads should run at lower power levels. These tasks are then carried out using minimum power levels, reducing power consumption when the server is not doing useful work.
A different approach to application-level resource management is taken by Californian startup Librato. Its Silverline workload manager is a software-as-a-service (SaaS) offering that dynamically balances workloads with reference to business priorities and server utilization. Silverline gathers applications together in “containers” and then ensures that these collected applications are run in a way that maximizes utilization while ensuring that there is no reduction in the performance or availability. For example, one container may include a business application while another holds background system applications. The first container will always be given priority in terms of access to systems resources, while the second will be allowed to make use of any excess capacity outside of peak load periods. In essence, Silverline acts as an “uber-scheduler” based on defined criteria and business policies. Librato Silverline does not shut down servers, rather, it “harvests” unused capacity while servers are not dealing with business-critical applications to process applications and procedures that are not time critical.
Improvements to cooling and power infrastructures and the move to virtualized environments currently dominate discussions on the improving energy efficiency in the data center, but going forward we will need to focus more closely on the applications that the data center is running and how they are optimized. 1E and Librato are among the suppliers paving the way in this area, but we expect others to follow.
Article by Eric Woods, appearing courtesy the Matter Network.
|Tags: data center Efficiency energy costs green grid IT power infrastructure workload efficiency||[ Permalink ]|