climateprediction.net home page
RUNNING 2 HADAM3P WU\'S TOGETHER

RUNNING 2 HADAM3P WU\'S TOGETHER

Message boards : Number crunching : RUNNING 2 HADAM3P WU\'S TOGETHER
Message board moderation

To post messages, you must log in.

AuthorMessage
Profile JIM

Send message
Joined: 31 Dec 07
Posts: 1102
Credit: 20,168,736
RAC: 848
Message 37810 - Posted: 18 Aug 2009, 6:15:09 UTC

Hi, Everyone:

Recently, while reattaching to the project I accidentally downloaded 2 Hadam3p WU’s to my dual core laptop (Windows 7 Core2Duo, 1.5Ghz with 2 GB of RAM). I had read all the dire warnings about the 2 WU’s fighting over RAM resulting in massive slowdowns if the 2 were run together. I ran both together anyway. What I observed was only slight slowdown and no major problems.

Normally, when running a Hadam3p WU with a another type of WU such as Hadsm3mh the Hadam3p finishes in about 162 hours. The 2 Hadam3p’s run together finished in 177 hours. While it is 15 hours longer, it isn’t the major slowdown I was lead to expect. The 2 WU’s running together (along with all the other running Windows processes) consumed only 60% the available RAM (according to the Moo0 SystemMonitor). The difference in the s/TS was about half a second from 6.09s/TS (run with another type of WU) to 6.67s/TS (run together).

Did other people have much greater slowdown or other problems running 2 Hadam3p’s together?

ID: 37810 · Report as offensive     Reply Quote
Profile mo.v
Volunteer moderator
Avatar

Send message
Joined: 29 Sep 04
Posts: 2363
Credit: 14,072,953
RAC: 41
Message 37814 - Posted: 18 Aug 2009, 10:16:28 UTC
Last modified: 18 Aug 2009, 10:33:36 UTC

On my C2Duo 6600 I get about a 10% speed reduction if I run two together, which isn\'t disastrous. I honestly don\'t know how much the speed would be reduced on other types of dual-core, and there are quite a few different types.

On an Intel quad Les reported while we were testing them in beta that each extra HadAM3P he loaded and ran simultaneously slowed them all down by a similar amount. Ie the slowdowns were cumulative. So running 3 or 4 side-by-side on a quad is probably not a good idea, though again it will depend on the particular computer.

We did notice a multicore hyperthreaded machine running a big load of them (maybe 6 or 8) all at 24 - 25 sec/ts(!). So we have to repeat the warning from time to time. Of course, if you check the speeds once you shouldn\'t need to do so again.

My single-core AMD, which usually is about 40% slower than my C2Duo, had quite a hard time with HadAM3P and ran one more slowly than I expected. I just don\'t let it have any more.
Cpdn news
ID: 37814 · Report as offensive     Reply Quote
Profile Thyme Lawn
Volunteer moderator

Send message
Joined: 5 Aug 04
Posts: 1278
Credit: 15,779,438
RAC: 7
Message 37815 - Posted: 18 Aug 2009, 10:30:03 UTC

The degradation seems to be very dependent on the processor type. As Jim observed there\'s probably not that much difference on dual core systems, but it becomes progressively worse as you move to quads and above.

My Q6600 2.40GHz XP system is averaging around 3.85 s/TS for a single HadAM3P task on CPDN main project. That\'s a bit faster than the 4.12 s/TS I had for a single task on CPDN Beta, but when I scaled up to 4 HadAM3P tasks on Beta the slowest one ran at 5.71 s/TS.

Geophi got a similar effect on a Q9550 (2.45 s/TS -> 3.95 s/TS) but a lesser effect on a Phenom X4 940 (2.65 s/TS -> 3.65 s/TS). As Mo has pointed out the effect was much greater with more cores.
"The ultimate test of a moral society is the kind of world that it leaves to its children." - Dietrich Bonhoeffer
ID: 37815 · Report as offensive     Reply Quote
old_user1
Avatar

Send message
Joined: 5 Aug 04
Posts: 907
Credit: 299,864
RAC: 0
Message 37818 - Posted: 18 Aug 2009, 11:28:26 UTC - in response to Message 37815.  

I guess it\'s a bottleneck with the memory usage, as these higher res models will have even more memory bandwidth requirements as each timestep goes through the whole big array a few times doing calculations etc.
ID: 37818 · Report as offensive     Reply Quote
Kenneth Larsen

Send message
Joined: 26 Aug 04
Posts: 59
Credit: 438,133
RAC: 0
Message 38062 - Posted: 6 Oct 2009, 16:29:02 UTC

Sorry to revive an old topic, but does this warning apply to hadsm models too? I\'ve currently downloaded three of them and was wondering if it would be wise to run them simultaneously (phenom II 940 cpu) or one at a time?
I have 4GB ram and so far haven\'t had any trouble memory-wise, the OS uses only a few hundred MBs (linux).
ID: 38062 · Report as offensive     Reply Quote
Profile mo.v
Volunteer moderator
Avatar

Send message
Joined: 29 Sep 04
Posts: 2363
Credit: 14,072,953
RAC: 41
Message 38063 - Posted: 6 Oct 2009, 17:23:12 UTC
Last modified: 6 Oct 2009, 17:27:37 UTC

The warning doesn\'t apply to HadSM which use very little memory compared with HadAM3P. They\'re good for less well-endowed computers and would be ideal to run alongside HadAM3P if we could get both types. A multicore computer can be fully loaded with them.

But a different warning applies. A small proportion of HadSMs still turn into iceworlds with monochrome graphics, slow-processing on Intel or fast-processing on AMD. Part of the nature of this problem is understood but not how to eliminate it. Iain Inglis is using data collected at the moment iceworlds start and can show their geographical locations. To help with this investigation members need to regularly back up their Boinc Data folder, keep an eye on the model graphics and if an iceworld develops abort the model then report on the forum for details about how to restore the backup and collect the data for Iain. While experienced crunchers are running so many HadSMs I hope there\'ll be more reports of iceworlds, particularly fast ones on AMD.
Cpdn news
ID: 38063 · Report as offensive     Reply Quote
Kenneth Larsen

Send message
Joined: 26 Aug 04
Posts: 59
Credit: 438,133
RAC: 0
Message 38065 - Posted: 7 Oct 2009, 14:23:01 UTC

Thanks for the warning, I\'ll keep that in mind. Unfortunately today I trashed the 3 models because of a bug in BOINC under Linux (it\'s known) and had to unsuccessfully edit the cilent-state.xml by hand.... Anyway, I now have 4 new models happily running. :)
Proud owner of the CPDN Wow-Mug!
ID: 38065 · Report as offensive     Reply Quote

Message boards : Number crunching : RUNNING 2 HADAM3P WU\'S TOGETHER

©2020 climateprediction.net