New BRP CUDA Apps 1.07 / 1.08
Bernd Machenschalk
Moderator
Administrator
Joined: 15 Oct 04
Posts: 4330
Credit: 251470102
RAC: 36353
18 Feb 2011 18:13:25 UTC
Topic 195659
(moderation:
A new series of CUDA Apps for Windows and Linux has been released (version 1.07). These should fix the CPU / GPU validation issues.
As for the experimental Linux App that requires the 270 beta driver, this requires an update of the scheduler. I plan to do this on Monday.
BM
BM
New BRP CUDA Apps 1.07 / 1.08
)
Hi Bernd,
good news. Was it intentional to set linux cuda tasks to 0.15 CPU?
edit: I am confused, the name of the app (here) says: BRP3cuda32nv270? And I am certain, that I installed 260.* driver...
edit2: As a side note, on my 9800GT runtime increased from 5000s to 6000s, but CPU time descreased from 5000s to 3000s.
Michael
Team Linux Users Everywhere
RE: Was it intentional to
)
As per Oliver Bock here:
Sorry, the Linux App
)
Sorry, the Linux App BRP3cuda32nv270 which was meant for the 270 driver only (and doesn't need a full CPU core) was already shipped due to an error in the current scheduler. Should be fixed (or at least worked around) now.
BM
BM
RE: Sorry, the Linux App
)
OK, I see. Nethertheless they seemed to run fine. (edit: and at least one validated.)
Michael
PS: Could someone point me to a thread discussing the initial problem with 260.*? I somehow missed it.
Team Linux Users Everywhere
Look here, where I got the
)
Look here, where I got the advice to install this driver.
The 27.18 beta driver reduced my error quote. I also added a remark at the end of this thread.
Hi! The initial problem
)
Hi!
The initial problem was this:
In a CUDA app, you can configure what the CPU thread should do while it is waiting for a computation on the GPU to finish (e.g. because it needs the results for a computation on the CPU, or because the next GPU computation needs the finished results as an input).
The most natural choices are:
-(actively) wait in a loop on the CPU: CPU core running this will show 100% load
- release the CPU: tell the operating system it can give the CPU(core) to another process. When the GPU is finished, the OS can again assign a CPU core to the CPU thread to continue.
Most BOINC GPU apps choose the second method: While it is increasing overall runtime of the GPU task a bit (because of a slighly longer delay when switching back to the CPU thread) it is freeing a CPU core to be used for CPU apps. Overall productivity should increase this way.
Ok, so what was the problem? On Linux, the drivers before 270.* had a bug that would cause sporadic errors in the computation of the BRP3 app only if you choose the second method (see above) of waiting for the GPU. This was detected already during the development phase of the app and did not occur "in the field", because the Linux app (as a workaround) selected the first method (full CPU required).
CU
HB
RE: edit2: As a side note,
)
i confirm: with 270 app cpu load decreased to 50-60% but time to complete a WU increased from ~4000 sec. to ~5000 sec. (GTX 560)
and that's mean what overall output of machine decreased =(
RE: RE: edit2: As a side
)
I can confirm that as well. Some numbers:
GForce9800 (for one WU):
Full CPU app: total runtime ~5570s, CPU time ~5500s, GPU load:~85%
270 app: total runtime ~6870s, CPU time ~2400s, GPU load ~66%
GTX470 (for two WUs at the same time):
Full CPU app: total runtime ~4600-4900s, CPU time ~4600-4900, GPU load:~85%
270 app: total runtime ~6000s, CPU time ~2800s, GPU load ~60%
So, the total runtime has increased, the CPU time and GPU load have decreased.
I still have to test if 3 WUs on the 470 are now possible will yield any gain in over all throughput.
Michael
*headscratch* Well, I found
)
*headscratch*
Well, I found the new Binary Radio Pulsar Search v1.07 (BRP3cuda32nv270) workunits today in the morning. And i found it was not the BRPS v1.06 (BRP3cuda32fullCPU) and also not the BRPS v1.07 (BRP3cuda32fullCPU). And I found this thread. And I installed the nvidia Beta driver after reading some posts. So far, so good.
My first nv270-workunit was partly crunched with the stable 260-er driver and then I installed the beta. It's this thing here: http://einsteinathome.org/workunit/93425612. Well, it's funny to see that the 270-er WU got crunched with the 260- driver, but this not my question.
I like to know that will happen now. Do I need to change something (e.g. install a e@h beta application) to receive more nv270-workunits? Or should I do nothing because I'm already using the nvidia beta and the sheduler will send the right workunits to my cruncher?
Okay, it seems that there is no e@h beta application, so my question is more hypothetical. But anyway: it would be nice to get some more information if possible. Thanks. :-)
Stephan
Hi! The crucial message is
)
Hi!
The crucial message is this from Bernd:
http://einsteinathome.org/node/195659&nowrap=true#110620
While you were still using the 260* driver, you should not have received the *270nv app, that was a mistake of the scheduler.
Now that you have installed the NVIDIA 270 driver, you will again receive work for the special nv270 app variant (the so called beta app, as it is targeted to a beta driver) once the scheduler is fixed next week. No further action is required for you.
If you should decide to go back to the stable 260 driver, you can do so, and then the scheduler should send you work for the "fullCPU" app variant automatically.
I hope this helps a bit to clarify the situation.
CU
HB