ABP2 CUDA applications

Richard Haselgrove
Richard Haselgrove
Joined: 10 Dec 05
Posts: 2,040
Credit: 688,884,747
RAC: 943,873

RE: Hi Bikeman, I got this

Message 96313 in response to message 96311

Quote:

Hi Bikeman,
I got this error message after created cc_config.xml file by copying and paste your lines:
Dim 17 jan 22:31:32 2010 Unrecognized tag in cc_config.xml:
Dim 17 jan 22:31:32 2010 Missing end tag in cc_config.xml

GPV


Make sure you compose the file with the simplest possible plain-text editor, and save it as plain (ASCII) text - no unicode or anything like that.

Bikeman (Heinz-Bernd Eggenstein)
Bikeman (Heinz-...
Moderator
Joined: 28 Aug 06
Posts: 3,516
Credit: 468,481,246
RAC: 58,553

Ahhhh....sorry, I misspelled

Ahhhh....sorry, I misspelled the ending tags.

It should be like this (sorry for that):

[pre]


1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1

[/pre]

GPV67
GPV67
Joined: 26 Mar 05
Posts: 10
Credit: 17,892,243
RAC: 367

Hi All, I corrected cc_config

Hi All,
I corrected cc_config file.
I now see what's wrong: not enough Graphical RAM.

Dim 17 jan 23:20:04 2010 Einstein@Home [cpu_sched_debug] 2030_54170_48896_0079_G51.72+00.82.C_4.dm_545_1: insufficient GPU RAM (387MB < 450MB)
Dim 17 jan 23:20:04 2010 [cpu_sched_debug] Request CPU reschedule: insufficient GPU RAM

So I have to wait 1GB Graphical card...

The bad thing is that a lot of ABP2 CUDA units are downloaded, even they cannot be computed due to lack of GPU memory!

Cheers,
GPV

Bikeman (Heinz-Bernd Eggenstein)
Bikeman (Heinz-...
Moderator
Joined: 28 Aug 06
Posts: 3,516
Credit: 468,481,246
RAC: 58,553

RE: Hi All, I corrected

Message 96316 in response to message 96315

Quote:

Hi All,
I corrected cc_config file.
I now see what's wrong: not enough Graphical RAM.

Dim 17 jan 23:20:04 2010 Einstein@Home [cpu_sched_debug] 2030_54170_48896_0079_G51.72+00.82.C_4.dm_545_1: insufficient GPU RAM (387MB < 450MB)
Dim 17 jan 23:20:04 2010 [cpu_sched_debug] Request CPU reschedule: insufficient GPU RAM

So I have to wait 1GB Graphical card...

The bad thing is that a lot of ABP2 CUDA units are downloaded, even they cannot be computed due to lack of GPU memory!

Cheers,
GPV

Hi
Thanks for letting us know, that's interesting.

In theory, the available GPU memory is checked also BEFORE fetching the work. Then again, I understand the ABP2 workunits are the same for CPU and GPU variants, so (again, in theory) BOINC could at least run the WUs with the CPU app again. I think this is something that BOINC and BOINC only can decide, it would not be good if the app itself would fall-back on CPU computation because it would be transparent to BOINC and all the resource allocation (here: 1 CPU core plus 1 GPU) and flops estimations would be terribly wrong then.

I'll forward the info to the devs, thanks again for obtaining the debug info.

CU
H-B

P.S.: [B^S] Elphidieus, is it the same for your Mac?

Richard Haselgrove
Richard Haselgrove
Joined: 10 Dec 05
Posts: 2,040
Credit: 688,884,747
RAC: 943,873

RE: RE: Hi All, I

Message 96317 in response to message 96316

Quote:
Quote:

Hi All,
I corrected cc_config file.
I now see what's wrong: not enough Graphical RAM.

Dim 17 jan 23:20:04 2010 Einstein@Home [cpu_sched_debug] 2030_54170_48896_0079_G51.72+00.82.C_4.dm_545_1: insufficient GPU RAM (387MB 450MB - thus allowing the task to be issued.

But the new test checks that the GPU has enough available RAM, when the time comes to run it, and in this case the test fails.

So the next question for GPV is: what is using (512 - 387) = 125MB of your video RAM, when BOINC/CUDA isn't?

GPV67
GPV67
Joined: 26 Mar 05
Posts: 10
Credit: 17,892,243
RAC: 367

RE: RE: In theory, the

Message 96318 in response to message 96317

Quote:
Quote:


In theory, the available GPU memory is checked also BEFORE fetching the work. Then again, I understand the ABP2 workunits are the same for CPU and GPU variants, so (again, in theory) BOINC could at least run the WUs with the CPU app again. I think this is something that BOINC and BOINC only can decide, it would not be good if the app itself would fall-back on CPU computation because it would be transparent to BOINC and all the resource allocation (here: 1 CPU core plus 1 GPU) and flops estimations would be terribly wrong then.

I'll forward the info to the devs, thanks again for obtaining the debug info.


This may be the effect of a new 'feature' in BOINC v6.10.25 and later: changeset [trac]changeset:19864[/trac]. The scheduler may have checked that the video card had - in absolute terms - enough RAM to run the task: 512MB > 450MB - thus allowing the task to be issued.

But the new test checks that the GPU has enough available RAM, when the time comes to run it, and in this case the test fails.

So the next question for GPV is: what is using (512 - 387) = 125MB of your video RAM, when BOINC/CUDA isn't?


Hi,
I don't know what is using 125MB video RAM, I don't start any graphical program during that tests.
What I know is that Mac OS X 10.6 includes openCL capabilities, and the system itself can use GPU for some Mac internals.

Harri Liljeroos
Harri Liljeroos
Joined: 10 Dec 05
Posts: 1,230
Credit: 1,299,427,642
RAC: 1,720,355

Hi! The following is probably

Hi!
The following is probably more of a BOINC problem than Einstein but anyway here's what is happening to me:

In order to get those ABP2 CUDA tasks returned as fast as possible I suspended some Einstein WU's on Friday as requested in another thread. Everything seemed to go OK. Unfortunately there was a power supply failure during weekend and the computer was shutdown by itself (laptop). On Monday when I finally got the thing running again I noticed some of the suspended WU's were due to return on the 20th so I unsuspended them and two of them were started to crunch in high priority mode (this is a Core2Duo T9500), after a minute one of the CUDA ABP2's were also started and now I have totally three WU's crunching at the same time. On the CPU one APB1 3.12 and one S5R6 3.01 in high priority mode and one ABP2 3.03 with CPU+GPU.

I do not have any local configuration overrides in the computer. The settings are 100% of CPU's and 100% of CPU time. BOINC is 6.10.18 on WinXP32.

I don't think this is a normal situation.

Oliver Behnke
Oliver Behnke
Moderator
Administrator
Joined: 4 Sep 07
Posts: 889
Credit: 25,165,240
RAC: 0

RE: RE: So the next

Message 96320 in response to message 96318

Quote:
Quote:
So the next question for GPV is: what is using (512 - 387) = 125MB of your video RAM, when BOINC/CUDA isn't?

Hi,
I don't know what is using 125MB video RAM, I don't start any graphical program during that tests.
What I know is that Mac OS X 10.6 includes openCL capabilities, and the system itself can use GPU for some Mac internals.

This memory is typically used by your desktop environment. The amount used depends on your screen resolution, number of monitors, enabled 3D features, etc...

If a given machine doesn't have enough available GPU memory permanently, you want to set up a new venue (home, work, school) in you web preferences for it and opt-out of E@H GPU work.

HTH,
Oliver

 

Einstein@Home Project

Gundolf Jahn
Gundolf Jahn
Joined: 1 Mar 05
Posts: 1,079
Credit: 341,280
RAC: 0

RE: I don't think this is a

Message 96322 in response to message 96319

Quote:
I don't think this is a normal situation.


Does a restart of the BOINC client or a reboot change anything?

Gruß,
Gundolf

Computer sind nicht alles im Leben. (Kleiner Scherz)

Harri Liljeroos
Harri Liljeroos
Joined: 10 Dec 05
Posts: 1,230
Credit: 1,299,427,642
RAC: 1,720,355

RE: RE: I don't think

Message 96323 in response to message 96322

Quote:
Quote:
I don't think this is a normal situation.

Does a restart of the BOINC client or a reboot change anything?

Gruß,
Gundolf

Nope, neither operation helped. Situation is the same.

Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.