High Performance Computing Vendors Gain Energy Advantage

SGI is a high-performance computing (HPC) hardware manufacturer based in Hayward, California, the result of the 2009 acquisition of the original Silicon Graphics, Inc. by Rackable Systems. For several years, the company has had an asset-management platform called SGI Management Center. Now, for select products, Management Center embeds Intel Data Center Manager (DCM) technology, opening up a range of new capabilities that makes the SGI offer more compelling.

SGI’s Data Center Innovations

SGI products are used in some of the most advanced calculations taking place on Earth. Astronomers at NASA and biologists working on modeling genes at the Technical University of Denmark are among SGI’s customers. In addition to traditional HPC products, the company also offers an innovative set of rack-mounted servers and storage for large data center clients.

Energy management has always been a top concern for SGI. One of the innovations of the legacy Rackable Systems side of the business was the development of “half-depth” servers, which make efficient use of rack space by consolidating cooling air and heat release through a central chimney, rather than isolate the “cool” and “hot” sides of the units, as many other vendors do, according to David Sundstrom, director of product marketing at SGI.

Traditionally, both groups of customers had energy concerns, although the HPC customers were more interested in “breaking the laws of physics”—trying to get as much power and performance out of the machines as possible. But power has become more expensive, particularly in Europe, and computing demands have begun to outstrip the capacity of the power supply for buildings and the capabilities of utilities that support them. About two years ago, SGI began to see a convergence in demand for power management between its enterprise data center and HPC customers. In both cases, “the systems now have the ability to draw power in excess of the building power envelope,” Sundstrom says. “Power is now the dominant operational cost of very large systems.”

Many customers, particularly those in the HPC market, dealt with the issue by running higher-powered computations at night and throttling back on consumption during the day, when power is more expensive. The SGI Management Center (MC) administration platform provides real-time analysis of more than 120 metrics for each server and provides a common interface across multiple architectures in the SGI family. Users can also manage tens of thousands of server nodes from a single point of control, including powering the servers up or down, in what Sundstrom calls “coarse-grained power management.” In other words, servers, or entire clusters of servers, can be turned on or off quickly and with a contextual, consolidated level of understanding.

The disadvantage of this approach was the time it took to reboot servers. This could take several minutes—assuming that no issues during reboot—and thus was not ideal for highly variable environments, which sometimes require workloads to be shifted in seconds.

This changed in fall 2010, when SGI MC began to embed Intel DCM.

The Role of Intel DCM

Intel DCM offers partners such as SGI the ability to assist with power capacity planning by measuring energy consumption and usage by device. Intel DCM also enables partners to measure the inlet temperature of devices, by improving the thermal profile of the data center and allowing partner solutions to proactively identify and rectify potential failures.

In addition, Intel DCM allows dynamic control of machines through real-time connections. By incorporating Intel DCM, SGI now offers “fine-grained, dynamic power management” in the SGI MC for a number of its Intel–based servers, Sundstrom says.

Data center customers can now apply policies to throttle servers, which will reduce clock cycles and thus power demand, rather than switching them completely on or off. This provides many more options for apportioning tasks to machines or groups of machines, based on power and computing demand constraints. Combined with SMC’s capability to dynamically manage logical and physical machine clusters, users now have a much more complete and agile level of control over their data center landscapes.

“It’s a key breakthrough that really gives us a completely different option for managing power that we didn’t have before,” Sundstrom says.

Benefits of Embedding Intel DCM

Using SMC with Intel DCM means data center managers can aggregate and control servers in ways that are unique to their specific environments.

For managing physical clusters, data center managers might want to compare the power consumption of one rack versus another, or individual machines against each other, or troubleshoot hotspots in the data center.
For logical clusters, managers can also slice and dice their understanding and control across a group of servers organized to support a calculation or a particular application, or compare the relative consumption of storage, administration, or calculation nodes.

One SGI customer running Hadoop was in danger of exceeding its power envelope in the data center. Using SMC with Intel DCM, the customer is now able to scale the power states of each server and distribute the workload proportionately, providing a consistent level of throughput while staying well clear of the power ceiling.

Because it was able to leverage Intel’s substantial research and development efforts, instead of trying to develop a similar capability on its own, SGI is now able to compete more effectively against much larger vendors that have power-management solutions as part of their offers.

“Between our hardware, including our servers with Intel motherboards and power-management enabled chipsets, and our joint software set—SMC incorporating Intel DCM—we are able to drive a much more cost-effective solution for power management,” Sundstrom says. “Now there’s another game in town.”

As part of partner applications such as SGI Management Center, Intel DCM adds significant value, both for data center managers looking to get a handle on the energy profile of their data centers and for hardware vendors looking to increase the value of their offerings.

This blog was written by CITO Research and sponsored by Intel Data Center Manager.

Related Articles

Seeing Is Believing in Data Centers

Eliminating Massive Amounts of Energy Micro-Waste

Power Management at the Data Center: From Capacity to Control