Although energy costs are the fastest-rising cost element in the data center, the power battle hasn’t been lost. There are still many opportunities to improve efficiency. These include cooling optimization using hot and cold aisles, increasing rack density, turning on/off machines on demand, and balancing load in the data center to optimize cooling and reduce power consumption.
All these opportunities can potentially be achieved with Intel® Intelligent Power Node Manager, a technology embedded into Intel chips in a select group of servers. Some of most common scenarios where Intel Intelligent Power Node Manager can be applied beyond monitoring include:
- Increasing compute density—enforcing power limits based on power reported and populating racks with more servers using the previously stranded power capacity in the rack
- Linking cooling to actual demand—coordinating Intel Intelligent Power Node Manager power and thermal data with data center cooling controls to help ensure that adequate, but not excessive, cooling is provided, minimizing cooling costs
- Dynamically balancing resources—using migrations tools to move workloads to racks with available power headroom, and using Intel Intelligent Power Node Manager’s power capping to help ensure the rack budget is not exceeded
In addition to these optimization scenarios, Intel Intelligent Power Node Manager can be applied to increase availability, applying power capping in case of power outage and reducing the overall consumption with some performance penalty.
With the launch of the Intel® Xeon® E5 processor, the second generation of Intel Intelligent Power Node Manager (aka. 2.0) has been released. It is designed to improve monitoring and control granularity and to allow implementation of a range of usage models, as depicted here:
These scenarios go from simple real-time power monitoring to integrated data center power management practices. Expected higher payoffs for power management require higher investment and process maturity to deploy.
You don’t necessarily have to step up to the top, or even one of the more advanced usage models. Some situations could be enough for usage model No. 1, Real Time Server Power Monitoring. There may be no reason to invest beyond this point.
In usage model No. 2 (Power Guard Rail) and No. 3 (Static Power Capping), Intel Intelligent Power Node Manager allows you to pack servers more densely in a rack by imposing a guaranteed power limit.
Consider this scenario: In a traditional method, we usually take the specification of the power supply rating from the server manufacture, e.g. 650W, and test in a lab the real power consumption using a power meter. We then discover that 400W is reasonable to be used. In a typical 4KW power envelope, we usually populate the rack with 10 servers (i.e. 4.000W/400W = 10). Using Intel Intelligent Power Node Manager in the same server, measurements indicate that for a defined workload, the power consumption rarely exceeds 250W. Using that as an aggressive power/server budget, and enforcing 4KW for a global cap, i.e. the entire rack, only on rare situations could the consumption exceed the 4KW envelope, and will not exceed that amount due to Intel Intelligent Power Node Manager policy. In this scenario, we can then populate a rack with 16 servers instead of 10 (i.e. 4.000W/250W = 16), for an increase of 60 percent in rack density.
The Static Power Capping usage model employs more aggressive capping. There will be some performance impact during peak, but this should be OK as long as the service level agreement (SLA) is met. The effect is to increase infrastructure utilization.
Usage model No. 4 (Dynamic Power Capping) implements continuous capping for additional power savings. The capping level is determined by the application performance monitor driving a power management policy. This scheme may not be practical if the performance monitoring facility is not available.
For instance, in virtualized environments, where hosts run a variety of applications, it is difficult to isolate a meaningful indicator representing the application mix. For want of a better indicator, monitoring CPU utilization has been surprisingly useful in some settings. The idea is to impose a cap on a server based on the current CPU utilization in that server. The actual capping level, in watts, is derived heuristically from offline experiments with representative workload mixes, yielding energy savings of 10 to 15 percent over a daily workload cycle.
In usage model No. 5 (Hybrid Usages), the practical capping range is limited to about 30 percent of peak power in light configurations. If the goal is energy saving, non-operating states, such as hibernation, must be added to Intel Intelligent Power Node Manager policies. This is possible in virtualized cloud environments that allow dynamic consolidation of workloads into a pool of active machines and the shutting down of unused machines.
What’s new with Intel Intelligent Power Node Manager 2.0
The following table compares the features in each version of Intel Intelligent Power Node Manager.
I would love to hear from you if you have a specific usage case that is not covered.