IBM to Offer Automatic Power Throttling on Servers

Published: May 23, 2006

by Timothy Prickett Morgan

Innovation always lags necessity, and sometimes a tough problem that needs correcting takes a long time to solve. There are no easy fixes for the power and thermal issues that have been plaguing data centers since the dot-com boom. It has taken years for chip makers and server makers to not only face the problems, but to begin the task of engineering new machines that are not just more efficient, but actually have energy management as part of their core feature set.

Today, as part of its rollout of new servers based on Intel's "Dempsey" dual-core Xeon DP processors and their associated "Bensley" server platforms, IBM is announcing a new feature for its rack-based System x X64-based servers called PowerExecutive, which takes a slightly different tack on lowering the energy consumption on servers from what has been done to date.

According to IBM, electricity costs have nearly doubled since 1994, and many people believe that it won't be long (perhaps three to five years) before the cost of supplying electricity and cooling to a server will be as large or larger than the cost of the server itself. That's why Intel scrambled several years ago to completely redo its server processor line based on its Pentium M cores.

Laptops and notebooks have had advanced power management features for many years, which allows end users to configure their processors, disk drives, and screens to drop into a lower power state under certain conditions so they can prolong their battery life. There is no reason to run a processor clock or spin a disk drive at full speed if an end user is just sitting there, not doing anything but reading a document. These active power management features are, as the name suggests, proactively trying to reduce the amount of electric power that end users consume on the fly, but when end users start banging away on the machines, these components rev up to full speed and then they start consuming more power. Often a lot more power. Over the past few years, these power-saving technologies have been woven into some server designs, and they are now key features of the upcoming server processors from both Intel and rival Advanced Micro Devices.

The problem with such active power management features is that they do not think at the system or rack level, like data center managers do, and they do not allow managers to set absolute maximum power budgets that a server or a rack of servers can consume. That's what IBM's PowerExecutive feature is about. Think of it as "cruise control" for your servers, so you don't break the energy limit.

In the new Dempsey servers that IBM is rolling out today, which embody the second generation of the company's so-called Xtended Design Architecture for X64 processors and the System x (formerly xSeries) servers that employ them, Big Blue has created a set of electronics on the server motherboards and systems software that interfaces with it and can monitor actual power usage in a machine and allow system administrators to set energy consumption thresholds in the servers. The PowerExecutive tools allow administrators and data center planners to first figure out how much energy they are consuming as real workloads are running and plot that out over time. The data gathered by PowerExecutive would also allow something that is sorely needed in the data center: a reckoning of performance delivered and the electricity consumed. Because of the power curves (as in a logarithmically sloping curve) that govern electronics, sometimes goosing performance just a little bit can double the amount of electricity consumed and heat generated. This is why Intel has, in fact, shifted to dual-core chips and backed off the clock speed with the Dempseys. Intel can deliver twice the performance (at least on multithreaded applications) and consume a little bit less power compared to single-core Xeon DPs.

"People understand the power that goes into the data center," explains Stuart McRae, worldwide marketing manager for IBM's System x product line, "but they don't know what happens once the power gets into the data center." The System x division, working with scientists at the IBM Research labs, came up with features to add to servers that allow software to not just monitor power usage, but to also throttle it. These features were tested to a certain extent in IBM's BladeCenter blade servers, and are now being rolled into its rack-based servers as they are upgraded to new processor and chipset technologies.

While the new hardware supports the PowerExecutive 2.0 software, that software itself is not going to be fully ready until the third quarter, which is also well after the next-generation "Woodcrest" Xeon Core processors will be shipping in Intel's Bensley platforms. The current iteration of the PowerExecutive software, release 1.0, can monitor power usage, but it won't be until the 2.0 release ships in Q3 that the power throttling, either at the server or rack level, will be possible. It won't be until the "Tulsa" Xeon MP processors ship sometime in the second half of 2006 that servers using these chips will have full PowerExecutive functionality.

IBM is not charging for PowerExecutive, but rather using it as a competitive weapon against other server makers. It will have hooks into its Director systems management software System x servers, and will also integrate into broader Tivoli systems management programs and various server virtualization hypervisors to allow the interplay of dynamic logical and virtual partitions, service-level agreements for applications, and electricity consumption such that power usage is capped while server utilization is maximized across a bunch of servers.

McRae would not say when this PowerExecutive functionality would make its way into IBM's Power-based System i servers, which run i5/OS as well as Linux and AIX, or its System p servers, which predominantly run AIX but can also support Linux and i5/OS. The current generation of JS20 BladeCenter blades, which have 64-bit PowerPC processors, supports the PowerExecutive features, so the issue is not one of architecture, but going for the market that is most sensitive to heating and cooling. PowerExecutive is already supported on IBM's LS20 Opteron-based blade servers, and will presumably be added to the kicker to the rack-based eServer 326 server. Server sprawl has always been much more of an issue in the X86 and X64 markets, which is why IBM is focusing on blades and X64 at first.