GPU crunching question

[Neoman]

Just ran the new seti file, got a -9 result overflow....it took exactly 6 minutes on my outdated, overclocked athlon 64 3400+ "old socket 754", GeForce 7600 OC....

I noticed on the second part of the run, my cpu goes from 138 degrees F, to 150 degrees F 
Is that normal? Even overclocked I have never seen anything above 140 on this pc.....

For those that are overclocking and dont have a temp monitor, I would be careful.......

That cleared my doubt, because my graphic card sits on water, i can hardly see a temp increase. I was doubting that the card is not fully utilised as the temp was always 40°C.
Thanks for the info.

[Pepo]

Run on notebook with C2D T7200, ATI X1600, XP Pro SP2, without additional CPU load from other apps.

---
Two runs with the older version ("next technology preview" from 26.5., which talked a lot), once with 2GHz - 2:55 user + 1:06 kernel time = 4:02, once with 1GHz - 6:03 user + 2:04 kernel time = 8:08.
In both cases CPU consumption was around 35% during the first FFT phase (with additional 3.5% as a punishment consumed by csrss.exe), followed by some minute-two second phase (FindSpikes?) at around 10% CPU load (but punished with 8-17% loaded csrss.exe, depending on CPU speed) and finaly very short phase at 40-50%.

It seemed like the CPU load did not depend on the CPU frequency (on the same HW), but it could be it was limited by one core (only one thread was used). The consumed time does, pretty linearly.

---
Then two runs with the newer version (the more quiet one from 29.5.), once with 2GHz - 2:45 user + 0:38 kernel time = 3:23, once with 1GHz - 5:38 user + 1:09 kernel time = 6:47.

In both cases CPU consumption was around 48-49% during the whole run (with no punishment by csrss.exe), followed by some 20-30 seconds of third phase (memory copying) with no consumed kernel time. It seemed like the CPU load was really limited by one thread + one core. The consumed CPU time went up exactly linearly (+100%) with the 2 x CPU frequency (the same was in timings in stderr.txt, except v_ChirpData - only +25% - no that much CPU frequency dependent?).

---
Physical memory consumption spiked in all cases to nearly 100MB during first seconds at the very beginning, but was rather tiny (3-4 MB) till the end (no idea how it was during the final copying at the very end, virtual memory climbed from 95 to 115 MB and no kernel time was used during this short period).

Both state.sah generated by the two different apps were different - mostly peak_power, ra, bs_score, bs_bin, bs_fft_ind, partially (with less significant bits) time, decl, freq, detection_freq. Both generated a -9 result overflow.

Peter

[pepperammi]

Pentium D 945 @3.6Ghz, GF 7950gt 512mb, Vista x64, 3gb dual channel mem.
cpu at about 50-80%, 40-55% on the app and rest seemed to be on a svchost.exe which would go from 0-50%.
Second half at ~50% and that's all from the app. Still a  -9 overflow.
Time bit quicker I think at roughly 8 minutes

[attachment deleted by admin]

[Culibrk]

Thanks Duke for your efforts, must say I admire your coding skills.
vista32, 6600C2D @3.33Ghz, 7900GT/512MB @450/650 - wu done under 6mins and -9 err`s others at the end.


Fakticky si Tvoji prace vazim a stejme jako zbytek sveta se z ni tesim.





*** sorry for speaking czech, it's just that Duke deserves an "almost native" support

[attachment deleted by admin]

[Devaster]

for now i have small problems with vista activation  but i think i will solve it....

je pekne "pocut" aj skoro rodny jazyk ....

[khanna]

I saw on this thread something about a GPU client based on RapidMind. Is that still under development? I'm interested because the same could be used on the PS3/Cell.

Gaurav

[popandbob]

I saw on this thread something about a GPU client based on RapidMind. Is that still under development? I'm interested because the same could be used on the PS3/Cell.

Gaurav

Go back 1 page and see/use the test app for yourself.

BoB

[Devaster]

I saw on this thread something about a GPU client based on RapidMind. Is that still under development? I'm interested because the same could be used on the PS3/Cell.

Gaurav

yes, it can be used because code is now writed to use any available platform, but must be compiled under linux...

bu i think there is better way to use power of cell with specialised compilers

[Pepo]

je pekne "pocut" aj skoro rodny jazyk ....

Keď už nie naozaj "počuť", tak aspoň ho vidieť 
Radi poslúžime, ak to poteší a pridá chuť do práce

Peter

[|MatMan|]

Opteron 165 @ 2x 2,6 GHz, Geforce 6800 @ 16/5, Windows XP SP2, Forceware 160.03

one CPU core at 100% all the time, WU took 3 min to complete, stopped with -9 error as stderr.txt stated

the client stopped after fft computation (I guess), here the console output:

Code: [Select]

after init analyze
20000001: Generating glsl program
        Version (SPLIT 4)(2D FETCH_1D_REPEAT ARRAY NEW_INPUT)(1D TEXI ACCESSOR)
2D FETCH_1D_REPEAT ARRAY NEW_INPUT)(1D TEXI ACCESSOR)
20000001: Generating glsl program
        Version (SPLIT 4)(2D FETCH_1D_REPEAT ARRAY NEW_INPUT)(1D TEXI ACCESSOR)
2D FETCH_1D_REPEAT ARRAY NEW_INPUT)(1D TEXI ACCESSOR)(2D FETCH_1D_REPEAT ARRAY
EW_INPUT)(1D TEXI ARRAY)
20000001: Generating glsl program
        Version (SPLIT 4)(2D FETCH_1D_REPEAT ARRAY NEW_INPUT)(2D FETCH_1D_REPEA
 ARRAY NEW_INPUT)
20000001: Generating glsl program
        Version (SPLIT 4)(2D FETCH_1D_REPEAT ARRAY NEW_INPUT)(2D FETCH_1D_REPEA
 ARRAY NEW_INPUT)(2D FETCH_1D_REPEAT ARRAY NEW_INPUT)(1D TEXI ARRAY)
20000001: Generating glsl program
        Version (SPLIT 4)(2D FETCH_1D_REPEAT ARRAY NEW_INPUT)
after fft 0 131072 size:8
after fft 10000 131072 size:8
after fft 20000 131072 size:8
after fft 30000 131072 size:8
after fft 40000 131072 size:8
after fft 50000 131072 size:8
after fft 60000 131072 size:8
after fft 70000 131072 size:8
after fft 80000 131072 size:8
after fft 90000 131072 size:8
after fft 100000 131072 size:8
after fft 110000 131072 size:8
after fft 120000 131072 size:8
after fft 130000 131072 size:8
20000001: Generating glsl program
        Version (SPLIT 4)(2D FETCH ARRAY)
after powerspectrum
10000001: Copying from user-managed memory.
        Array arr2f2 was the last owner of this memory but the data is required
by other Arrays.
10000002: Memory copy
        From array  to array
Nice to see something else than F@H happening on the GPGPU front - keep up the good work!

[attachment deleted by admin]

[Devaster]

Please send your feedback to this topic: http://lunatics.at/7-gpu-crunching/feedback-from-gpu-client.new.html#new

[WR-HW95]

Weird,but my 8800GTX was on idle while i ran this.   

[Slawek]

Hi
Meybe this URL helps: http://developer.download.nvidia.com/compute/cuda/0_8/NVIDIA_CUFFT_Library_0.8.pdf

[Holly]

Feedback url doesn't work for me for some reason.

Ran the latest app on C2D Vista64, GF7300GS, 2GB dualchannel RAM. CPU load was at 45% - 60%.

Results attached (-9 occured as well with latest build)

[attachment deleted by admin]

[riha]

So, i found this page as well, read through the Forum, downloaded the app and tried oit out.

I also get error -9 in my log file.

This is the text i get in logfile, better posting it like this instead of attaching the file for only a few lines.


----
Can't set up shared mem: -1
Work Unit Info:
...............
WU true angle range is :  0.425877
Optimal function choices:
-----------------------------------------------------
                          name  timing    error
-----------------------------------------------------
              v_BaseLineSmooth 0.11469 0.00000
            v_GetPowerSpectrum 0.00023 0.00000
                   v_ChirpData 0.02190 0.00000
                  v_Transpose4 0.00730 0.00000
SETI@Home Informational message -9 result_overflow
NOTE: The number of results detected exceeds the storage space allocated.
--------------



I am as well using the latest posted version from this forum.

Some info on system:

Windows XP SP2, swedish
2 Gb RAM
Nvidia 7950GT 512Mb Gfx card