BM Seer Unofficial thoughts from an anonymous Sun employee

SPECpower_ssj too many measurements?

Monday Jun 09, 2008

A posting last week, clearly demonstrated that even small increases in utilisations provide HUGE savings".

Then I started looking at the data, it seems that one only really needs two points(!) {active-idle & 100%) to determine the watts used at any utilisation. Let's take a look at the HP DL580 SPECpower_ssj result.

%utilMeasured
Watts		Linear
Predict		watts
diff		%Diff
100%387w--0%
90%376w375.4w0.6w0%
80%368w363.8w4.2w1%
70%359w352.2w6.8w2%
60%347w340.6w6.4w2%
50%335w329w6w2%
40%322w317.4w4.6w1%
30%309w305.8w3.2w1%
20%294w294.2w-0.2w0%
10%280w282.6w-2.6w-1%
idle271w--0%

I'll look at more at at other SPECpower_ssj results. But it seems that SPEC should just simply add idle watts and wattage measurements at 100% utilisation to ALL SPEC benchmarks and not redesign benchmarks to measure watts at 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, and 90%. In the worst case, above the linear prediction was ONLY 2% different than actual watts!

I have long said SPEC should just at watt/perf to all of their benchmarks as currently designed.

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

[12] Comments
Like this post? del.icio.us | furl | slashdot | technorati | digg
Posted at 05:07AM Jun 09, 2008 in Benchmark Issues
Tags:
Comments:

I've heard internally that this is a "duh" and is very well known inside Sun. But I haven't heard anyone talk about it externally.

Posted by BM Seer on June 09, 2008 at 11:13 AM PDT #

Yeah, we also noticed this. But it only applies for today's hardware; future systems may be different.

Posted by Wes Felter on June 09, 2008 at 12:35 PM PDT #

hmmm... so the additional utilisation levels were added to quantify these "unknown non-linear" future systems, but you avoided adding watt/perf to every current benchmark to avoid showing the *known big effect* of the power draw of memory?

Posted by BM Seer on June 09, 2008 at 02:21 PM PDT #

I'm not on any SPEC committees, so I didn't do anything. These are certainly the early days of power benchmarking; maybe things will get simpler in the future as we begin to understand what matters and what doesn't.

Posted by Wes Felter on June 09, 2008 at 02:32 PM PDT #

You can't draw any conclusions from the published data except that the shape is nearly linear for 2 vendors' processors that have extensive power management capabilities. The curves are probably considerably different for processors and operating systems without power management (e.g. Sparc, Niagara, Solaris).

Posted by blogger on June 10, 2008 at 05:56 AM PDT #

I went through all of the data even with "extensive power management" all curves are linear. If anything "extensive" would mean most opportunities for linear curves (you know exotic weird things kicking in at different %util - but this just isn't happening).

What is huge is the number watts drawn buy non-SPARC systems at 64GB or 32GB, let's see those published results on standard BIOS, regular dimms, same procs benchmarks on other tests, etc.

Notice the whole point of the previous posting was that your so called "extensive power management" saves MUCH less than tiny changes in utilisation.

SPARC, UltraSPARC T1, T2, T2+ all have power management and Solaris (even the OpenSolaris distro) has power management as well. The curves no different, linear. Clearly bl*gger, your are trying to attack based on no info.

How about some real info:
The only thing extensive on published results "unique" configs:
http://blogs.sun.com/bmseer/entry/configs_used_for_specpower_ssj
http://blogs.sun.com/bmseer/entry/chopped_configs_and_specpowerssj2008
http://blogs.sun.com/bmseer/entry/specpower_ssj2008_power_benchmark_needs
...look at tags specpower, watts, wattage, etc. for lots more...

SPEC, SPECpower reg tm of Standard Performance Evaluation Corporation. Results from www.spec.org as of 12/11/07.

Posted by BM Seer on June 10, 2008 at 09:29 AM PDT #

Wes: sorry I put "you" but you used the phrase "we also noticed this" so I thought you were speaking for some collective involved in the benchmark.

In a previous comment you said: "why don't you join SPEC and try to overthrow it from within?"

So I sensed some connection, my apologies.

Also another comment on that, why shouldn't I broadcast out ideas (like an american) to get lots of people (300-1000 a day) thinking about the important issues instead of hanging out in a small committee where I couldn't discuss opinions in public?

Posted by BM Seer on June 10, 2008 at 09:37 AM PDT #

Sorry I wasn't clear in my previous comment. Some of my coworkers and I have a hobby of reading and analyzing the public SPEC scores.

I agree that it's good to discuss these things in public, although I am skeptical about how much influence us random bloggers can have on SPEC.

It is interesting that a few years ago people "knew" that power-performance curves would be quadratic or cubic but in reality they are pretty linear.

Posted by Wes Felter on June 10, 2008 at 12:02 PM PDT #

In my mind too much focus on power curves on "unique" configs and not enough getting power on a wide range of configurations at different memory sizes and processor GHz.

Carbon footprint is only altered slightly by exotic "extensive power management capabilities" Look above the watts used is _only 13%_ less going from 70% util to 30% util. Now if you can raise your %utilisation
from 30% util to 40% util in a dataceter of the above serers you could cut your power draw _64%!_

I'll be thinking about that we I head off into the heat-wave that is baking my city.

SPECpower_ssj2008:HP Proliant DL580 G5 (4-chip QC Xeon L7345 1.86GHz, LV DIMMS 1 tiny disk,...), 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.

Posted by BM Seer on June 10, 2008 at 02:29 PM PDT #

I can understand how products without "extensive power management" (don't you think "extensive" works as any number compared to 0) will look linear - they will have nearly the same power consumption at idle as at full load. How about some links to Sun's ">0" power management in any processors or OS? I read elsewhere that Sun has neither.

And speaking of "based on no data", what do you expect when Sun won't publish any SPECpower_ssj2008 numbers? How can anyone discuss Sun products as regards power without SPECpower_ssj2008? I know your answer already - the chant goes on - everyone must be running at 100% load all the time.

Posted by blogger on June 10, 2008 at 05:56 PM PDT #

I don't know when Sun will publish any benchmarks on tiny memory configurations, hack a the BIOS to turn off standard features, use a single 60GB drive, turn off redundant fans, use special dimms, use expensive low GHz CPUs, and always point people to 10%utilisation.

"I read elsewhere that Sun has neither." -- then you aren't forming your opinions from any truthful data.

Posted by BM Seer on June 11, 2008 at 09:57 AM PDT #

I asked for links to Sun power management and you provide nothing? I looked at all the T2 docs at sun.com/t2 and find no mention of power management in the tech specs or any of the pdfs.

This link http://opensolaris.org/os/community/laptop/power/ explains why we don't see Sun SPECpower_ssj2008 results on their x86 platforms: "The Solaris x86 power team is working on adding support for CPU frequency changing, CPU throttling and other power saving features. Solaris has supported power managment on SPARC for a long time and Solaris/x86 has some catching up to do."

Posted by Blogger on June 13, 2008 at 05:59 AM PDT #

Post a Comment:
Comments are closed for this entry.