Intel has unveiled the Intel Data Center Manager, a software development kit for monitoring power consumption on datacenter servers and adjusting power consumption on an as-needed basis.
The caveat? The servers have to utilize Intel’s Intelligent Power Node Manager embedded in Intel’s newly launched Xeon 5500 “Nehalem” generation of server chips to fully use this software.
The software can monitor intelligent power supplies, but to get the full benefit of thermal and power management, the Intelligent Power Node Manager has to be present on the hardware and exposed.
The software, released Thursday, is designed to give a more accurate look into the power being consumed in a server’s cabinet, which can hold a number of racks or blades. The problem, as Jon Khazam, vice president and general manager of the manageability and middleware division at Intel (NASDAQ: INTC) noted, is that people build to the wrong specs.
Each cabinet has a few power supplies, as much for redundancy as for powering the racks. If a cabinet has, for example, three power supplies with a combined 6,000 kilowatts, then the cabinet won’t be populated with anything beyond 6,000 kilowatts total power draw, and usually less. The maximum power is listed on the power supply’s nameplate.
The problem is people build to the maximum spec. If a rack-mounted server says it has a 300 kilowatt draw, then the most racks going into that cabinet is 20. The problem is, Khazam noted, servers don’t run at their maximum draw. They usually operate at some power level below that.
“There’s been the general challenge of dealing with power and constraints of power,” he told InternetNews.com. “They tend to design a datacenter at the maximum of the nameplate power spec, which ends up introducing a lot of overbuild into the datacenter, where there’s a lot more capacity to handle servers than they end up deploying.”
The Data Center Manager works in two areas; on the individual rack and on the datacenter at large. On an individual rack, it gives an accurate measure of the overall average power draw, not the maximum draw. It shows individual units and the total power draw in the entire cabinet. Administrators can then limit each rack’s power draw. For example, a 300 watt max rack might be limited to 150 or 200 watts. This allows for more compute density to be added.
Power sharing that works
When one rack in the cabinet is running at the upper limits of its power capacity, the software can then examine the other racks, find ones that are drawing much less power than they have allocated, and give the power to the racks that need it.
In one example, Intel worked with Chinese search engine provider Baidu and found it could increase cabinet density by 20 to 40 percent just by populating the cabinet according to average draw, not theoretical maximum. The research is detailed in a Intel white paper.
On a larger scale, power management can be aggregated not just among individual racks in a cabinet but across the whole datacenter. Data Center Manager’s console lets admins set thermal and power policies for all systems, so if there are computers needing more power while others are idle, the idle ones give up their available juice for the ones that need it.