Jul 31 2008

Beyond the Hype

Agencies can gain efficiencies and save money through monitoring and managing IT power use, so find your patch of green.

Green IT has become such a hot topic that there’s beginning to be a backlash from techies tired of hearing about it as the new, new thing.

The most surprising comment: “Does green IT have anything to do with energy efficiency?” This just goes to show that the real meaning behind the IT power measurement and management movement is getting lost in the hype.

Illustration: Elizabeth Hinshaw
George Ou

Whether it’s through government mandate or corporate responsibility, green IT should be about saving power and getting a return on investment. But that requires gathering some knowledge first and then acting wisely on that information. There are also different dynamics at work when looking for and making changes on the client and server sides of the equation. But in either case, there are some common denominators: Learn and understand the metrics of power; consider what today’s systems can provide in the way of energy efficiency; go for the obvious by reducing consumption when no one’s using the agency’s systems; and evaluate technology options against your processing demands.

On the Client Side

Measuring Computer Efficiency

Figuring out how to be green has been a challenge for most organizations because the field is relatively new and the metrics for measuring energy efficiency are lacking.

Currently, one of the key metrics that the Energy Star rating focuses on is the 80 Plus rating, an energy-saving incentive program funded by electric utilities for power supplies (80plus.org).

That metric alone really can’t tell you how efficient a computer is. It often can’t even tell you how efficient the power supply is because the power supply is too large for the computer it’s powering. The 80 Plus rating calls for power supplies to be at least 80 percent efficient at 20 percent, 50 percent and 100 percent output loads, but desktop computers and servers have taken such a dramatic drop in power consumption that they’re loading larger power supplies at only 5 to 10 percent.

An energy-consumption metric in the absence of work performed is relatively meaningless because 100 150-watt low-performance servers are far less efficient than 10 300-watt servers that can do the same work as those 100 servers. The Standard Performance Evaluation Corp. created the SPECpower standards to craft just such a performance-per-unit energy measurement. The first version of SPECpower, called SPECpower_ssj2008, made its debut in December (www.spec.org/power_ssj2008). It measures Server Side Java operations per second, per watt (or operations per joules).

In May, SPEC began working on a SPECweb version of SPECpower that tests web server operations per joule, and eventually the SPECpower Committee plans to issue versions of SPECpower for workstation performance.

SPECpower offers fine-grain power measurements from 0 to 100 percent loads in 10 percent increments and identifies the work that can be done per unit of energy at those load levels (www.spec.org/power_ssj2008/results). Even if your organization is not specifically interested in Server Side Java metrics, the power measurements are still useful because they give you a ballpark figure on peak power consumption. The idle power measurement is relevant to all workloads because there is no work at 0 percent, but peak power measurements may vary because your specific workload may stress the CPU and memory subsystem differently than Server Side Java.

Ideally, you should conduct your own power measurements and create your own performance benchmarks on your actual workload to help with your purchasing decisions because no standardized benchmark can be representative of all workloads.

Evaluating Client Computing Efficiency

Although performance per watt may be important to the power user, the vast majority of office workers need machines that work well enough. That means for general-purpose office computing, the most important metric is power consumption. To get the least power consumption on the client side, notebook systems are natural candidates because they were designed with long battery life in mind. But notebooks may not be what your organization standardizes on, so energy-efficient desktops remain important.

In 2006, chipmakers began producing very efficient CPUs. In 2005, Intel’s Pentium 4 and Pentium D CPUs shot power-consumption levels into the stratosphere: A typical Intel desktop computer could consume 200 watts while idle. AMD, meanwhile, led the desktop and server race. Intel reversed the situation in mid-2006 by closing the gap and upping the ante on energy efficiency.

Today, both companies can boast 65-nanometer, dual-core computers that operate at less than 50 watts when idle. Intel chips can maintain peak system power consumption levels below 95 watts; AMD peaks above 122 watts. But even more amazing is a high-performance Intel 45nm, 2.83-gigahertz X3360 quad-core desktop or server that can hit 53 watts idle and 107 watts peak. How that translates in terms of power is that Intel’s 45nm, quad-core technology has reached power-consumption levels equivalent to its 65nm dual-core technology and lower than AMD’s 65nm, dual-core technology.

Enabling and Managing Suspend State

Because many office computers get only eight to 10 hours of use on weekdays and none on weekends, what an organization does with an unattended computer is even more critical to saving energy.

The first thing to check is whether S3 suspend state is supported on the hardware. That means the motherboard must support S3 suspend state. S1 suspend state is almost worthless because the power consumption is only slightly better than a powered-on idle state. For example, there are older P4 systems that only operate in S1 suspend state that “sleep” at 150 watts.

By comparison, the S3 sleep state consumes roughly 2 watts of power. Not only do you want a motherboard that supports S3 suspend, but you need to make sure that that the system’s BIOS has it enabled. The operating system must also support S3 suspend and be enabled for it to be effective. Last, the drivers and applications loaded on the computer must not interfere with entry to or exit from S3 suspend state.

Trying to set power-saving states manually is highly impractical, so find ways to do it through the OS.

With Microsoft Windows Vista, you can configure group policies from Active Directory. That’s not the case with Windows XP and 2000, but the Environmental Protection Agency has created a tool called EZ GPO that solves the shortcoming in XP and 2000 and lets you maintain power-management settings centrally. It can be downloaded at www.energystar.gov/index.cfm?c=power_mgt.pr_power_mgt_ez_gpo.

The flip side of managing suspend state is handling wake-up requests. Many system management programs, such as Microsoft System Center Configuration Manager (formerly SMS), have built-in functionality that can perform remote wake-ups. Your IT team needs to use these programs. If it’s not using SCCM, SMS or some other management system to support remote wake-ups, there are free utilities that can be scripted.

On the Server Side

Measuring the Right Thing

Server efficiency differs from client-side efficiency because the interest is not how little power a server consumes but how much work can be done for a given amount of energy. For this, the SPECpower benchmarks are the best in the industry.

The only time an organization will care simply about power consumption and not performance is when running nonvirtualized servers with limited workloads, but that’s wasteful because such servers are probably ideal candidates for consolidation through virtualization. If it’s not possible to consolidate these servers for some reason, try to use single-socket, low-power servers that consume less than 50 watts when idle.

Servers that have low typical workloads but high, time-sensitive peaks can’t be put on the lowest power servers with limited performance levels, so the goal then is to use servers with high performance-per-watt ratings.

Taking Advantage of Virtualization

There is a saying that the most energy-efficient server is the server that doesn’t physically exist. This is the promise that virtualization holds, and it’s something that most organizations have yet to adopt widely.

When selecting virtualized servers, the defining criteria should again be high performance/watt ratings because the higher the server performance, the more virtual servers it can host. Theoretically, AMD’s Barcelona-based quad-core CPUs currently have the upper hand in virtualization because they support nested paging, but no production virtualization solution at this time supports that feature.

Reducing Your Footprint

Blade servers generally have the highest efficiency because of their shared power supplies and cooling systems. The problem is that they have more limited configurations and proprietary backplanes, and most server farms are not using them yet.

That’s not to say that your IT department should avoid blade servers if they meet your requirements; organizations should take a serious look at them. These servers not only save power, but they allow higher densities, which give you vastly more efficient data-center cooling solution options, such as in-rack liquid cooling.

Trying Different Routes

Emerging options to blade servers are servers that plug into the rack infrastructure with in-rack DC power distribution for airflow and power consumption superior to traditional server racks.

The racks themselves take AC power, but redundant DC power distribution to the servers can save a lot of power compared with the inefficient dual-power supplies in traditional servers.

Another option in this space is the twin server. These 1U 1.75-high servers contain a single power supply and two dual-socket servers in the same chassis. That means it’s possible to have 80 dual-socket eight-core servers in a single 40U rack, which is about the same density as a blade server.

For many organizations, running twin servers is an affordable alternative. But with any high-density solution that has 80 servers in a single rack, the peak power consumption will hit 20,000 watts, which requires some type of active airflow at the rack level in addition to traditional raised-floor cooling.

Regardless of whether it’s fashionable or not, saving energy and money is simply good business, and it’s savings that will last. Because older technology often uses excess energy, it’s fairly easy to reduce consumption without compromising user experience or spending a lot of money.