Strategies for efficiently powering and cooling modern high-density data centers
Data centers are critical to our daily lives, but they continue to get bigger and more power-hungry. Now we want them to be green as well. The good news is that energy-efficient design can create a better environment for the computing equipment and reduced energy costs — they don’t have to be energy wasters, despite the fact that they require substantial amounts of power. Specialized power and cooling equipment is available that gets the job done, but special techniques are required to get the most out of it.
Legacy techniques, if properly applied, cool approximately 5kW per rack. The problem is many cabinets today are drawing 15kW to 40kW, which may soon rise to 60kW. It’s easy to see why overheating is such an issue. It’s also obvious that these densities can’t be addressed with conventional air conditioners.
The old practice of ringing the room with computer room air conditioners (CRACs) doesn’t work across the board. Placing CRACs at right angles causes uneven air flow and creates under-floor vortices that reduce pressure and cooling effectiveness while consuming unnecessary energy. However, CRACs still have a place. They should be placed perpendicular to cabinet rows and aligned with the hot aisles, unless specific methods have been employed to isolate and control the return air. That may seem counter-intuitive, but when no other means of air control is employed, this placement minimizes the amount of cold air that a CRAC can pull back into itself
A boatload of new tools is available on the market to handle every situation efficiently, including the ability to cool really high-density equipment for high-performance computing. The key is to know what the various systems are and when to use them. There are two basic principles for energy efficiency that apply whether cooling is done from below a raised floor or overhead:
• Keep hot and cold air from mixing.
• Maximize the temperature of the air returning to the air conditioners.
Mixing results in warmer air being delivered to the servers and cooler air returning to the air conditioners. You can avoid mixing by controlling return air, which results in more efficient equipment cooling, higher cooling capacity from the air-conditioner coils, and less concern about humidity and condensation. Return air can be channeled with overhead ducts, via a ceiling plenum, by “containing” either the hot or cool aisles to keep air from mixing or a combination of these.
The highest efficiency is achieved with close-coupled or source-of-heat cooling. This means cooling units are installed with the computing equipment right next to it. This greatly reduces fan energy needed to push air under the floor or through ducts, delivers the air right where it’s needed in the right quantity at the correct temperature, and pulls the hot air back in before it has a chance to go anywhere else.
Close-coupled cooling can be done with in-row or overhead cooling units, with rear door pre-coolers, or a combination of approaches. Some of these devices circulate chilled water or condenser water through the cooling units; some use refrigerant. However, all are superior to conventional cooling, if deployed
appropriately. Beyond these methods are systems that circulate coolant directly to the processors — or those that immerse the servers in cooling fluid.
There is no longer any reason to keep the data center igloo-cold. Substantially more energy is saved by operating in accordance with the “Thermal Guidelines for Data Processing Environments” published by ASHRAE TC 9.9. This document allows equipment intake temperatures up to 27°C (80.6°F) on a continual basis, and even higher for a few days if necessary, without voiding warranties or significantly increasing fan speeds.
The 27°C limit was chosen because above this temperature fans speed up dramatically, resulting in significant energy waste. (Doubling fan speed draws eight times the energy.) The increased power draw of thousands of server fans could quickly offset the savings from higher temperature operation. Good practice today is to control air delivery to around 75°F — much better than the 55°F air that was the standard for decades.
This higher temperature operation is suitable for legacy and new computers. It enables more hours of free cooling — the use of outside air instead of mechanical refrigeration to remove heat through either water-side economizers (using the air to remove the heat from circulating water) or through air-to-air heat exchangers (air-side free cooling).
Water in the Data Center
Virtually all high-density systems use water somewhere in the cooling chain. It may be run directly to the cabinets or hardware, or circulated through control units, which then distribute refrigerant to the cooling devices. Water is far more efficient than air in removing heat. As temperatures continue to rise, the use of liquid in the data center will become more prevalent. Although this prospect gives most IT people angina, it should not be a concern so long as the piping is designed and installed correctly, and leak detection is adequately employed. Pipe leaks are actually very rare.
It is likely that many data centers will eventually have equipment that requires water circulation, if they don’t already. Designing piping with extra connections in strategic locations means the data center is ready for rear door coolers, water-cooled cabinets, directly cooled servers, or whatever new form of cooling comes along.
Heat Originates with Power
While cooling may offer the greatest opportunities for energy savings, nearly all the heat in a data center is the result of power consumption, so power reduction must also be considered. Data center power systems are mostly concerned with reliability, which still trumps efficiency in most cases, even though that is no longer necessary. The first “power” issue relates back to cooling design.
Motors use a great deal of electricity, particularly those on fans, pumps, and chillers. There are two fundamental solutions available: variable-frequency drives (VFDs) on fan, pump, and chiller motors; and electronically commutated (EC) fan motors. VFDs adjust motor speed to match cooling need as determined by sensor information, so devices don’t run at full speed when demand is less. Through the use of VFD control, it is often possible to run every part of a redundant system at lower speed under normal conditions, and actually use less energy than the older approach of activating only the primary units and leaving the redundant ones shut down until needed. This method also ensures that redundant units are always operational, and that there is no time lag for redundancy to take over when a primary unit fails. EC motors (which are becoming common on air-conditioner “plug fans”) can also reduce energy consumption by as much as 30% over conventional motors because of their more efficient design.
A major and historic energy waster is an oversized uninterruptible power supply (UPS). Most UPS units in use today are of the double-conversion design, which means incoming alternating current (AC) power is rectified to direct current (DC), which charges the batteries and is then converted back to AC through an inverter. Power is lost as heat through each step in the change process, and further losses occur through every transformer in the power chain.
Although UPSs have been designed more efficiently in recent years, even the best units are rated at only 95% to 97% at full load, which means they waste 3% to 5% of the power delivered to them. (That’s as much as 50kW on a 1MW system or 1,200kWh per day.) If UPSs run in the lower part of their capacity range (in the order of 30% of rated load), their efficiency can drop dramatically — into the 80% efficiency level or below range. This problem is particularly acute when running redundant systems, because a “2N” installation requires that each half (UPSs, PDUs, etc.) be loaded to no more than 50% of capacity so that either can assume full load when needed. But if UPS systems are grossly oversized to begin with (usually done because of poor load estimates or to accommodate theoretical long-term future growth), it is easy for actual usage (particularly on a redundant system) to drop down as far as 15% to 20%, which results in very low efficiency and enormous energy waste and cost.
However, there are good solutions to this dilemma:
- “Right size” the UPS using systems that enable incremental growth.
- Use newer “intelligent line interactive” UPSs that can reach 98% to 99% average efficiency.
There are also systems being run on high-voltage DC to avoid double conversion completely, although those are still somewhat controversial in many circles.
Regardless of UPS type and configuration, the computing hardware should be run on higher voltage. At 208V, the power supplies run more efficiently. In addition, fewer conductors are used, and current draw is reduced, which translates into the use of less copper. The only drawback is more challenging phase-balance, because each load appears on two of the three phase wires. Incorporating good monitoring can help solve that problem as well as achieve more efficient operation and realize maximum capacity and efficiency from the UPS.
McFarlane is a principal of Shen Milsom & Wilke LLC, heads the data center design and consulting practice, and is considered a leading authority in the design of the physical space and its power and cooling. Based in New York, he can be reached at firstname.lastname@example.org.