Sustainable ComputingFiltering the Greenwashing
The Need for Numbers
Measuring Energy Usage
Measuring in Production
The Microsoft EnterpriseEngineering Center
There are an overwhelming number of products and solutions being marketed as green, as saving energy, as being more efficient. All this marketing hype creates confusion in the market as to what is really eco-friendly. Even after evaluating the specifications on various products, it is difficult, if not impossible, for an IT pro to
determine what equipment should be used when environmental impact is a key concern. When you see all the demonstrations, large energy savings are always highlighted, and this leads you to think that the return on investment (ROI) makes the upgrade easier to justify. After all, the energy savings should reduce the total cost of ownership (TCO).
The move by companies of all types to label seemingly everything as environmental and exploit the current interest in green solutions has led to the concept of "greenwashing," which refers to the over-promising of environmental benefits. So what is the truth of energy savings? This isn't as clear cut as, say, installing new energy-efficient lightbulbs in your home.
As interest in sustainable IT efforts increases and the market for environmentally friendly IT equipment expands, many people and organizations jump to the end result of deploying energy-efficient laptops, desktops, and servers, and using virtualization to reduce energy consumption. Yet few organizations run energy audits to determine the true benefits of what they have purchased.
While the ideal scenario is to actively measure in your production environment, that can also be expensive. If you aren't ready to start measuring in production, you can still move forward by performing your energy audit earlier in the process, doing so in your performance and evaluation labs. (Most companies have some lab or group responsible for testing and evaluating equipment before making a purchase.) There you can add energy performance as one of your test criteria and then take those results into account when making purchasing decisions, rather than relying on the numbers provided by manufacturers. So as you test, create your own device power-consumption database. Then you can ignore the greenwashing and see for yourself what works.
Of course, I should point out that if you want highly accurate numbers for operation under your true load, you'll need to monitor in your production environment. The quicker you begin monitoring your production environment, the better for your organization and your bottom line. This process will be critical to your long-term success.
The Need for Numbers
In an ideal situation, there would be independent tests that evaluate the performance per watt of hardware devices under realistic user loads. But the industry is in the early stages of developing and rolling out tests like these. And those now available, such as the SPEC Power benchmark, are still in their infancy. How well these energy tests will work once they're ready remains to be seen. But as with any test, the manufacturers will learn how to modify their equipment for optimal test results.
Experienced IT pros have learned to question tests conducted by testing labs. And just like car companies make sure they send their best cars for road tests, IT vendors will be sure to send their most efficient equipment configurations for energy tests.
Consider virtualization tests being used to compare a group of over-provisioned servers where no server consolidation has been run to demonstrate the benefits of virtualization. Vendors neglect to mention issues such as potential rebound effects from energy savings projects like virtualization. Essentially, this refers to when energy efficiency decreases costs and demand increases in response, driving up energy use and increasing use. When you see a solution, such as virtualization, marketed as a cure-all for energy savings, you should be suspicious. You really have to see how everything works together and what the ripple effects will be over time.
So how do you discover the right path for your organization? There are plenty of consultants ready to run assessments for a price. But there are some major downsides to this approach. Do you really want to pass along the expertise of being efficient to someone outside your organization? Do you want to rely on a consultant who wants to maintain a long-term contract and create a dependency on their service?
You could turn this into a huge project and get dozens of people involved to determine your environmental sustainability strategy. Avoid the temptation to just pick a spot in your environment and start measuring. While the sentiment is in the right place, the approach doesn't make sense, as this is just random action. You need to find the truth of where your energy consumption lies and where you can make an impact.
Measuring Energy Usage
You need to think in a new way about what is working. IT pros know when a piece of hardware isn't working if, say, a service goes down. But what about when a piece of hardware isn't working properly in terms of power consumption? This probably goes under the radar as most organizations have no data to establish their baseline energy efficiency.
How do you get these numbers? Some people suggest the need for an energy czar. That, however, requires someone who can navigate the company organization and is technical on power issues. A simpler, more approachable method is to add energy-measurement tools to existing functions in performance labs.
There are a range of energy-measuring devices. Unfortunately, at this time there is no perfect device for an IT performance lab. At the low end, you can go with a Watts Up Pro. This is a USB device that provides inline power measurement for 120V. Another device is the Smart-Watt, a networkable power consumption device with a 100-240 voltage and 15-30 amp range. Smart-Watt devices are also available with temperature and humidity sensors to measure environmental factors. Then there are the industrial power analyzing devices, such as the Extech Appliance Tester 380801 and Fluke 345 Power Clamp. You should have no problem finding devices like these to fulfill the needs of your lab tests.
If you're new to power measurement, it's important that you understand the power factor concept. The power factor of an AC electrical system is defined as the ratio of the real power to the apparent power, and is provided in a number between 0 and 1. Real power (watts) is the capacity for the circuit for performing work. Apparent power (VA) is the product of the current and voltage on the circuit.
You may be wondering why this is important. Take a look at Figure 1. The photograph shows an energy-monitoring device from Smart Works being used to compare the energy efficiency of a laptop, a lightbulb, and a capacitor. A lightbulb has a 50-watt load and 50 VA, for a power factor of 1.0. A capacitor with a pure capacitive load draws a 2-watt load and 193 VA, for a power factor of 0.01. The laptop is using 22 watts of power but has a 48 VA reading, for a result of a .47 power factor. The laptop's low power factor is due to an inefficient power supply design and was likely the result of manufacturing cost targets.
Figure 1 Power consumption for a laptop, a lightbulb, and a capacitor (Click the image for a larger view)
With so much interest in energy efficiency, vendors are now looking to improve the performance of power supplies under expected conditions. If you compare the power factor of existing equipment to that of new equipment, you may find you can gain back some power capacity simply by choosing solutions with more efficient power supplies and higher power factors.
Measuring in Production
The goal of measuring results in the performance lab is to predict performance in the production environment. The problem with measuring energy consumption in production is the cost required to measure all devices. One strategy, however, to reduce the cost of measuring power in production is to measure at the power distribution unit (PDU) and aggregate multiple equipment readings. Figure 2 shows a sample energy reading taken from a PDU with multiple pieces of the same equipment on one unit. Because average power consumption per server can be calculated, this means of measurement provides you with production power-consumption numbers.
Figure 2 Measuring at the power distribution unit (Click the image for a larger view)
As each piece of equipment is measured, the power information should be integrated into your configuration management database. If this is too difficult, you can create your own database or at least an Excel® spreadsheet that lists devices and their power consumption. As you start to accumulate more devices, you can fill in estimates for others to create an overall power-capacity calculation. Keep in mind that a new energy-efficient server will consume less than half its power load while idle versus at peak. And if you have old servers running obsolete solutions, the old servers will consume significant amounts of power even if they're idle. This is one of the easiest targets for regaining power capacity.
If you want to continue your improvement efforts, the next step is to calculate the power per rack being used in your datacenter. You must also be sure to understand your power and cooling capacities as you alter the environment. Over time, you will start to view your equipment in terms of its power needs and not just the space it requires. Space is an easy concept—it's static and visual. IT pros can readily discuss 1U, 2U, and 4U servers, but start talking about 200-watt servers versus 450-watt servers and you'll find the same IT professionals are not as comfortable with the conversation. This is a new language for IT equipment, and it's evident in today's datacenters. Many have plenty of space for more equipment even though they have reached their max power capacities.
The Microsoft EnterpriseEngineering Center
The Microsoft Enterprise Engineering Center (EEC) recently added power-measurement capabilities to its facility. Located on the Microsoft corporate campus in Redmond, Wash., the EEC (microsoft.com/windowsserver/evaluation/eec/default.mspx) is a state-of-the-art proving ground for the most complex computing environments.
With more than $40 million in hardware and networking equipment (see Figure 3), the EEC can tackle the most complex recreations of real enterprise production environments. The group partners with many leading networking, storage, and client/server solution providers to offer a mix of cutting-edge technologies and legacy platforms. The goal is to ensure each test provides an accurate reflection of the customer's current environment. The latest service being developed in the facility is reporting on energy consumption for solutions to provide a benchmark of performance per watt.
Figure 3 One of many rows in the EEC lab (Click the image for a larger view)
Over the past few years the EEC has seen many datacenter operators run out of power and cooling for their equipment. State-of-the-art equipment often comes with higher power densities, which put more stress on the facilities. The EEC staff has learned firsthand the costs and time involved in upgrading the power and cooling infrastructure. So to operate more efficiently and be more cost-effective, the EEC has added the ability to measure power per device.
Some of the tools and equipment being used by the EEC were not shipping when this article was written. Because this is a new solution, the EEC will continue to develop better techniques, working with customers, vendors, and Microsoft development teams. The EEC shares the methods it has been developing thus far directly with customers for early feedback, allowing the customers to create their own benchmarks.
And don't assume the results will be the obvious. The EEC has shared some of the interesting results the group has seen when using energy-monitoring features in the EEC performance labs:
- 1.Turning off a device doesn't necessarily reduce energy consumption as much as you might expect (see Figure 4). In one case involving server hardware, the EEC discovered a device that actually consumed 100 watts when turned off but still plugged in. This surprised many, and the EEC went over the setup many times. They eventually used an infrared thermometer to measure inlet and outlet temperature and verified that the device did, in fact, consume 100 watts while off.
- Software can have a significant impact on power consumption. On identical networking switches, with identical hardware and BIOS configurations, running different networking software displayed a 21 percent difference in power consumption. High-end solutions with more processes and features enabled, like security and monitoring tools, often consume more than their simpler low-end counterparts.
- In virtualization scenarios, the EEC has measured power consumption versus I/O utilization and CPU utilization to determine when a given piece of hardware maximizes its performance per watt. The EEC found that a narrow focus on CPU utilization could lead to too many virtual machines loaded on a physical machine, actually decreasing the overall performance per watt.
- Higher-density devices, as you might expect, have more power and cooling issues. When deploying higher- density systems, your power and cooling facilities staff should be consulted as early as possible. These devices may be good candidates for their own power-monitoring devices in production if you know the environment will be power constrained.
- Dual power supplies can consume considerably more power than a single power supply.
- Seemingly identical pieces of hardware with identical configurations can have significantly different power consumption. Observed differences were significant enough to make the EEC staff double-check hardware to ensure they were really configured the same.
- The watt ratings on the product plate are not actual consumption numbers, but rated capacity for power supplies.
- Maintaining a database of energy consumption tests and results per device and subcomponent is essential for retaining knowledge and comparing data.
- Different configurations of equivalent amounts of RAM consume different amounts of energy. Fewer DIMMs typically consume less energy—for example, 4 x 2GB DIMMs versus 8 x 1GB DIMMs. But there have been some cases where fewer DIMMs consumed more energy.
Figure 4 Power-consumption comparison of on versus off (Click the image for a larger view)
By adding the capability to measure power consumption in your performance labs, you can start to accumulate your own database of energy consumption per device with accurate numbers for your real loads. When you deploy those solutions, you should audit the results in a closed-loop feedback to determine the accuracy of your performance lab tests. As the EEC learned by running its own measurement tests, there are some really interesting details you can look for in order to help you filter out the greenwashing and discover the truth.
This method of measuring in performance labs will influence the market overall to begin identifying what truly is most energy efficient. And over time manufacturers will find it necessary to provide more accurate data about their energy-savings claims. Likewise, in time IT professionals will talk about watts per device as commonplace purchasing criteria. This needs to be a common practice, especially for companies that buy large quantities of servers. However, you can lead the charge by starting today to think about power as one of your valuable IT resources.
Keeping Cool in the Datacenter
Datacenter cooling offers a huge potential for reducing energy consumption. It is astounding how much heat can be generated in a datacenter and how much energy is used to keep hardware cooled. But if you want to manage your cooling successfully, fix problems, and develop more efficient cooling solutions, you'll need a temperature-monitoring solution. Consider the solution the Microsoft datacenters use.
Microsoft Research built a temperature sensor network for the datacenters that allows for improved temperature control and also enables evaluation of various cooling improvements. For instance, one Microsoft datacenter was evaluating end-of-aisle air curtains to improve hot and cold air separation. After the curtains were installed, some servers started to send overheat alarms. Naturally, the operation engineers increased the air flow from the cooling system to provide more cool air. To their surprise, however, more servers sent overheat alarms. And all these servers were at the bottom of the rack—the bottom, of course, is usually the coolest area from a raised-floor cooling system.
Using the sensor network, the engineers confirmed that the racks were cooler up higher, with the bottom of the rack the hottest. And they soon figured out that hot air was being drawn from the hot aisle between the bottom of the rack and the flooring—a result of Bernoulli's principle. They easily fixed the overheating by sealing the bottom of the rack and reducing the air flow speed.
This is just the sort of data the Microsoft Enterprise Engineering Center gathers and analyzes when doing performance testing. So the EEC recently notified Microsoft Research that they were ready for a deployment test. Within a day the system was deployed to 10 racks, and the installation took just one hour to complete. The EEC now is able to study and better understand cooling issues and their relationship to hardware performance.
Of course, simply monitoring isn't a solution in itself. The real gain is in your ability to find problem areas that you can fix, make changes, and evaluate various solutions to see if they have the result you are expecting. After all, you don't want to be caught off guard when your new cooling solution unexpectedly causes your racks to overheat.
Dave Ohara has 26 years of experience in technology. He is now working with multiple companies implementing green initiatives.
© 2008 Microsoft Corporation and CMP Media, LLC. All rights reserved; reproduction in part or in whole without permission is prohibited.