Modified SETI MB CUDA + opt AP package for full GPU utilization

[Richard Haselgrove]

Yeah, and here's what David Anderson is thinking of doing:

...
The NVIDIA people didn't have a fix or workaround for this.
So I'm going to change the server so that it won't send
CUDA jobs to GPUs slower than 60 GFLOPS
(I'm not sure this is the magic number,
but on my machine the GPU is 50 GFLOPS,
computed as clockRate * multiProcessorCount * 2,857).

-- David

From a comment somewhere in the BOINC sources, his board is a Quadro FX 3700.
                                                                          Joe

That card has the same number of shaders (112) as my 9800GT (but costs six times as much - hehe).

I can't find a clock speed for it - but if I plug the shader clock for the 9800GT (1500 MHz) into the formula, I get 48 GFLOPS: seems like the right ballpark.

If they're going to set the limit 25% above a 9800GT, then GPU crunching is dead until the next generation of hardware.

[alollini]

I try with V5a ... that seems to work
however I have still only 4 tasks running, that may be better with 5

if the overall performance is better with real_number_of_cores + 1

where do i enter this setup ? in what file ? what line ?

[Raistmer]

Yeah, and here's what David Anderson is thinking of doing:

...
The NVIDIA people didn't have a fix or workaround for this.
So I'm going to change the server so that it won't send
CUDA jobs to GPUs slower than 60 GFLOPS
(I'm not sure this is the magic number,
but on my machine the GPU is 50 GFLOPS,
computed as clockRate * multiProcessorCount * 2,857).

-- David

From a comment somewhere in the BOINC sources, his board is a Quadro FX 3700.
                                                                          Joe

That card has the same number of shaders (112) as my 9800GT (but costs six times as much - hehe).

I can't find a clock speed for it - but if I plug the shader clock for the 9800GT (1500 MHz) into the formula, I get 48 GFLOPS: seems like the right ballpark.

If they're going to set the limit 25% above a 9800GT, then GPU crunching is dead until the next generation of hardware.

Sorry but I consider this approach as meaningless. We speak about doing SETI on Android platform, ON CELLPHONES (!!!) and throw away 50GFLOPs of power? Nonsense!
So, there will be custom builds that will work on all CUDA-enabled GPUs and will just abort VLARs as my current V5a/b do.  Much better if true bug-fix will be developed than such inefficient workaround... Even falling back to CPU processing approach much better than this.

[Raistmer]

I try with V5a ... that seems to work
however I have still only 4 tasks running, that may be better with 5

if the overall performance is better with real_number_of_cores + 1

where do i enter this setup ? in what file ? what line ?

search for cc_config.xml file description. You need ncpus field.
I will add minimal cc_config.xml example to first post of this thread.
(done. You can take that sample and edit it if you have duo or single-core CPU instead of quad)

[alexeft]

Hi Reistmer, thanks for all the effort you put (along with Crunch3r)

I'm using 5b and seems to work fine, had no crashes yet.

What I notice is that it goes slow at the beginning of the WU and then accelerates rapidly. Why is that so?
Also, the timer is totally wrong.

Thanks again!

[Richard Haselgrove]

search for cc_config.xml file description. You need ncpus field.

http://boinc.berkeley.edu/wiki/Client_configuration

[Raistmer]

What I notice is that it goes slow at the beginning of the WU and then accelerates rapidly. Why is that so?

This behavior is not specific to V5b build AFAIK. At the beginning of task (and on each restart) some CPU pre-processing should be done (and GPU initial feeding by data) - so few seconds there will be no % done increase.

[alexeft]

No, I'm not talking about those few seconds.

In the beginning it goes at 0.05% per second.
In the middle it goes at 0.1% per second.
Toward the end it goes at 0.3% per second.

Not specific to 5b, all versions did that!!!

[Raistmer]

No, I'm not talking about those few seconds.

In the beginning it goes at 0.05% per second.
In the middle it goes at 0.1% per second.
Toward the end it goes at 0.3% per second.

Not specific to 5b, all versions did that!!!

Don't know - think it's algorithm dependent feature. Maybe someone else can better satisfy your curiosity.

[FloSoft]

Hi,

updated my cuda-app today to 5b and I get client errors: http://setiathome.berkeley.edu/result.php?resultid=1121446826

the "stock" cuda app works, but only says 2-3minutes cpu-time, but computes for 30minutes

[Maik]

updated my cuda-app today to 5b and I get client errors: http://setiathome.berkeley.edu/result.php?resultid=1121446826

Hmm.. Whats your gfx-driver version?

btw.
still have non-VLAR's that go stuck:

• 14.01.2009 06:27:43 > 16no08ah.27587.13978.5.8.44_0 runtime 1295s -> stuck ?
  
  WU true angle range is :  0.433241
  Total GPU memory 536543232    free GPU memory 485399296
  
  Restarted at 32.00 percent.
  Flopcounter: 21582314122380.687000
  Spike count:    0
  Pulse count:    0
  Triplet count:  0
  Gaussian count: 0
  called boinc_finish

[li]14.01.2009 07:55:12 > 16no08ah.27587.13978.5.8.79_0 runtime 686s -> stuck ?

WU true angle range is :  0.433241
Total GPU memory 536543232    free GPU memory 485399296

Restarted at 18.62 percent.
Flopcounter: 21582326705508.687000
Spike count:    0
Pulse count:    0
Triplet count:  0
Gaussian count: 0
called boinc_finish[/li][/list]
[li]14.01.2009 11:44:02 > 16no08ah.10277.3344.6.8.64_0 runtime 351s -> stuck ?

WU true angle range is :  0.433138
Total GPU memory 536543232    free GPU memory 485399296

Restarted at 6.82 percent.
Flopcounter: 21586649651890.055000
Spike count:    7
Pulse count:    0
Triplet count:  0
Gaussian count: 0
called boinc_finish[/li][/list]
[/list][/size]Any suggestions?

[attachment deleted by admin]

[alollini]

that seems to work, I have 4 Astropulse at work and 1 seti enhanced cuda.

the 9600 GSO is warm ! (that is better than cold)

i hope the work units are done in a useful way (and not hidden errors).
my config is AMD quad 9550 + 9600 GSO driver 181.x win XP sp3 32 bits
peak power (12V power supply for sun use was 112w before cuda, now 147... 152w)
has anyone done a performance estimate of different gpu vs performance of one core ?

1122162365   395044125   14 Jan 2009 10:33:21 UTC   14 Jan 2009 12:36:15 UTC   Over   Success   Done   72.72   14.59   pending
1122162329   395044166   14 Jan 2009 10:33:21 UTC   14 Jan 2009 12:36:15 UTC   Over   Success   Done   73.05   14.59   pending
1122162326   395044109   14 Jan 2009 10:33:21 UTC   14 Jan 2009 12:36:15 UTC   Over   Success   Done   67.58   14.59   pending
1122162324   395044111   14 Jan 2009 10:33:21 UTC   14 Jan 2009 12:36:15 UTC   Over   Success   Done   71.41   14.59   pending
1122162322   395044103   14 Jan 2009 10:33:21 UTC   14 Jan 2009 12:36:15 UTC   Over   Success   Done   79.33   14.59   pending
1122162320   395044160   14 Jan 2009 10:33:21 UTC   14 Jan 2009 11:47:52 UTC   Over   Success   Done   379.63   14.62   pending
1121074263   384447812   13 Jan 2009 08:25:12 UTC   14 Jan 2009 11:47:52 UTC   Over   Success   Done   112.72   49.74   37.31
1121074214   394542483   13 Jan 2009 08:25:12 UTC   14 Jan 2009 11:56:27 UTC   Over   Success   Done   67.94   14.15   pending
1121074189   394542488   13 Jan 2009 08:25:12 UTC   14 Jan 2009 11:47:52 UTC   Over   Success   Done   68.89   14.15   pending
1121074187   394542493   13 Jan 2009 08:25:12 UTC   14 Jan 2009 11:47:52 UTC   Over   Success   Done   68.31   14.15   14.15
1121074186   394542482   13 Jan 2009 08:25:12 UTC   14 Jan 2009 11:47:52 UTC   Over   Success   Done   73.73   14.15   14.15

advice: paint your computer boxes in red, those will go faster 

[Raistmer]

Hi,

updated my cuda-app today to 5b and I get client errors: http://setiathome.berkeley.edu/result.php?resultid=1121446826

the "stock" cuda app works, but only says 2-3minutes cpu-time, but computes for 30minutes

Update driver.

[Raistmer]

Any suggestions?

Only one: Try to catch hang with stock 6.07 too and if it will hang - report to Q&A forum - it should be reported back to devs in this case.
And try to keep tasks too for standalone testing.

[Maik]

Everytime a task gets stuck and my script notice that it will do a copy
I've a bunch of them now ^^
Omgh , could someone attach stock app 6.0.7 please.
Ive a full cache ...