First CUDA App for Windows available for Beta Test

Richard Haselgrove
Richard Haselgrove
Joined: 10 Dec 05
Posts: 2,140
Credit: 2,770,592,658
RAC: 909,861

RE: So, that sez ( ?? ) -

Message 94129 in response to message 94128

Quote:

So, that sez ( ?? )

- started blurb_451_1
- started blurb_476_1
- blurb_476_1 falls over
- blurb_451_1 falls over

..... concurrent.

Cheers, Mike.


You missed out:

- started blurb_451_1
- started blurb_476_1
- blurb_476_1 falls over
- Resuming blurb_451_1
- blurb_451_1 falls over

..... consecutive. (Recent BOINCs don't tell you when they're preempting something, unless you use debug logs).

Your version of BOINC will always try to run CUDA jobs in FIFO order. With extreme prejudice: if an earlier job becomes available, it will immediately pre-empt any later job. I suspect your second allocated job finished downloading first, and started to run: then got preempted by the first job, when its download finished second.

That caused problems in v6.6.36 - if the pre-empt comes very early in the task life, before the first checkpoint, the preempted task was left in VRAM. Not a good choice. That's now been fixed in BOINC v6.6.37 (memory is cleared, ready for the new task).

With your 1GB card, that shouldn't have been a problem, but it would be worth trying to run one at a time (maybe a reboot first, to fully reset the graphics card). If that doesn't work, it's back to the driver hunt.

Mike Hewson
Mike Hewson
Moderator
Joined: 1 Dec 05
Posts: 6,537
Credit: 286,524,051
RAC: 93,547

RE: You missed out: -

Message 94130 in response to message 94129

Quote:

You missed out:

- started blurb_451_1
- started blurb_476_1
- blurb_476_1 falls over
- Resuming blurb_451_1
- blurb_451_1 falls over

..... consecutive. (Recent BOINCs don't tell you when they're preempting something, unless you use debug logs).


Whoops .... :-)

Quote:

Your version of BOINC will always try to run CUDA jobs in FIFO order. With extreme prejudice: if an earlier job becomes available, it will immediately pre-empt any later job. I suspect your second allocated job finished downloading first, and started to run: then got preempted by the first job, when its download finished second.

That caused problems in v6.6.36 - if the pre-empt comes very early in the task life, before the first checkpoint, the preempted task was left in VRAM. Not a good choice. That's now been fixed in BOINC v6.6.37 (memory is cleared, ready for the new task).

With your 1GB card, that shouldn't have been a problem, but it would be worth trying to run one at a time (maybe a reboot first, to fully reset the graphics card). If that doesn't work, it's back to the driver hunt.


Yup, extreme prejudice all right.
You're a veritable gold mine, nay Niobium mine ( that's more valuable ), of information Richard! ;-)

I will :

- slip in 6.6.37
- reboot
- shake 'n bake

Cheers, Mike.

I have made this letter longer than usual because I lack the time to make it shorter ...

... and my other CPU is a Ryzen 5950X :-) Blaise Pascal

Richard Haselgrove
Richard Haselgrove
Joined: 10 Dec 05
Posts: 2,140
Credit: 2,770,592,658
RAC: 909,861

RE: You're a veritable gold

Message 94131 in response to message 94130

Quote:
You're a veritable gold mine, nay Niobium mine ( that's more valuable ), of information Richard! ;-)


That's the advantage of the Einstein approach to development. Arriving on the scene second (or fifth, if my maths is right), you get the benefit of all the mistakes we've been making at SETI for the last eight months :-)

(and avoid some of the mistakes the BOINC developers have made along the way, and all the time we've wasted working things out for ourselves).

Here's another nugget: if you have any spare (preferably unstarted) 308 ABP1 tasks, but the server won't send you any 307 CUDA ones because of quota, you can apply the alchemists' stone and turn CPU into CUDA. (Involves editing client_state.xml: anyone not happy with doing that, stop reading now).

Stop BOINC, make backup, yadda yadda, usual precautions.

Find the section and the section for the task you want to transmogrify. They both have a version number: change 308 to 307 in both of them. Make sure you've chosen matching sections - the 'blurb' must match exactly.

Into the section (only - not the ) - insert a complete new line

cuda

as seen in app_info.xml: convention has it that the line goes immediately after the version number.

Save and restart - best of luck with the new driver (at least that will have taken care of the reboot.....)

samuel7
samuel7
Joined: 16 Feb 05
Posts: 34
Credit: 1,579,363
RAC: 0

Hi Mike! If you can get

Message 94132 in response to message 94128

Hi Mike!

If you can get the CUDA app running with that driver version please report.

To this point, I haven't seen anyone reporting a successful run on a 64-bit Vista/Win7. Please correct me if I'm wrong.

BeemerBiker reported his failure with some screen captures here. I got the same Runtime Error dialog on my Vista64 (driver 190.38).

Edit - To clarify: As far as BOINC is concerned the task is "running" until the Visual C++ error dialog is closed. Then it throws the BOINC error with log notes 'Computation finished' and 'Output file xxx absent'. This could explain the first task being preempted and erroring out only after resumption in your BOINC log.

Mike Hewson
Mike Hewson
Moderator
Joined: 1 Dec 05
Posts: 6,537
Credit: 286,524,051
RAC: 93,547

@Richard: RE: That's

Message 94134 in response to message 94131

@Richard:

Quote:

That's the advantage of the Einstein approach to development. Arriving on the scene second (or fifth, if my maths is right), you get the benefit of all the mistakes we've been making at SETI for the last eight months :-)

(and avoid some of the mistakes the BOINC developers have made along the way, and all the time we've wasted working things out for ourselves).


Ah, you are wise beyond your years GrassHopper :-) ;-D

Quote:
Here's another nugget ......Save and restart - best of luck with the new driver (at least that will have taken care of the reboot.....)


Ah, well the servers are having a little lie down at present. So I have duly mounted this piece in The Shiny Things Cabinet within The Good Room, 'til the server migraine passes. :-)

@Samuel:

Quote:
To this point, I haven't seen anyone reporting a successful run on a 64-bit Vista/Win7. Please correct me if I'm wrong.


Neither did I, so I thought I'd throw my box into the melee. It's a new home-built system barely run in yet, and I thought it was time to teach it ice skating. :-)

Quote:
BeemerBiker reported his failure with some screen captures here. I got the same Runtime Error dialog on my Vista64 (driver 190.38).


That's precisely how mine errored out.

Quote:
Edit - To clarify: As far as BOINC is concerned the task is "running" until the Visual C++ error dialog is closed. Then it throws the BOINC error with log notes 'Computation finished' and 'Output file xxx absent'. This could explain the first task being preempted and erroring out only after resumption in your BOINC log.


Indeedy do. Extreme prejudice alright! I wouldn't want to be standing in the same line at the supermarket checkout with it. Thanks for the tips. :-)

While I'm about I'll re-post my current settings, as per BOINC Manager's understanding:

Quote:
Starting BOINC client version 6.6.36 for windows_x86_64
log flags: task, file_xfer, sched_ops
Libraries: libcurl/7.19.4 OpenSSL/0.9.8j zlib/1.2.3
Data directory: D:\ProgramData\BOINC
Running under account Mike
Processor: 8 GenuineIntel Intel(R) Core(TM) i7 CPU 920 @ 2.67GHz [Intel64 Family 6 Model 26 Stepping 5]
Processor features: fpu tsc pae nx sse sse2 pni mmx
OS: Microsoft Windows 7: x64 Edition, (06.01.7100.00)
Memory: 5.99 GB physical, 11.98 GB virtual
Disk: 927.50 GB total, 910.03 GB free
Local time is UTC +10 hours
CUDA device: GeForce 9800 GT (driver version 18635, compute capability 1.1, 1024MB, est. 60GFLOPS)
Einstein@Home Found app_info.xml; using anonymous platform
Not using a proxy
Einstein@Home URL: http://einstein.phys.uwm.edu/; Computer ID: 1992248; location: (none); project prefs: default
Einstein@Home General prefs: from Einstein@Home (last modified 30-Jun-2009 21:37:50)
Einstein@Home Host location: none
Einstein@Home General prefs: using your defaults
Reading preferences override file
Preferences limit memory usage when active to 3067.56MB
Preferences limit memory usage when idle to 5521.61MB
Preferences limit disk usage to 5.00GB

Cheers, Mike.

[ edit ] I couldn't find 6.6.37, it wasn't on this download page. It's in testing at the moment ?

I have made this letter longer than usual because I lack the time to make it shorter ...

... and my other CPU is a Ryzen 5950X :-) Blaise Pascal

Mike Hewson
Mike Hewson
Moderator
Joined: 1 Dec 05
Posts: 6,537
Credit: 286,524,051
RAC: 93,547

RE: Ah, well the servers

Message 94135 in response to message 94134

Quote:
Ah, well the servers are having a little lie down ....... 'til the server migraine passes. :-)


I apologise profusely. In future I will watch my words far more carefully.

Cheers, Mike.

I have made this letter longer than usual because I lack the time to make it shorter ...

... and my other CPU is a Ryzen 5950X :-) Blaise Pascal

stormdog
stormdog
Joined: 3 May 05
Posts: 5
Credit: 282,118,623
RAC: 2,988

Hello, After downtime, all

Message 94136 in response to message 94135

Hello,

After downtime, all my CUDA units error out instantly

6.6.36

������� �� ������� ����� ��������� ����. (0x3) - exit code 3 (0x3)

Activated exception handling...
[09:29:00][4928][INFO ] Starting data processing...
[09:29:00][4928][INFO ] Using CUDA device #0 "GeForce 9600 GT" (307.20 GFLOPS)
[09:29:01][4928][INFO ] Checkpoint file unavailable: status.cpt (No such file or directory).
------> Starting from scratch...
[09:29:01][4928][INFO ] Header contents:
------> Original WAPP file: p2030_54100_59289_0040_G33.20+02.29.C_2.wapp
------> Sample time in microseconds: 128
------> Observation time in seconds: 268.9792
------> Time stamp (MJD): 54100.686215277776
------> Number of samples/record: 512
------> Center freq in MHz: 1440
------> Channel band in MHz: 0.390625
------> Number of channels/record: 256
------> Nifs: 1
------> RA (J2000): 184322.603388
------> DEC (J2000): 11719.910885
------> Galactic l: 33.1606
------> Galactic b: 2.3811
------> Name: G33.20+02.29.C
------> Lagformat: 0
------> Sum: 1
------> Level: 3
------> AZ at start: 356.962
------> ZA at start: 17.0593
------> AST at start: 0
------> LST at start: 0
------> Project ID: p2030
------> Observers: JuliaDeneva
------> File size (bytes): 16190754
------> Data size (bytes): 16179201
------> Number of samples: 2097152
------> Trial dispersion measure: 198 cm^-3 pc
------> Scale factor: 6439.28
[09:29:02][4928][INFO ] Seed for random number generator is -1139401116.
[09:29:03][4928][INFO ] Derived global search parameters:
------> f_A probability = 0.04
------> single bin prob(P_noise > P_thr) = 1.2977e-008
------> thr1 = 18.1601
------> thr2 = 21.263
------> thr4 = 26.2923
------> thr8 = 34.674
------> thr16 = 48.9881
[09:29:04][4928][ERROR] Error during CUDA host->device data transfer (unspecified launch failure)

]]>

No changes to harware or software were done on my side during downtime. Here are task numbers 136757595, 136757589, 136757526, 136757502, 136757435, 136757416, 136757295 and there are more.

Regards,
Andrew

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

RE: Hello, After downtime,

Message 94137 in response to message 94136

Quote:

Hello,

After downtime, all my CUDA units error out instantly...


Did you try a reboot to clear graphics memory?

Gruß,
Gundolf

Computer sind nicht alles im Leben. (Kleiner Scherz)

stormdog
stormdog
Joined: 3 May 05
Posts: 5
Credit: 282,118,623
RAC: 2,988

RE: Did you try a reboot to

Message 94138 in response to message 94137

Quote:
Did you try a reboot to clear graphics memory?

It helped.

Thanks,
Andrew

Bernd Machenschalk
Bernd Machenschalk
Moderator
Administrator
Joined: 15 Oct 04
Posts: 4,273
Credit: 245,219,601
RAC: 12,936

We posted a new CUDA App for

We posted a new CUDA App for Windows that fixes a bug and should be a little faster on the CPU, too. See here.

BM

BM

Comment viewing options

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