Monday Jun 23, 2008
Last friday I blogged about an article on Duke University's Larrick & Soll's research:
Posting a vehicle’s fuel efficiency in “gallons per mile” (GPM) rather than “miles per gallon” (MPG) would help consumers make better decisions about car purchases and environmental impact, researchers from Duke University’s Fuqua School of Business report in the June 20 issue of Science magazine.
The main issue is that people usually make comparisons by linear improvement in miles/gallon, but this leads most to errors. Switching to gallons/mile (and as
I said for servers watt/performance) avoids these problems.
If one does the calculations correctly of course it doesn't matter, but on a quick look one can be mislead. For example, (do this quickly!) if one climbs
10 miles a hill and gets 10mpg and then coasts down the hill for 10 miles getting 100 mpg, how many mpg does one average? If you didn't come up with an answer of
18mpg (or nearly double the uphill rate), then you should consider looking at the reciprocal calculation.
If on that same hill that same car gets 1 gal/10 miles (10mpg) uphill and 0.1gal/10miles (100mpg), then it is easy to that coasting downhill can only come close doubling your fuel efficiency. Even if you doubled your fuel efficiency on the downhill section to 200mpg (0.05gal/10 miiles) you can see that your average fuel efficiency doesn't change much.
As I've said before on servers it is also critical to understand watt/performance on a wide variety of benchmarks, Sun understands this. This way you avoid benchmarks were vendors only highlight small-memory and low-GHz configurations.
Finally increase your server utilisation (even a small amount) and closely look at power-performance (watt/perf).
Thursday Jun 05, 2008
Intel's recommendation for 2-socket Xeon QC is 16GB of memory, I'm assuming 4-socket QC is 32GB when I look at this and other docs:
http://www.itworldcanada.com/documents/server-virtualization.pdf
Kingston memory suggests 4-way servers are best for virtualization, see:
www.kingston.com/gov/PDF_files/Kingston_Memory_and_Roadmap.pdf
Actually I've heard 64GB to 128GB being used for many companies for 2-socket to 4-socket consolidation servers.
This fits the memory sizes of most SPEC & TPC performance benchmarks.
What is the one benchmark that vendors only post results on 4GB to 16GB?
SPECpower_ssj. This seems very very very silly to me. Is this an example of vendor self-interest vs. good for the customer? Are vendors trying to not show watts at reasonable memory configs?
Why do I harp on SPECpower_ssj? because it avoids some major issues...
The challenge now ...is not processors. Or power supplies. Or storage. It's memory. Users simply want too much of it. The ratio we're seeing now is the memory taking over [processor power requirements 2 to 1." Roger Schmidt, chief thermal architect and DE of IBM server and workstation division - Computer World April 30, 2007
I think small-memory on servers is generally silly, I even know people who are buying 4GB laptops. You can't convince me that 4GB servers are the way to go for most commercial uses...
SPEC, SPECpower reg tm of Standard Performance Evaluation Corporation. Results at www.spec.org
Wednesday Jun 04, 2008
Servers are very different than laptops, DUH! Therefore one must benchmark power
in a very different way. We all know that laptops spend lots of time
at low utilisation or idle (waiting for your input). Most servers
should not be used at idle or anywhere near idle. Kick out any vendor
that only wants to show watt savings at low utilisation, they are trying to
get you to take your eye of the real truth.
In the good ol' wasteful days of yore, you could have your servers running at 20%, drive your huge SUV alone down the block to pick up buy an incandescent bulbs.
Today, you need to lower your costs and carbon footprint by turning off wasteful low-utilised servers, buying efficient servers and running them in at least an efficient part of the power curve, say 50-60% util or more.
Demand that every vendor add power measurement to every benchmark right now!
I fear many are redesigning benchmarks to add power only to emphasize low utilisation. In my opinion (I speak for myself not necessarily Sun), low utilisation measurements are just smoke and mirrors and disinformation. As I pointed out last week you can save many times more watts per unit of work by just raising your utilisation a bit.
Why do I mention this which appears obvious...
This morning I had breakfast with a friend. Always enjoyable catching up with friends. He mentioned that a customer of his wanted to make buying power decisions "based on the idle watts of a server"!?! The customer was prompted to ask that by a vendor. After a realistic conversation continued the customer now feels completely mislead by the other vendor. , you lost.
Now off to a tasty lunch...
Friday May 30, 2008
The SPECpower_ssj2008 benchmark goes out of its way to measure servers at low-utilisation (5 of 11 datapoints are in the wasteful active-idle to 40% range - the worst range for servers). SPECpower_ssj actually shows this. Customers needs to demand that power measured on all benchmarks, now.
Let's look at HP DL580 G5 4-socket (for SPECpower_ssj2008 HP only used the low-GHz 1.86GHz Xeon, tiny 16 GB with special DIMMs, and a hacked non-standard BIOS). Regardless it can still be used to prove my point.
Increasing utilisation even a small amount provides HUGE improvements in watts-per-unit-of-work. Increasing utilisation a tiny 10% improve your watt/work an amazing 33-163% ! If you increase utilisation more you save even more:
| %Util Improvement | Savings in watts-per-unit-of-work |
| 10 | 33-163% |
| 20 | 45-257% |
| 30 | 62-353% |
<| 40 | 84-437% |
| 50 | 114-511% |
| higher% | even more! |
The biggest savings occur when you stop running at low utilisation.
We need to do everything we can to discourage low-utilisation!
The first column, in the table below, is the %utilisation you start at, the rows than show you the %savings if you increase your utilisation. So for example if you were at 20% utilisation (2nd row from the bottom) and increase the %utilisation to 40% (3rd column from the right) you save 140% per unit of work!
| %u | Perf/Power | 100% | 90% | 80% | 70% | 60% | 50% | 40% | 30% | 20% |
| 100% | 929 | | | | | | | | | |
| 90% | 867 | 33% | | | | | | | | |
| 80% | 793 | 45% | 35% | | | | | | | |
| 70% | 712 | 62% | 50% | 37% | | | | | | |
| 60% | 626 | 84% | 71% | 56% | 41% | | | | | |
| 50% | 538 | 114% | 99% | 82% | 64% | 46% | | | | |
| 40% | 451 | 155% | 138% | 117% | 96% | 75% | 54% | | | |
| 30% | 357 | 223% | 200% | 174% | 148% | 121% | 94% | 64% | | |
| 20% | 243 | 374% | 341% | 303% | 264% | 224% | 185% | 140% | 89% | |
| 10% | 129 | 793% | 731% | 659% | 586% | 511% | 437% | 353% | 257% | 163%
| |
IBM bloggers accuse me of many things. LET ME BE CLEAR: These are my personal opinions and NOT the opinions of Sun. This blog is NOT the source for official opinions.
Disclosure statement
SPECpower_ssj2008:HP Proliant DL580 G5 (4-chip QC Xeon L7345 1.86GHz), 546 overall ssj_ops/watt,
359,523 ssj_ops and 387 watt at 100% target load,
325,931 ssj_ops and 376 watt at 90% target load,
291,991 ssj_ops and 368 watt at 80% target load,
255,512 ssj_ops and 359 watt at 70% target load,
217,222 ssj_ops and 347 watt at 60% target load,
180,262 ssj_ops and 335 watt at 50% target load,
145,079 ssj_ops and 322 watt at 40% target load,
110,173 ssj_ops and 309 watt at 30% target load,
71,409 ssj_ops and 294 watt at 20% target load,
36,070 ssj_ops and 280 watt at 10% target load, and Active Idle 271 watts.
SPEC, SPECpower reg tm of Standard Performance Evaluation Corporation.
Results from www.spec.org as of 12/11/07.
In a more realistic configuration the HP DL580 G5, from HP's own power calculators, a HP DL580 G5 with four QC Xeon 2.93GHz Tigerton and 64 GB memory should draw 1,072watts. HP DL580 power consumption from HP Power Calculator system configured with 4 x2.93GHz processors, redundant PSU, 16 x 4GB DIMMs, 8 x 36GB SAS drives,1 x PCI card, 80% utilisation on 9/10/07: http://h30099.www3.hp.com/configurator/powercalcs.asp
Friday May 16, 2008
Sun has shown watts on UltraSPARC for years!:
2008: UltraSPARC T2+
2007: UltraSPARC T2
2005: UltraSPARC T1 & T2000 blogs with power-performance
Sun is now showing measured watts with measured performance for X64 servers as well, the rumor mill inside Sun is saying we are going see a lot more soon:
2008: X64 Virtualization
2008: X64 Java
Real measured watts on a variety of workloads is critical to truly inform customers.
Friday Apr 18, 2008
The Uptime Institute has a whitepaper called "New Product Review: Self-Contained Computer Room in a Shipping Container from Sun Microsystems". You can find a link to this whitepaper at: http://uptimeinstitute.org/wp_pdf/(TUI3023)SunModularDataCenterOp_weblock.pdf
It talks about the design, their review, and the efficiency of the design.
Wednesday Jan 09, 2008
It is so easy to measure you wonder why some vendors hide their wattages.
Sun as US T2:
2007: UltraSPARC T2
2005: UltraSPARC T1 & T2000 blogs with power-performance
The only thing I can think of is that the numbers are embarrassing in
real configurations.
I a previous posting I mentioned that HP configurations used in other benchmarks have reasonable sized memory:
64GB: HP DL580 G5 (4 quad-core Xeon 2.933GHz)
64GB: HP DL585 G2 (4 dual-core Opteron 3GHz)
32GB: HP DL380 G5 (2 quad-core Xeon 2.66GHz)
32GB: HP DL380 (2 quad-core Xeon 3GHz)
32GB: HP BL685c (4 dual-core Opteron 3GHz)
32GB: HP BL460c (2 quad-core Xeon 3GHz)
Looking around at HP's own power calculators we find that an
HP DL580 G5 with four QC Xeon 2.93GHz Tigerton and 64 GB memory
should draw 1,072watts but how much does is actually draw?
We carefully measured one in our lab running database workloads at
60-80% util at just above 800watts. Fair, NO gaming. It is so easy to
do this right, try it yourself.
Note:HP configures 1200w supplies with their 4 Xeon QC DL580 G5. Everyone
knows that even with efficient power supplies you
design near the rating so that you keep the costs down and keep the
efficiency up. An engineer would be a bit daft to put a 1200w supply on a
server they expect to draw only 300watts! 
HP DL580 G5 power consumption from HP Power Calculator system configured with 4 x2.93GHz processors, redundant PSU, 16 x 4GB DIMMs, two GbE, four 146GB SAS disks, two PCI-E Dual-Port FC, 80% utilisation on 1/8/07: http://h30099.www3.hp.com/configurator/powercalcs.asp
Thursday Jan 03, 2008
Luiz André Barroso and Urs Hölzle of Google have written an article called:
The Case for Energy-Proportional Computing. You can find the IEEE Computing article here.
They correctly point out that memory (and disk) are a key factor in power used. Too bad that new SPECpower benchmark only favors tiny-memory configurations, and vendors do not publish measured watts on real-size memory configurations.
They miss the fact that low utilization wastes more power than energy-proportional designs will get them,
as we've pointed out
here
Disclosure statement
SPECpower_ssj2008. SPEC, SPECpower reg tm of Standard Performance Evaluation Corporation. See www.spec.org.
Wednesday Jan 02, 2008
Timothy Prickett Morgan in the IT Jungle missed the point a bit when writing about SPECpower_ssj2008.
It is not that you can't use power supply ratings to determine power usage. This difference really
proves that these are chopped configs that are being used just for making numbers on SPECpower_ssj2008.
Vendors who are trying to build power-efficient systems don't make mistakes like using small configs on power supplies that are too big. Power supplies are most efficient when the watts drawn are near
the power supply ratings. If HP thought these small-memory chopped configs were typical they would have put smaller power supplies in them.
Have power supplies that are ONLY at 22% load (269/1200) is smoking gun aimed right at showing this is not typical config...
"it is no surprise at all that the best metrics for this are coming from streamlined two-socket machines so far. An HP ProLiant DL160 G5 with two quad-core Xeon E5450 processors running at 3 GHz with 16 GB of main memory and a single 80 GB 7.2K SATA drive was able to max out at 308,022 ops/sec on the test; the average idle power draw on this machine was 160 watts and when running full speed (99.2 percent of CPU capacity), it drew 269 watts. This machine had a 1200 watt power supply, which tells you how useless these maximum draw ratings on power supplies can be in determining power usage."
http://www.itjungle.com/bns/bns121807-story01.html
Disclosure statement
SPECpower_ssj2008: HP Proliant DL160 G5 (2-chip QC Xeon E5450 3GHz), 698 overall ssj_ops/watt. SPECpower_ssj2008:HP Proliant DL580 G5 (4-chip QC Xeon L7345 1.86GHz), 546 overall ssj_ops/watt, 359523 ssj_ops and 387 watt at 100% target load, 255512 ssj_ops and 359 watt at 70% target load, and 71409 ssj_ops and 294 watt at 20% target load. Results from www.spec.org as of 12/11/07.
Tuesday Dec 18, 2007
A new SPEC benchmark called SPECpower_ssj2008 was announced last week..
I've taken a little time to look at the new results and benchmark rules.
Quite simply, I don't know why SPEC didn't add power measurement
to every benchmark. It is quite easy to add rules that
puts one SUT (system under test) on a power meter and a few rules
to accurately measure power. That should have been done months ago,
it is so much easier than creating a new benchmark.
So why didn't SPEC do this? I have no idea. I can only guess some
vendors to show off their power-management software at different utilisation
levels. As we've talked about here for a very long time, running servers at low-utilisation (below 50% utilisation) is the worst way in the world to waste power... Even with the most extreme power-saving HW/SW.
With a different benchmark only for power-benchmarking it is really possible
to game results. Did anyone?
- HP submitted a SPECpower_ssj2008 on the DL 580 G5 QC Xeon. Cool
now we can see what the DL580 G5 does. But HP PICKED 1.86GHz Xeon L7345 for
SPECpower_ssj and a very DIFFERENT processor (2.9GHz Xeon) for performance on SPECweb2005, well at least they are not confusing things...
- HP's DL 580 G5 SPECpower_ssj2008 documentation does NOT EVEN mention processor GHz or type!
Why?
- The only SPECpower_ssj2008 results were on 4GB, 8GB, 16GB configs.
Memory is a huge power draw. So why configure such small systems for power?
Why do these vendors use 32GB and 64GB configs on performance benchmarks?
- The HP DL580 G5 used one 60GB 5400 RPM SATA drive. My laptop has better
- HP's DL580 G5 Power supplies used are 2x1200 watts. The 1.86GHz
Xeon DL580 G5 draws 387 watts at 100% on SPECpower_ssj2008.
Since all vendors know that you use power supplies near their rating
for best energy efficiency, why does HP not sell properly-size power supplies?
Maybe because HP typically expects to sell configurations that are different
than SPECpower_ssj2008 DL580 configs?
- Is running at low utilisation efficient? SIMPLE answer NO!
Let's look at SPECpower-ssj2008 HP DL580 G5. Comparing 70% utilisation
with 20% utilisation we see that one wastes 3 TIMES more watts for the
same amount of work at 20% utilization than at 70% utilization!!! 3 = round(2.93) = (294w/71409ssj_ops) / (359w/255512ssj_ops)
- SPECpower_ssj allows default BIOS settings to be changed to make better
results by turning off prefetch (also see SPECjbb & prefetch) - yeah, right, every customer does that
get real.
So what serves the industry? Using the same configurations that you
benchmark for performance and power-performance.
Sun does this and has been doing this for years:
2007: UltraSPARC T2
2005: UltraSPARC T1 & T2000 blogs with power-performance
The only thing I really can take away from SPECpower_ssj2008 is that running at low utilization is silly, read carefully, and avoid silly configurations.
Disclosure statement
SPECpower_ssj2008: HP Proliant DL160 G5 (2-chip QC Xeon E5450 3GHz), 698 overall ssj_ops/watt. SPECpower_ssj2008:HP Proliant DL580 G5 (4-chip QC Xeon L7345 1.86GHz), 546 overall ssj_ops/watt, 359523 ssj_ops and 387 watt at 100% target load, 255512 ssj_ops and 359 watt at 70% target load, and 71409 ssj_ops and 294 watt at 20% target load.
SPECpower_ssj2008: Colfax CX2266-N2 (2-chip AMD Opteron DC 2216HE 2.4GHz), 203 overall ssj_ops/watt. SPEC, SPECpower reg tm of Standard Performance Evaluation Corporation.
Results from www.spec.org as of 12/11/07.
HP ProLiant DL580 G5 (16 cores, 4 chips) 30261 SPECweb2005. SPEC, SPECweb reg tm of Standard Performance Evaluation Corporation. Results from www.spec.org as of Oct 8, 2007.
HP DL580 power consumption from HP Power Calculator system configured with 4 x2.93GHz processors, redundant PSU, 16 x 4GB DIMMs, 8 x 36GB SAS drives,1 x PCI card, 80% utilisation on 9/10/07: http://h30099.www3.hp.com/configurator/powercalcs.asp HP DL385G2 power consumption from HP Power Calculator for system configured with 2 x AMD 2220 2.8GHz processors, redundant PSU, 8 x 4GB DIMMs, 2 x HBAs and 2 x 146GB SAS drives, 80% utilisation on 6/4/07: http://h30099.www3.hp.com/configurator/powercalcs.asp
Monday Oct 01, 2007
Topics such as space and power/cooling,that we write about here, are truly important for datacentres. The article below covers a survey by ONStor (a network attached storage company).
"ONStor Green Data Center survey reveals 63% of organizations have run out of space, power, or cooling capacity without warning
read more at:
http://www.byteandswitch.com/document.asp?doc_id=134163
Tuesday Sep 25, 2007
The BBC reports that "Large majorities in many countries now believe human activity is causing global warming, a BBC World Service poll suggests."
It goes on to say, "A sizeable majority of people agreed that major steps needed to be taken soon to address global warming."
More at: http://news.bbc.co.uk/1/hi/world/7010522.stm
Monday Sep 24, 2007
There is a ZDnet article that goes into issues with hosted datcenters called, ""Building a data center circa 2007"
It goes into power issues and now they are coming to dominate real estate (and performance) perspectives.
This clearly goes beyond hosted datacenters. All IT Management (or their bosses) should really focus on getting power budgets into IT so the right decisions are being made.
Thursday Sep 20, 2007
More good advice on datacenters was given by Mark Monroe(Sun) in at
eWEEK.com article entitled, "Finding the Economic Green in Green Computing" By Chris Preimesberger September 18, 2007. You can read about it at:
http://www.eweek.com/article2/0,1895,2184747,00.asp
In this blog we covered other important factors:
http://blogs.sun.com/bmseer/entry/the_register_needs_to_ask
and remember measure it in watts/performance, just like $/perf.
Friday Sep 14, 2007
I know AMD & Intel love to focus on the power used by CPUs
in their high-stakes battle to gain server chip dominance. Both
started talking TDP (Thermal Design Power) and getting people to judge
systems based on TDP.
...wait a minute, buying a system based on the power of a CPU is
a bit of misdirection unless the CPU is most of the power. This isn't
the case any more. So CPU TDP should be ignored, unless you are
designing your own product and are just buying CPUs.
First of all system power is what datacenters care about, so Intel and
AMD should be talking about system power in realistic memory configurations.
Memory draws lots of power these days.
Second,TDP was created so the manufactures of servers manufacturers
would know much power the chip consumes in worst-case maximum-power cases
so they could design power supplies, cooling, etc. That just isn't
useful to datacenter managers.
AMD's marketing only slightly improved the situation by telling
customers of SYSTEMS to look at the processor's ACP (Average CPU
Power).
Two problems:
- Focus on CPU power to avoid talking system power, but system power is
what one needs to know, in average case to estimate electrical bills.
- Focus on average CPU power not server maximums that datacenters need
to design cooling on (see: http://blogs.sun.com/bmseer/entry/watts_a_matter_with_their.
ACP of a CPU ...hmmm, do you know what a pain it is to just measure a CPU. AMD
in their whitepaper , has to isolate the power consumed by the processor
and that consumed by the motherboard -- this requires motherboard modifications and special instrumented server platforms!.
way to much work to get a marketing message, all we want is server power on a variety of memory configuations and full-speed CPUs and actually running at good datacenter utilizations!
WARNING: Everyone loves to talk performance of high-GHZ CPUs and low-power of low-GHz CPUs, so watch for the confusing marketing and much worse "bait&switch." Also watts per core is useless marketing, it is the watt/perf for a system that counts. Also any vendor trying to sell power-efficiency on high-performance
systems should report watt/performance along with their world record performance on that system.
|
|
|
|
4G in a laptop? looking at the price difference to...
I agree that 4G in laptop makes sense: nowadays mo...