Seti@Home optimized science apps and information
Optimized Seti@Home apps => Windows => GPU crunching => Topic started by: Pepi on 26 Feb 2010, 10:17:49 pm
-
This stock application is faster on my machine then opt app.
Difference is speed is from 10-15% per results with low credit ( fast result) to 1-5 % per result with high credits (aprox 110 credits)
It also seem that this stock app works better on my GT 240 compared to second comp with GT9800 ( or maybe it is driver thing ( 196.21 WHQL)
-
Did you apply the 2.3 DLL's (to the opt install) Pepi ? (that installer was made before 2.3 was released, and an updated installer is in testing that will do it for you, if selected)
[Edit:] Since you're of squire status you may like to test it for us ? Making beta test posts about them shortly, but the installers are already in the beta download area.
-
Yes , both cuda 2.3 dll are installed :)
After night of crunching I can confirm: GT 240 with stock application and 196.21 are faster in any WU compared to opt app. But IT IS NOT CASE when using same combination of drivers and GT 9800 ( Green). In that case GT9800 (Green) are fast equally with 195.62 and opt app.
Of course I will test it, just now I download 0.35 beta and put on both machines.
Great work guys!!!!
Update: Reschedule work likes charm: all results from 6.09 are transformed to 6.08 without problem. Installer makes great job!
-
Cheers for the installer test!. Yep sounds like the newer driver 'funkiness' is playing a hand in the situation somehow. GT240, being a recent card may be a special case the newest drivers work best for, but others with older cards have reported noticeable slowdown with drivers beyond 190.38 & 191.07. Anyway, will keep an eye on things as nVidia will likely be juggling driver & SDK characteristics for a while. Keep using what's fastest for the card you have :)
Jason
-
For me there was little difference in speed on crunching using 191 or 195 Nvidia drivers on 9800 GT ( Green) with OPT app. But yesterday I was seen one user that uses 196.21 with CUDA 6.09 stock app, and he beats my crunching time on GT 240 with opt app. So I made clear install, use only software that BOINC download to my comp, and yeah, my time was compared to his times, so in that case 6.09 stock app was better then Opt app.
However the most unusual thing was that my card is 10% OC compared to his card, but I am not 10% faster :(
So more thing to find and learn :)
-
I agree ;D, It's a long learning curve this GPU computing stuff ;)
-
I agree ;D, It's a long learning curve this GPU computing stuff ;)
I must add one word: VERY long ;D
-
This stock application is faster on my machine then opt app.
Difference is speed is from 10-15% per results with low credit ( fast result) to 1-5 % per result with high credits (aprox 110 credits)
It also seem that this stock app works better on my GT 240 compared to second comp with GT9800 ( or maybe it is driver thing ( 196.21 WHQL)
I see the same difference. Just eyeballing 5 similar tasks before and after, my 9800gtx and gts250 dualie seem to run slightly faster with the stock 6.09 app (10%). The real advantage with the opt seems to be that the low vlar's are discarded. Those easily up the time by 5x . I am running up enough points in colatz & milkyway ati that I am going to leave the stock seti app in tho I am open to doing any testing.
vista-64, 6.10.43, dual opteron 285, v12 tested, cuda 2.3
-
Hmm.. funny.. and I'm confused.. ;)
I was funny and made again a test, because I love it.. :D
I compared MB_6.08_CUDA_V12_VLARKill_FPLim2048 with setiathome_6.09_windows_intelx86__cuda23 .
The MB_6.08_.. with CUDA V2.3 .
The MB_6.09_.. with the orig. DLed CUDA 2.3 .
I had let run some bench WUs..
[fastest in yellow]
MB_6.08_CUDA_V12_VLARKill_FPLim2048.exe
TaskName: PG0395.wu 46.984 secs Elapsed 16.109 secs CPU time
TaskName: PG0444.wu 43.672 secs Elapsed 16.219 secs CPU time
TaskName: PG1327.wu 31.266 secs Elapsed 15.094 secs CPU time
TaskName: PG0395.wu 46.688 secs Elapsed 16.328 secs CPU time
TaskName: PG0444.wu 43.672 secs Elapsed 16.453 secs CPU time
TaskName: PG1327.wu 31.750 secs Elapsed 15.188 secs CPU time
TaskName: PG0395.wu 46.078 secs Elapsed 15.781 secs CPU time
TaskName: PG0444.wu 43.594 secs Elapsed 16.344 secs CPU time
TaskName: PG1327.wu 31.688 secs Elapsed 15.172 secs CPU time
AVG:
TaskName: PG0395.wu 46.583 secs
TaskName: PG0444.wu 43,646 secs
TaskName: PG1327.wu 31,568 secs
setiathome_6.09_windows_intelx86__cuda23.exe
TaskName: PG0395.wu 48.172 secs Elapsed 17.953 secs CPU time
TaskName: PG0444.wu 42.516 secs Elapsed 17.078 secs CPU time
TaskName: PG1327.wu 33.047 secs Elapsed 16.906 secs CPU time
TaskName: PG0395.wu 46.891 secs Elapsed 17.703 secs CPU time
TaskName: PG0444.wu 43.000 secs Elapsed 17.109 secs CPU time
TaskName: PG1327.wu 33.109 secs Elapsed 16.844 secs CPU time
TaskName: PG0395.wu 46.938 secs Elapsed 17.094 secs CPU time
TaskName: PG0444.wu 43.563 secs Elapsed 18.078 secs CPU time
TaskName: PG1327.wu 32.781 secs Elapsed 16.688 secs CPU time
AVG:
TaskName: PG0395.wu 47,334 secs
TaskName: PG0444.wu 43,026 secs
TaskName: PG1327.wu 32,979 secs
I had let run some normal WUs and was surprised the stock 6.09 is faster than opt. 6.08 .
528 to 572 sec.
This mean ~ 8 % faster!
All 4 WUs with 0.422538 AR:
Opt. 6.08:
http://setiathome.berkeley.edu/result.php?resultid=1566007975
http://setiathome.berkeley.edu/result.php?resultid=1566007971
Stock 6.09:
http://setiathome.berkeley.edu/result.php?resultid=1566008022
http://setiathome.berkeley.edu/result.php?resultid=1566008020
Hmm.. the stock 6.09 is faster than opt. 6.08 ? Why?
Only at this AR, or maybe all others also?
O.K., the real WUs confirm the bench WUs, but.. maybe.. all ARs faster with stock 6.09 ?
Ahh.. on my OCed QX6700 @ 3.14 GHz with manufacturer OCed GTX260-216.
The CUDA WU preparation on CPU is ~ 14 sec for opt. 6.08 and ~ 16 sec. for stock 6.09 .
The opt. 6.08 use ~ 70 MB system RAM, stock 6.09 use ~ 100 MB system RAM.
BTW. Where I can DL stock SETI@home and SETI@home BETA apps?
EDIT: The upper test was with WinXP 32bit SP3. nVIDIA driver 190.38 .
-
Your test is much more thorough than mine. I simply picked 5 wu's that are in the same range and compared against 5 from the stock app that also seemed to be in the same range. I did not want to process the same wu on different apps as I recall it was a PITA the last time I did that.
Anyway, besides the obvious angle, there is other differences that can affect the comparision. The system I ran the test on has 2 opteron dual cores but I set aside one of the cores for feediing my two gpu's. I doint think that Raistmer's needs to be fed because he has code that raises the priority. Perhaps on a system with all 4 cores allocated, raiising the priority can show a big improvement. I like to run that aqua and more recently the quantumfire and would rather have all 4 cores avaiilable for those projects.
I am also running crunch3r's seti linux cuda, tho right now, there are no wu's aviailable. That one uses 2.1 lib and I tried 2.2 but got an invalid signature. I didnt bother with 2.3.
-
What I noticed is next thing: (on AMD computers) if you only use GPU for crunching you will get approx 20% slower results if Cool and Quiet is enabled in BIOS.
And that 20% is used in result preparation and in GPU crunching. So take this information in consideration.
-
I am also running crunch3r's seti linux cuda, tho right now, there are no wu's aviailable. That one uses 2.1 lib and I tried 2.2 but got an invalid signature. I didnt bother with 2.3.
Post more info on the linux forum we have here if you want help.
-
If I add the inclluded section to my appinfo file will it then properly download version 6.09, cuda23 work to crunch on my GTX 275 NVIDIA machines?
-
interesting !
I will try on my 8600 GTS.
-
Is there anything else I need to do besides add that section to the app_info? I would like to give it a try on my 9500GT but with all that's going on on SETI I'm a little afraid to make changes right now.
-
Is there anything else I need to do besides add that section to the app_info? I would like to give it a try on my 9500GT but with all that's going on on SETI I'm a little afraid to make changes right now.
You actually have to download the app and any DLLs you don't yet have! Otherwise you get the "No url specified for download of...." error message.
But once you've downloaded the files you can stick them into the existing app_info section: there's nothing magical about a version number under anonymous platform.
-
Well, leave it to me, I screwed up royally!!! Lost my three day cache and am now back with what I started with. Forgot to back up too. I did manage to go from installer .35 to installer .36. Managed to pick up six tasks so I'm not completely idle. Two of them ran on the GPU, two started out and finished on the CPU, two got moved from the GPU to the CPU by the rescheduler and are running now. With the tasks page as messed up as it is I don't have any idea how they did. ::)