...which is strange because the runtime is anything but excessive.
Hmm...here's a question for the BOINC experts: how does BOINC decide, for a CUDA app, when the max. time is reached? I know that usually the Workunits will contain a estimated maximum number of floating point operations that processing the result could reasonably consume at worst. Now, with a CUDA app that takes 1 CPU plus 1 GPU in parallel, what does this tell BOINC? Will BOINC try to limit CPU time using the CPU benchmark, or will it use the estimated GPU performance (usually two orders of magnitudes greater than that of a single CPU core!!)??
For those of you suffering from "maximum elapsed time exceeded" error mesages, a solution is discussed in the thread on the Windows CUDA App http://einsteinathome.org/node/194499&nowrap=true#99856, it's the same problem for Linux.
For those suffering from "signal 11" (segmentation fault) crashes of their ABP1 CUDA apps with Linux drivers 185.x and 190.x, I stumbled over a possible workaround for the moment, but I'm a bit reluctant to share it here in public because I'd rather like to have this tested by one or two other volunteers before. If you are interested, drop me a PM.
For all of you who report CUDA-related problems, please do always include the BOINC version you are using. Please make sure that you use at least BOINC 6.6 because 6.4 still has several known CUDA-related issues.
For all of you who report CUDA-related problems, please do always include the BOINC version you are using. Please make sure that you use at least BOINC 6.6 because 6.4 still has several known CUDA-related issues.
Thanks,
Oliver
At least on my Debian 5 (stable) systems BOINC6.6 has the even worse issue,
that it does not connect to any project, therefore there seems to be no
newer useful version than 6.4 :-(
But this might be different with other distributions or experimental versions
of Debian or with the next stable version of Debian ...
At least on my Debian 5 (stable) systems BOINC6.6 has the even worse issue,
that it does not connect to any project, therefore there seems to be no
newer useful version than 6.4 :-(
But this might be different with other distributions or experimental versions
of Debian or with the next stable version of Debian ...
Interesting. Which 6.6.x BOINC version did you try? Did you use one of the official binary downloads or did you compile it yourself?
Interesting. Which 6.6.x BOINC version did you try? Did you use one of the official binary downloads or did you compile it yourself?
Cheers,
Oliver
Well I tried 6.6.36 as provided directly from BOINC on three computers
(two without GPU, one with GPU).
I did not install it over a previous version, just in a new directory as
for previous versions too to avoid possible problems with residuals from
previous versions.
Well, on two other computers I still use 5.10.21 because three computers use the
same home directory - and this seems not to work anymore with BOINC version 6,
but those computers have no GPU, therefore no need to update ;o)
What can be considered as stable or convenient for a normal user seems to
depend on the number and the arrangement of the computers used for BOINC at
the same time ;o)
I think, within Debian, version 6.4.5 is considered to be the current
experimental/unstable version, stable is 6.2.14.
However, what is considered as a stable or unstable NVIDIA driver at Debian,
was already to old for the GPU in the newest computer, therefore I had already
something to do to get the GPU work with the newest driver considered by
NVIDIA to be stable/worth to publish ;o)
Maybe there is a similar incompatibility with newer BOINC versions...
For all of you who report CUDA-related problems, please do always include the BOINC version you are using. Please make sure that you use at least BOINC 6.6 because 6.4 still has several known CUDA-related issues.
Thanks,
Oliver
At least on my Debian 5 (stable) systems BOINC6.6 has the even worse issue,
that it does not connect to any project, therefore there seems to be no
newer useful version than 6.4 :-(
But this might be different with other distributions or experimental versions
of Debian or with the next stable version of Debian ...
I use the stable version of Debian Lenny and 6.6.36 doesnt have problem connecting to any projects. Installed it by coping the files from the official boinc download into /usr/bin, iirc.
Im still using the 2.6.26-1-686-bigmem kernel, i see you have 2.6.26-2-686-bigmem but i doubt thats why.
After some connection problems again I managed now to get BOINC 6.6.41 work
including CUDA for Intel Core i7, GeForce GTX 260, driver 185.18.36
(the mashine already mentioned above) - hopefully with useful results now ;o)
RE: Most recent invalid is
)
It finished with "Maximum elapsed time exceeded"
Gruß,
Gundolf
Computer sind nicht alles im Leben. (Kleiner Scherz)
RE: RE: Most recent
)
...which is strange because the runtime is anything but excessive.
Hmm...here's a question for the BOINC experts: how does BOINC decide, for a CUDA app, when the max. time is reached? I know that usually the Workunits will contain a estimated maximum number of floating point operations that processing the result could reasonably consume at worst. Now, with a CUDA app that takes 1 CPU plus 1 GPU in parallel, what does this tell BOINC? Will BOINC try to limit CPU time using the CPU benchmark, or will it use the estimated GPU performance (usually two orders of magnitudes greater than that of a single CPU core!!)??
Thanks for any insights
CU
Bikeman
RE: I'm getting comp errors
)
I'm up over 50% failure rate now. Funny thing it doesn't seem to be the GPUgrid FFT error but a mix of things.
Think I need to put E@H on hold for a bit and see if the Collatz 64b Linux app works on this card.
UPDATE: Collatz seems to work once I figured out the right symlink to add. 2 of 2 since symlink fix and now under v6.10.11.
Hi! For those of you
)
Hi!
For those of you suffering from "maximum elapsed time exceeded" error mesages, a solution is discussed in the thread on the Windows CUDA App http://einsteinathome.org/node/194499&nowrap=true#99856, it's the same problem for Linux.
For those suffering from "signal 11" (segmentation fault) crashes of their ABP1 CUDA apps with Linux drivers 185.x and 190.x, I stumbled over a possible workaround for the moment, but I'm a bit reluctant to share it here in public because I'd rather like to have this tested by one or two other volunteers before. If you are interested, drop me a PM.
CU
Bikeman
Hi, For all of you who
)
Hi,
For all of you who report CUDA-related problems, please do always include the BOINC version you are using. Please make sure that you use at least BOINC 6.6 because 6.4 still has several known CUDA-related issues.
Thanks,
Oliver
Einstein@Home Project
RE: Hi, For all of you who
)
At least on my Debian 5 (stable) systems BOINC6.6 has the even worse issue,
that it does not connect to any project, therefore there seems to be no
newer useful version than 6.4 :-(
But this might be different with other distributions or experimental versions
of Debian or with the next stable version of Debian ...
Hi Olaf, RE: At least
)
Hi Olaf,
Interesting. Which 6.6.x BOINC version did you try? Did you use one of the official binary downloads or did you compile it yourself?
Cheers,
Oliver
Einstein@Home Project
RE: Hi Olaf, Interesting.
)
Well I tried 6.6.36 as provided directly from BOINC on three computers
(two without GPU, one with GPU).
I did not install it over a previous version, just in a new directory as
for previous versions too to avoid possible problems with residuals from
previous versions.
Well, on two other computers I still use 5.10.21 because three computers use the
same home directory - and this seems not to work anymore with BOINC version 6,
but those computers have no GPU, therefore no need to update ;o)
What can be considered as stable or convenient for a normal user seems to
depend on the number and the arrangement of the computers used for BOINC at
the same time ;o)
I think, within Debian, version 6.4.5 is considered to be the current
experimental/unstable version, stable is 6.2.14.
However, what is considered as a stable or unstable NVIDIA driver at Debian,
was already to old for the GPU in the newest computer, therefore I had already
something to do to get the GPU work with the newest driver considered by
NVIDIA to be stable/worth to publish ;o)
Maybe there is a similar incompatibility with newer BOINC versions...
RE: RE: Hi, For all of
)
I use the stable version of Debian Lenny and 6.6.36 doesnt have problem connecting to any projects. Installed it by coping the files from the official boinc download into /usr/bin, iirc.
Im still using the 2.6.26-1-686-bigmem kernel, i see you have 2.6.26-2-686-bigmem but i doubt thats why.
Team Philippines
After some connection
)
After some connection problems again I managed now to get BOINC 6.6.41 work
including CUDA for Intel Core i7, GeForce GTX 260, driver 185.18.36
(the mashine already mentioned above) - hopefully with useful results now ;o)