Advanced search

Message boards : Graphics cards (GPUs) : returning to gpugrid, suggestions on setup please, thanks

Author Message
alephnull
Send message
Joined: 8 Jul 09
Posts: 13
Credit: 306,850,267
RAC: 0
Level
Asp
Scientific publications
watwatwatwatwatwatwatwatwatwatwat
Message 20677 - Posted: 16 Mar 2011 | 23:28:12 UTC

hi.

im trying to get set up again on gpugrid. recently got a pair of gtx570 cards and was wondering if i could get some help and insight into my setup. it doesnt really seem right to me for some reason. ive read a lot of information about cc_config, app_info, windows 7 overhead, etc. and have tried various combinations of things but i still get high cpu usage and also wanted to verify how badly i may have messed things up with my current configuration. any suggestions or insight would be greatly appreciated.

cc_config.xml:

<use_all_gpus>1</use_all_gpus>
<report_results_immediately>0</report_results_immediately>

i read to set <report_results_immediately> to 1 but i was hoping i could keep this at 0 for now.

env var SWAN_SYNC = 0

app_info.xml:

<app_info>
<app>
<name>acemd2</name>
</app>
<file_info>
<name>acemd2_6.13_windows_intelx86__cuda31.exe</name>
<executable/>
</file_info>
<file_info>
<name>cudart32_31_9.dll</name>
<executable/>
</file_info>
<file_info>
<name>cufft32_31_9.dll</name>
<executable/>
</file_info>
<file_info>
<name>tcl85.dll</name>
<executable/>
</file_info>
<app_version>
<app_name>acemd2</app_name>
<version_num>613</version_num>
<plan_class>cuda</plan_class>
<avg_ncpus>0.26</avg_ncpus>
<max_ncpus>0.26</max_ncpus>
<flops>1440690000000</flops>
<coproc>
<type>CUDA</type>
<count>1</count>
</coproc>
<file_ref>
<file_name>acemd2_6.13_windows_intelx86__cuda31.exe</file_name>
<main_program/>
</file_ref>
<file_ref>
<file_name>cudart32_31_9.dll</file_name>
</file_ref>
<file_ref>
<file_name>cufft32_31_9.dll</file_name>
</file_ref>
<file_ref>
<file_name>tcl85.dll</file_name>
</file_ref>
</app_version>
</app_info>

i set up "school" prefs to use 87.5% cpu and set my computer to school and issued the update

here are some lines from BOINC. ive removed the messages pertaining to the other projects:

2011-03-16 18:41:16 Starting BOINC client version 6.10.58 for windows_x86_64
2011-03-16 18:41:16 Config: use all coprocessors
2011-03-16 18:41:16 Config: GUI RPC allowed from:
...
2011-03-16 18:41:16 log flags: file_xfer, sched_ops, task
2011-03-16 18:41:16 Libraries: libcurl/7.19.7 OpenSSL/0.9.8l zlib/1.2.3
2011-03-16 18:41:16 Data directory: D:\BOINC_DATA
2011-03-16 18:41:16 Running under account ...
2011-03-16 18:41:16 Processor: 8 GenuineIntel Intel(R) Core(TM) i7 CPU 920 @ 2.67GHz [Family 6 Model 26 Stepping 5]
2011-03-16 18:41:16 Processor: 256.00 KB cache
2011-03-16 18:41:16 Processor features: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss htt tm pni ssse3 cx16 sse4_1 sse4_2 syscall nx lm vmx tm2 popcnt pbe
2011-03-16 18:41:16 OS: Microsoft Windows 7: Ultimate x64 Edition, Service Pack 1, (06.01.7601.00)
2011-03-16 18:41:16 Memory: 5.99 GB physical, 11.98 GB virtual
2011-03-16 18:41:16 Disk: 159.54 GB total, 155.35 GB free
2011-03-16 18:41:16 Local time is UTC -4 hours
2011-03-16 18:41:16 NVIDIA GPU 0: GeForce GTX 570 (driver version 26658, CUDA version 3020, compute capability 2.0, 1249MB, 1530 GFLOPS peak)
2011-03-16 18:41:16 NVIDIA GPU 1: GeForce GTX 570 (driver version 26658, CUDA version 3020, compute capability 2.0, 1248MB, 1530 GFLOPS peak)
...
2011-03-16 18:41:16 GPUGRID Found app_info.xml; using anonymous platform
...
2011-03-16 18:41:16 GPUGRID General prefs: from GPUGRID (last modified 16-Mar-2011 18:33:37)
2011-03-16 18:41:16 GPUGRID Computer location: school
2011-03-16 18:41:16 General prefs: using separate prefs for school
2011-03-16 18:41:16 Preferences:
2011-03-16 18:41:16 max memory usage when active: 3067.55MB
2011-03-16 18:41:16 max memory usage when idle: 6135.11MB
2011-03-16 18:41:16 max disk usage: 79.77GB
2011-03-16 18:41:16 max CPUs used: 7
2011-03-16 18:41:16 don't use GPU while active
2011-03-16 18:41:16 suspend work if non-BOINC CPU load exceeds 25 %
2011-03-16 18:41:16 (to change preferences, visit the web site of an attached project, or select Preferences in the Manager)
2011-03-16 18:41:16 Not using a proxy
...
2011-03-16 18:41:16 GPUGRID Restarting task 163-KASHIF_HIVPR_maca_wo_ba1-16-50-RND7576_0 using acemd2 version 613
2011-03-16 18:41:16 GPUGRID Restarting task 480-KASHIF_HIVPR_wo_ba1-17-50-RND3763_0 using acemd2 version 613


locally, i have set "Use GPU Always"

this system is currently oc @ 3.6ghz. the video cards are presently running stock speeds.

ive also set my computers to be visible on the website. im currently only trying to see if i can get this one host set up properly then will add the other hosts i have:

2 x gtx 275
2 x gtx 460
2 x gtx 580

running now, gpugrid still takes 94% amd 96% on two cores for each video card. as of right now i have

wt -- (et) -- cpu% -- %complete
00:47:04 -- (00:44:40) -- 94% -- 15.2%
00:47:49 -- (00:46:26) -- 96% -- 15.3%

according to BoincTasks. the elapsed time seems awefully high @ 0.26C + 1NV i would think or am i completely off here?

im guessin that roughly equals to about 320 minutes (5.3) hours per wu. does that sound about right?

i also do not understand why two cpu cores are running close to 100%. if someone can explain this to me id greatly appreciate it.

thanks for any insight you could provide. id really like to get my other three machines crunching gpugrid again but want to make sure im not completely messing all this up.

Profile skgiven
Volunteer moderator
Project tester
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3538
Credit: 1,012,918,807
RAC: 1,278,156
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 20679 - Posted: 17 Mar 2011 | 7:08:02 UTC - in response to Message 20677.

Your driver choice and settings look quite good, but you might want to change a few Preferences:

Consider increasing your max memory usage when active from 3067.55MB to 4 or 5GB, depending on what other projects you are running on that system.

Reduce the CPUs from 7 to 6 (free up one thread per GPU).

Set Boinc to "Use GPU while active".

Increase suspend work if non-BOINC CPU load exceeds 25 % to 75% or just set it to zero, to remove this setting.


About 320 minutes (5.3) hours per wu is about right for your GPU on that operating system and with the settings you have. The long tasks (if you select to run them as well) will take around 8 to 12h on that card, and get better credit.

The swan_sync environmental variable causes the high CPU use, but speeds up the task quite a bit. With high end GPU's such as the GTX570 it is very much worth while getting as much as possible out of one GPU even at the expense of a CPU thread per GPU.

PS. I would be inclined to not use the GeForce 9800 GTX+ for GPUGrid, especially given your arsenal.

Good luck,

alephnull
Send message
Joined: 8 Jul 09
Posts: 13
Credit: 306,850,267
RAC: 0
Level
Asp
Scientific publications
watwatwatwatwatwatwatwatwatwatwat
Message 20690 - Posted: 17 Mar 2011 | 22:04:35 UTC - in response to Message 20679.

skgiven,

thanks for the reply. i finally was able to catch up on some stuff and make the changes you proposed. today it seemed the workday would never end...

here are the changes i made

-- bumped up the max memory usage to 90% in use / 100% not in use
-- reduced cpu cores to 6/8 (75%)
-- "suspend work if non-BOINC CPU load exceeds ..." is now set to 0
-- changed <report_results_immediately> to 1 in cc_config for now ad will keep this setting for the next month during this trial phase for gpugrid settings

you suggested "Set Boinc to "Use GPU while active"". i did not do this but i override that setting locally anyway. that way if i use another computer where this should be true and i forget to change it for that instance, ill be ok. i forget little stuff like that a lot so its more a failsafe for my spaciness.

what you said about the higher cpu usage using swan_sync is acceptable. as such, i have decided that this machine, while running gpugrid with these "school" preferences, will not run any other cuda based apps on it.

im glad the times seem to be inline with what should be expected. im interested to see how this performs over a longer period of time so im gonna let this run in this configuration for a month to see how it works out and decide from there how to set up the other systems.

looking at my app_inf.xml file, is it set up to run the longer tasks as you mentioned or would i have to add additional entries? i dont mind running those tasks but im not sure if the settings there are correct for that.

this machine still has some other project's cuda wu but ive stopped new work for those projects. when those run out, we will see the actual turnaround for gpugrid.

youre also right to remind me of that machine with the 9800 card in it. i think ill reserve that card for older, less intense cuda apps and detach gpugrid from it.

again, thanks for the response and the help. im looking forward to seeing how well these cards perform for gpugrid and will eventually migrate these settings to those other systems too.

take care.

Profile skgiven
Volunteer moderator
Project tester
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3538
Credit: 1,012,918,807
RAC: 1,278,156
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 20693 - Posted: 18 Mar 2011 | 7:19:56 UTC - in response to Message 20690.

I can't see your devices anymore but one thing I didn't mention is that you should keep a low cache, especially if you are crunching for the long work units.

WRT your app_inf.xml

Firstly, you don't normally need to use this. It's the anonamous platform config, used to configure Boinc to work manually on operating systems that are not supported by the project. The project works on Windows!

You would need to add the long task application name to run long tasks.
You have,
<name>acemd2_6.13_windows_intelx86__cuda31.exe</name>

They are called,
Long runs (8-12 hours on fastest card) 6.13 (cuda31)

So perhaps,
acemdlong_6.13_windows_intelx86__cuda31.exe

<but I have not checked this>

Suggest you backup your app_inf.xml file and then just delete it, restart and configure the system directly through the web and Boinc Manager.

alephnull
Send message
Joined: 8 Jul 09
Posts: 13
Credit: 306,850,267
RAC: 0
Level
Asp
Scientific publications
watwatwatwatwatwatwatwatwatwatwat
Message 20694 - Posted: 18 Mar 2011 | 8:16:53 UTC - in response to Message 20693.

didnt think about that. i was running before without the app_info.xml file and remember seeing those long work units. the only reason i created it was to play around with the cpu settings but decided to free up two cores instead. i was also considering using it to run 2 tasks per card (gpu usage only about 80% for gpugrid with 1 wu) but read that not a great idea so ill refrain from that too. seeing as though this is set in the preferences, the app_info file is no longer required so it has been removed and the project will just run normally with the preferences you helped me finalize.

btw, my cache is set for 1 day of work only.

ill monitor the project over the next couple of days and see if those long wu return without the app_info file. im pretty sure they will. thanks for the reminder that this is no longer needed and for all the other help.

cya

alephnull
Send message
Joined: 8 Jul 09
Posts: 13
Credit: 306,850,267
RAC: 0
Level
Asp
Scientific publications
watwatwatwatwatwatwatwatwatwatwat
Message 20695 - Posted: 18 Mar 2011 | 8:27:30 UTC - in response to Message 20694.

by removing the app_info file, the four wu i had were discarded; sorry for that. i did get two new long wu so all is good. the setup seems to be good now. by tomorrow ill be able to see how fast there are able to be completed.

still got a couple milkyway, primegrid and collatz wu there to finish up before its 100% gpugrid. that should all be cleared by the end of tomorrow.

again, thanks.

rob

Profile skgiven
Volunteer moderator
Project tester
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3538
Credit: 1,012,918,807
RAC: 1,278,156
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 20696 - Posted: 18 Mar 2011 | 10:30:54 UTC - in response to Message 20695.
Last modified: 18 Mar 2011 | 10:34:58 UTC

I would keep a slightly lower cache, to be on the safe side. This is just to make sure you get the maximum credit. Long tasks might vary in runtime and I suppose you could just miss out if you got two tasks back to back on the same GPU that each took 12h (50% bonus for returning completed work inside 24h and 25% bonus between 24h and 48h).
0.5days should be fine for a GTX470 running 24/7 for example, and would even allow for a system restart or two. Using report tasks immediately reduces the task turn over time, so stick with that.
Different cards have different performances, so for a GTX580 cache size should not matter (the maximum number of tasks per GPU is 2), two of the longest tasks should return within one day, as long as the system is on 24/7. Obviously for a GTX460 things are different; an 8h task on a GTX580 might take between 20 and 22h on a GTX460. So you would want a very low cache if you ran the long tasks on such a card.

Good luck,

STE\/E [BlackOpsTeam]
Send message
Joined: 18 Sep 08
Posts: 357
Credit: 250,952,560
RAC: 0
Level
Asn
Scientific publications
watwatwatwatwatwatwatwatwatwatwat
Message 20697 - Posted: 18 Mar 2011 | 13:22:48 UTC

I set my AMD 1090T's to 4 Cores, seems to work okay that way. But my Quads without HT if I set to 3 or 2 Cores it dosen't seem to matter, the GPU Wu''s still want to use the full Core either way even though it say's using 0.28 CPU ???

I'm just leaving the Quad's at 3 Cores for now as I can't tell any difference so far ...

alephnull
Send message
Joined: 8 Jul 09
Posts: 13
Credit: 306,850,267
RAC: 0
Level
Asp
Scientific publications
watwatwatwatwatwatwatwatwatwatwat
Message 20698 - Posted: 18 Mar 2011 | 13:59:32 UTC - in response to Message 20697.
Last modified: 18 Mar 2011 | 14:07:24 UTC

thanks for the recommendation about the cache. its been changed to 0.5 days cause that makes sense.

last night, i suspended gpugrid and when i woke up, all other cuda tasks for the other projects were done. this machine is now going to run gpugrid exclusively on the cards and CPDN on the remaining cores. interested to see its performance for the next month.

as of now, those long wu are 35 minutes @ approx 5% complete. guessin about 697 minutes (11.6 hours) but we'll see the actual time they complete. if the 580s can complete them in 8 hours, thats a huge difference. then again you mentioned long wu seem to have various execution times. maybe these are the longer of the longs.

i do have HT on on this system. would it be better to turn this off or set max cpu usage to 50%? sounds as though max cpu at 75% with HT like i have now may be ok.

btw, all my systems run 24/7. crunching and working is all i use the machines for now. sorta phased out of gaming a couple years ago as work got busier so i dont need any special configurations for that anymore.

cya

STE\/E [BlackOpsTeam]
Send message
Joined: 18 Sep 08
Posts: 357
Credit: 250,952,560
RAC: 0
Level
Asn
Scientific publications
watwatwatwatwatwatwatwatwatwatwat
Message 20700 - Posted: 18 Mar 2011 | 14:16:17 UTC

The 580's can complete some of them in 8 hr's, in fact I've had some complete in less than 5 Hr's at Stock running Speeds. But some take 10-12 Hr's even on the 580's, there's no set time for them. If you could get all the same type Wu then the Times would be more consistent but that's not the case as you get various Long Running ones.

alephnull
Send message
Joined: 8 Jul 09
Posts: 13
Credit: 306,850,267
RAC: 0
Level
Asp
Scientific publications
watwatwatwatwatwatwatwatwatwatwat
Message 20701 - Posted: 18 Mar 2011 | 14:31:48 UTC - in response to Message 20700.
Last modified: 18 Mar 2011 | 14:33:09 UTC

The 580's can complete some of them in 8 hr's, in fact I've had some complete in less than 5 Hr's at Stock running Speeds. But some take 10-12 Hr's even on the 580's, there's no set time for them. If you could get all the same type Wu then the Times would be more consistent but that's not the case as you get various Long Running ones.


thats interesting to know. currently im testing out a system with the 570 cards. gonna let it cook in for a month to measure a solid performance trend and then ill test out the 580s for comparison.

after reading about other card performance, it seems to make more sense to run gpu grid only on the 570 and 580 cards. interestingly, i read the 460s seem to be no good for this project (more or less the same performance as the 275s) so ill reserve that machine for my other cuda projects.

i oc almost all my cpus but all my video cards run stock speeds. maybe one day ill learn to oc them but for now, im not knowledgeable enough to oc the video cards. cpus seem relatively easier, but maybe thats cause ive been doing that for a while.

Profile skgiven
Volunteer moderator
Project tester
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3538
Credit: 1,012,918,807
RAC: 1,278,156
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 20702 - Posted: 18 Mar 2011 | 18:22:31 UTC - in response to Message 20701.
Last modified: 18 Mar 2011 | 18:26:27 UTC

Steve, you might notice a difference in your CPU task times. I think they will rise if you are using 3 cores, but I understand wanting to use 3 CPU cores on a quad.

Rob, the GTX460 works reasonably well, but last time I checked it was more akin to a GTX260 than a GTX275 in terms of performance.

When I looked at the performance of GTX580's compared to GTX570's I worked out that you would need to add a 5% correction factor (after you accept the extra shaders and speed); the 580 being relatively better. To put it another way, GTX570 performance is 5% worse than the GTX580 after you allow for the difference in shader speed and count. Basically this is because it has been knobbled slightly. So expect a performance difference of +17% for the GTX580:
1.05x(512x772)/(480x732)=1.05x395264/351360=411075/351360=1.17
(shaders x freq.)

The GTX570 is still much better value for money; costs 50% less. Prices might start to change next week though.

STE\/E [BlackOpsTeam]
Send message
Joined: 18 Sep 08
Posts: 357
Credit: 250,952,560
RAC: 0
Level
Asn
Scientific publications
watwatwatwatwatwatwatwatwatwatwat
Message 20703 - Posted: 18 Mar 2011 | 18:43:05 UTC

So are you saying I should only use 2 Cores on the Quad's ???

Profile skgiven
Volunteer moderator
Project tester
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3538
Credit: 1,012,918,807
RAC: 1,278,156
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 20704 - Posted: 18 Mar 2011 | 18:54:48 UTC - in response to Message 20703.

It's up to you. I'm just saying that I would expect some reduction in CPU task performance anyway; using swan_sync will basically grab one CPU core/thread per GPU. If I was using a high end Fermi (CC2.0) I would always use swan_sync and free up a core/thread per GPU, but with a mid-range Fermi or a low end card it would sort of depend on what else I was crunching for.

Rob,

would it be better to turn this off or set max cpu usage to 50%? sounds as though max cpu at 75% with HT like i have now may be ok.

Leave HT on; one thread per GPU is plenty.

STE\/E [BlackOpsTeam]
Send message
Joined: 18 Sep 08
Posts: 357
Credit: 250,952,560
RAC: 0
Level
Asn
Scientific publications
watwatwatwatwatwatwatwatwatwatwat
Message 20706 - Posted: 18 Mar 2011 | 19:09:55 UTC

Okay, I changed back to 2 Core's on the Quads anyway ... Thanks

alephnull
Send message
Joined: 8 Jul 09
Posts: 13
Credit: 306,850,267
RAC: 0
Level
Asp
Scientific publications
watwatwatwatwatwatwatwatwatwatwat
Message 20707 - Posted: 18 Mar 2011 | 20:12:32 UTC - in response to Message 20702.

cool, thanks, ill keep HT on then.

...
Rob, the GTX460 works reasonably well, but last time I checked it was more akin to a GTX260 than a GTX275 in terms of performance.
...

thats ok then. i got those 460s on sale for relatively nothin, thats why i picked em up.

...
The GTX570 is still much better value for money; costs 50% less. Prices might start to change next week though.
...

totally agree: i actually bought the 570s after the 580s for just that reason.

590s out in 4 days. woohoo!

Post to thread

Message boards : Graphics cards (GPUs) : returning to gpugrid, suggestions on setup please, thanks