climateprediction.net home page
Performance on hadcm3igeo "coupled" models

Performance on hadcm3igeo "coupled" models

Message boards : Number crunching : Performance on hadcm3igeo "coupled" models
Message board moderation

To post messages, you must log in.

AuthorMessage
Profile Aaron Doucett

Send message
Joined: 1 Oct 05
Posts: 12
Credit: 10,041,430
RAC: 0
Message 41195 - Posted: 1 Dec 2010, 16:22:47 UTC

Hi, I've recently seen some of my computers running this new type of model. What is alarming to me is that while the famous models generally were taking around ~300 hours on (almost exclusively) Intel dual core machines, I am now seeing To Completion times of over 1000 hours without being in an "ice-world' state. The seconds per time step are over 2.0 in some cases and while I can tell these models are more complex, I am wondering if this is normal. I would rather not have to see a major drop in RAC if there is some kind of incompatibility with this model type.

I appreciate any info!!
ID: 41195 · Report as offensive     Reply Quote
Profile Iain Inglis
Volunteer moderator

Send message
Joined: 16 Jan 10
Posts: 1079
Credit: 6,903,221
RAC: 6,722
Message 41196 - Posted: 1 Dec 2010, 17:48:05 UTC
Last modified: 1 Dec 2010, 17:51:31 UTC

Over the years CPDN has run a number of model types. Generally speaking, volunteers like short models and scientists like long ones. Model run times have been coming down, so volunteers are getting something of what they want.

The credits allocated to each model type vary in order to produce roughly the same RAC (i.e. credit rate). This doesn't always work because models rely on different aspects of the computers on which they run - some models need lots of memory, others just need fast processing, and some models run faster on a particular combination of processor and operating system. It tends to even out in the end.

The geoengineering models are towards the longer end of the run-time range but, other things being equal (which they never are), the RAC shouldn't be that much different. The best thing is to compare the trickle rates for two model types on the same machine and multiply the trickle/day by the credit/trickle to get a RAC-like statistic.
ID: 41196 · Report as offensive     Reply Quote
Darmok

Send message
Joined: 29 Dec 09
Posts: 34
Credit: 18,395,130
RAC: 0
Message 41201 - Posted: 2 Dec 2010, 1:06:59 UTC - in response to Message 41196.  

I unfortunately disagree with Iain. I have a 2.33Ghz Xeon Quad and the oldest hadcm model indicates currently 300 hours completion and 690 remaining out of 770 hours estimated originally, which was of course totally incorrect. At this rate, this model will go well over 2000 hours, if not 3000. As for credits, there are 600-800 per day less than usual with Famous. I like running CPDN as I believe in the research and the cause but I hope there will be adjustments made. :)
ID: 41201 · Report as offensive     Reply Quote
Profile mo.v
Volunteer moderator
Avatar

Send message
Joined: 29 Sep 04
Posts: 2363
Credit: 14,611,758
RAC: 0
Message 41202 - Posted: 2 Dec 2010, 10:40:23 UTC

I didn't download any of these models when they were available and at the moment there are none.

Can anyone please tell us:

1. How many credits are you receiving per trickle?

2. What is the speed of the recent HadCM batch compared with HadCM downloaded much earlier on the same computer and OS? You would need to give us a link to the computer and its tasks please.
Cpdn news
ID: 41202 · Report as offensive     Reply Quote
Profile JIM

Send message
Joined: 31 Dec 07
Posts: 1152
Credit: 22,053,321
RAC: 4,417
Message 41204 - Posted: 2 Dec 2010, 20:13:16 UTC - in response to Message 41201.  
Last modified: 2 Dec 2010, 20:19:05 UTC

Hi, Darmok

The HadCM3 geo models are long, but, not as long as you think. You calculation that the CM geo model that you are running is going to take 2000 to 3000 hours to finish seems high.The real time is probably near 1100 to 1200 hours for your computer.

When I downloaded 2 of the CM3 geo models to my 2.2 GHz with 4 GB of RAM machine the “to completion” read 5660 hours! I realized that this was an over estimate. I was right. It only took 12 hours to crunch 1% and 120 hours to crunch 10%. At this rate it will finish in the neighborhood of 1200 hours. Your computer is slightly faster than mine so it should finish in a little less time.
ID: 41204 · Report as offensive     Reply Quote
3rkko

Send message
Joined: 12 Feb 08
Posts: 66
Credit: 4,877,652
RAC: 0
Message 41208 - Posted: 3 Dec 2010, 0:25:03 UTC

My credit rate for recent models are:
CM3IGEO 24.9 credit/hour, AM3P_EU 25.1, AM3P_SAF 25.2, Famous 27.9, AM3P_PNW 28.0

I have only one CM3 model that is currently at 307 hours & 28.8% done, which indicates 1066 hour total time. All older CM3 models seem to be already archived so I can't make any comparison. The computer is AMD Phenom2 X6, Win7_64, 8GB RAM.
ID: 41208 · Report as offensive     Reply Quote
Darmok

Send message
Joined: 29 Dec 09
Posts: 34
Credit: 18,395,130
RAC: 0
Message 41209 - Posted: 3 Dec 2010, 11:57:13 UTC - in response to Message 41202.  

HadCM3 sends 1 or 2 trickles per day compared to 20/day for Famous. The average credit per trickle is between 312 (older model) and 326 (3 models). The total credits, on a 2-day basis confirms 6.7 trikles sent daily for the 4 models. You can verify with tasks 12016174 - 12014849 - 12011495 - 12009677 on 1041368. I have 2Gb of RAM. Models are smooth so far.

For the runtime estimate, as the completion time is erroneous, it is of course difficult to assess. Again on my older model, there are currently 334 hours elapsed for 96 hours hours completed from the original estimate of 770 (or 674 remaining). It does indicate 30% completion, which confirms Jim's estimate of 1100 hours, but, if you base a rule of 3 calc on the 770 hours, the completion time becomes 2678 hours. This is interesting... As Les pointed out on another thread, the Ocean phase recalcs each 24 hour timestep, and I would add it messes up the completion estimate (and trickles?). Time will tell :)
ID: 41209 · Report as offensive     Reply Quote
Profile Iain Inglis
Volunteer moderator

Send message
Joined: 16 Jan 10
Posts: 1079
Credit: 6,903,221
RAC: 6,722
Message 41210 - Posted: 3 Dec 2010, 12:58:09 UTC - in response to Message 41209.  

... The total credits, on a 2-day basis confirms 6.7 trikles sent daily for the 4 models.
So that's roughly 29 credit/hour for FAMOUS vs 23 credit/hour for HADCM3-geo on that machine, which confirms the RAC discrepancy you reported earlier. Perhaps the credit rates are more appropriate for a single task running uncontended; the rates are usually set in beta testing.

The completion time estimates are a separate matter and frequently go awry on this project and others as well. When necessary I tend to use the elapsed/progress figures to get a %/hour estimate and then make a prediction based on that.
ID: 41210 · Report as offensive     Reply Quote
[B^S] mavau

Send message
Joined: 30 Aug 04
Posts: 142
Credit: 9,936,132
RAC: 0
Message 41216 - Posted: 3 Dec 2010, 19:39:55 UTC
Last modified: 3 Dec 2010, 20:22:26 UTC

Some more info on the models:
Recent average credit with 3 coupled models out of 8 tasks:



Completion estimates:


When the Famous models came out, my RAC went up. I suppose the new coupled models are useful too.
On the other hand, as I've said elsewhere,I think the dedication involved in the long models should be taken into account.
If only to make us feel better when we pay the power bill .-)

Forum search Site search
ID: 41216 · Report as offensive     Reply Quote
Les Bayliss
Volunteer moderator

Send message
Joined: 5 Sep 04
Posts: 7629
Credit: 24,240,330
RAC: 0
Message 41217 - Posted: 3 Dec 2010, 20:42:13 UTC

RAC is a decaying function, and, as the hadcm's have a lot fewer trickles per day, the rac will decay faster, and settle at a lower value.

It looks like "rac chasing" is starting to replace "credit chasing".



Backups: Here
ID: 41217 · Report as offensive     Reply Quote
Profile Aaron Doucett

Send message
Joined: 1 Oct 05
Posts: 12
Credit: 10,041,430
RAC: 0
Message 41234 - Posted: 5 Dec 2010, 23:22:09 UTC - in response to Message 41217.  

RAC helps the new guys who haven't been on the project as long but helps rank the power of your "cluster". I only started CPDN in October but look at the top participants list! :D

Also I helped my team climb in the top 10 slot worldwide.

I have no problem running models that take several days but if it's at the detriment of our RAC it's not so fun!
ID: 41234 · Report as offensive     Reply Quote
[B^S] mavau

Send message
Joined: 30 Aug 04
Posts: 142
Credit: 9,936,132
RAC: 0
Message 41239 - Posted: 6 Dec 2010, 20:14:21 UTC

I was only asking for balance between the different models.
I thought RAC reflected that.
Another issue, credit hasn't been updating for the past two or three days.
Has anyone else noticed?


Forum search Site search
ID: 41239 · Report as offensive     Reply Quote
Les Bayliss
Volunteer moderator

Send message
Joined: 5 Sep 04
Posts: 7629
Credit: 24,240,330
RAC: 0
Message 41240 - Posted: 6 Dec 2010, 20:30:00 UTC - in response to Message 41239.  
Last modified: 6 Dec 2010, 20:30:25 UTC

RAC may work smoothly on other projects, where credit is granted at the end of successful completions, but here it's granted all the way through, based on trickles.
The interval between trickles varies between different model types, so the RAC will rise or fall when a computer is switched between different types.

As different models are associated with different research groups, the lifetime of each model type is limited. So variations in RAC are inevitable.

I've noticed the lack of credit updates.
Backups: Here
ID: 41240 · Report as offensive     Reply Quote
skgiven
Avatar

Send message
Joined: 5 Jun 06
Posts: 28
Credit: 2,790,048
RAC: 0
Message 41363 - Posted: 27 Dec 2010, 16:21:17 UTC - in response to Message 41240.  

From the list of apps/models/experiments on offer (below), is there a definitive (or even rough) answer as to which tasks get the most credit (run the fastest/do the most) on which systems?

UK Met Office HadSM3 Slab Model
UK Met Office HadCM3 Coupled Model Experiment
UK Met Office HadAM3
UK Met Office HadSM3 Mid-Holocene
UK Met Office HadAM3P
UK Met Office FAMOUS
UK Met Office HadAM3P European Region
UK Met Office HadAM3P Southern Africa
UK Met Office HadAM3P Pacific North West

For example, which task would run the fastest on an Phenom CPU, which on a IC2Duo and which on an i7?

Also, which operating system is the fastest (while being stable)?
XP x64, XP x86, Vista/W7 (x64/x86), Linux (x64/x86).

The reason I ask is that I would like to configure my i7 system for maximum throughput (optimization) if possible. I would be happy to install the fastest operating system and run the fastest tasks for my CPU.

Thanks,
ID: 41363 · Report as offensive     Reply Quote
Profile geophi
Volunteer moderator

Send message
Joined: 7 Aug 04
Posts: 2167
Credit: 64,403,322
RAC: 5,085
Message 41365 - Posted: 27 Dec 2010, 16:56:28 UTC - in response to Message 41363.  

From the list of apps/models/experiments on offer (below), is there a definitive (or even rough) answer as to which tasks get the most credit (run the fastest/do the most) on which systems?

UK Met Office HadSM3 Slab Model
UK Met Office HadCM3 Coupled Model Experiment
UK Met Office HadAM3
UK Met Office HadSM3 Mid-Holocene
UK Met Office HadAM3P
UK Met Office FAMOUS
UK Met Office HadAM3P European Region
UK Met Office HadAM3P Southern Africa
UK Met Office HadAM3P Pacific North West

Tough question and lots of variables involved. Intel compiler versions and switches used are the biggest variables. For an i7 (or almost any relatively recent Intel processor):

Linux would be fastest in hadsm3 type models. However, those aren't being distributed now, and it's uncertain how many might be in the future.

For FAMOUS, Windows is definitely faster at this time.

For hadam3 and hadam3p models, there doesn't seem to be much difference between Windows and Linux. Pretty much equivalent.

For PNW, EU, SAF type models, Windows is faster.

For hadcm3 type models, Windows is currently faster, but we are beta-testing a new app version compiled with a different compiler, and this has sped up the Linux version to virtual equivalence.

So, at this time, Windows will give you the best performance overall.

For relatively recent AMD processors, Windows would be the same or faster than the Linux versions, even for hadsm3 type models.
ID: 41365 · Report as offensive     Reply Quote
skgiven
Avatar

Send message
Joined: 5 Jun 06
Posts: 28
Credit: 2,790,048
RAC: 0
Message 41366 - Posted: 27 Dec 2010, 18:08:41 UTC - in response to Message 41365.  

I’m mostly running FAMOUS tasks on the i7 and I guess that will remain the case for some time, so I will keep it on XP x86 for now and if I upgrade, in a month or two, I will just go for a 64bit server. My en route system will also get some win 64bit server. I'm guessing that as with most tasks x64 is slightly faster here too. For me this is all good news.

Many thanks.

ID: 41366 · Report as offensive     Reply Quote
Profile geophi
Volunteer moderator

Send message
Joined: 7 Aug 04
Posts: 2167
Credit: 64,403,322
RAC: 5,085
Message 41368 - Posted: 27 Dec 2010, 21:52:06 UTC - in response to Message 41366.  

I'm guessing that as with most tasks x64 is slightly faster here too. For me this is all good news.

I don't think there is any real performance advantage or disadvantage to x64...at least I haven't noticed any. All cpdn apps are 32 bit, which you can run in 64 or 32 bit BOINC. If the apps were compiled for 64 bit, there would indeed be some performance advantage, but given the vast amount of work the programmers have to do, and the VERY limited resources, that isn't coming anytime soon.
ID: 41368 · Report as offensive     Reply Quote

Message boards : Number crunching : Performance on hadcm3igeo "coupled" models

©2024 climateprediction.net