In-Depth
Reader Feedback: Readers Speak Out about Green IT Column
Readers share their reactions about Jon Toigo's analysis last week.
To the editor:
In his December 9, 2008 column, Jon Toigo made it sound like I wrote a misleading analysis about server electricity using bad data and incomplete assumptions. He said I ignored network and storage gear, relied on inaccurate forecasts from IDC, and based my power estimates per server on incorrect manufacturer numbers.
I understood perfectly well that network and storage gear use electricity, and noted that in my initial reports (http://enterprise.amd.com/Downloads/svrpwrusecompletefinal.pdf, http://www.amd.com/koomey). I focused first on servers because they are the largest portion of IT power in the data center, with the understanding that analysis of the other two components would follow later.
For the main part of my analysis I did rely on historical (not forecasted) server installed-base data from IDC because they have an extensive and widely used database of server installed-base data (even down to the model level). My colleagues there are perfectly capable of defending their data, but I found nothing in my research to indicate that those data were "mistaken" (although as with all data, there are some uncertainties buried within).
Finally, when I used manufacturer estimates of power use, I correctly adjusted for "typical" power use, consulted with experts throughout the industry, and benchmarked using multiple sources. I'm confident that the estimates I developed give a true picture of electricity use in the data center.
My analysis became one foundation of EPA's report to Congress (http://www.energystar.gov/datacenters), which added power used by networking and storage equipment to the totals (and made it clear that servers are indeed more important contributors to power use in data centers than either storage or networking). It also grew into a freely downloadable refereed journal article (http://stacks.iop.org/1748-9326/3/034008) and the work as a whole stands as the most widely cited source of estimates of power use for servers in recent years.
So, Mr. Toigo, please don't impugn my intellectual integrity with ill-founded criticisms. I wrote a whole book on intellectual integrity (Turning Numbers into Knowledge) and if I had done the things you claimed I did, I'd just be too embarrassed to leave my house.
Jonathan G. Koomey, Ph.D.Project Scientist, Lawrence Berkeley National Laboratory (not Livermore, as stated in Mr. Toigo's column)Consulting Professor, Stanford University
Mr. Toigo responds:
It troubled me when I saw Dr. Koomey's e-mail -- so much so that I did this follow-up work. My comments were based on Estimating Total Power Consumption By Servers In The U.S. And The World (http://enterprise.amd.com/Downloads/svrpwrusecompletefinal.pdf).
Mr. Koomey used IDC server install-base estimates extensively (see page 3, graphic on page 4, and pages 6 and 8).
Regarding my assertion that he reports power consumption by server type "as reported by vendors," on page 8, he writes:
Estimating power use for each server is not easy. The power use of electronic equipment varies with hardware configuration, usage, and environmental conditions. The power supplies for these devices are sized for the maximum loads expected when the server is fully configured, so the actual measured loads observed in typical installations are much lower than the rated power of the power supply.Unfortunately, measured data on energy use are not commonly made available for most servers, especially older ones (that is changing for more recent models -- see ASHRAE (2004) and the discussion of "Typical power use per unit" in the future work section). In addition, estimating power use is complicated because each server can have multiple disk drives, vary significantly in installed memory, and commonly have multiple processors and redundant power supplies."
I may have oversimplified his estimation of power consumption by server type, stating that he only used the manufacturer provided figure (the ASHRAE value), but he says he is working to get more accurate estimates in future work.
The bottom line is that in justifying my position, I would point to this on page 8: "I assign a power use per unit for each server based on measured data, on-line server configuration calculators, or estimates from manufacturer specification sheets." /p>
Regarding my assertion that he limited his study to servers, and did not include network equipment or storage, he writes on page 16: "This study focuses on the largest single component of data center power use: servers and the associated cooling and auxiliary equipment (C&A) needed to support them."
He also supplies a domain diagram showing what he excludes from the study, namely storage and network gear.
I stand by my column, but I will be speaking with Dr. Koomey soon. I think our readers would be interested in his thoughts about the results of the Federal and state regulatory activity around Green IT.
- - -
Hi, Jon.
I read your part 2 article today on Promises. I enjoyed it.
Being responsible for mainframe tape storage at my firm, I concur with your comments about Sun's inability to effectively get the word out about how green and cost effective their tape products are. In fact, I sent an e-mail earlier in the year to execs at both Sun and IBM encouraging them to get their act together and broadcast more into the market. I really hope they make some headway.
As we evaluated solutions this year, we continue to see tape as more cost effective for its workloads over the new "tapeless" disk-based DLMs from the disk vendors. Yet, I have had to spend time taking the arguments forward and needing to convince upper management that tape is still viable. The disk vendor's effective strategy has been to win them over. I'm still fighting some of the perceptions. (By the way, as another counter measure, I told Sun and IBM they need to get to my upper management and tout tape effectiveness directly to them.)
That being said, we're starting to see some workloads where these "tapeless" solutions could fit -- as a storage solution in and of itself and not strictly as a total tape replacement (spoiling the disk vendors' push that "their size fits all"). With a virtual tape look and feel plus mainframe connectivity, these solutions could meet the "tween" SLA needs for workloads such as customer statement storage -- faster than tape for a random recall but not as blazing as enterprise disk.
As they mature and incorporate their own unique SRM capabilities (such as movement of seldom-accessed data to spindles that spin down), more inexpensive compliance features, and more storage capacity, they will increase in attractiveness for certain workloads. Adding tape to get the "sludge" data out of the system may give the tape vendors the ability to tier the overall solution effectively and leverage their years of virtual tape experience, especially if a customer already has a tape infrastructure.
Anyway, you take care and keep pushing for SRM.
Regards,BT
[Editor's note: We are complying with the writer's request to publish only his/her initials, not a full name.]