Forum > Linux
Help installing CUDA version
sunu:
--- Quote from: BeemerBiker on 15 Feb 2010, 01:45:04 pm ---I got it working finally. Not knowing a lot about the inner workings of ubuntu, I added the phrase "include" in ld.so.conf where I should have just added /usr/local/cuda/lib64 by itself. There was only one line in ld.so.conf and that had "include" so I assumed I needed to use include for the cuda libraries.
--- End quote ---
Revert ld.so.conf to its previous condition, that is delete anything you've written. In /etc there is a directory called ld.so.conf.d . In there create a file with a .conf extension, for example cudalibs.conf. In that file write the path to the cuda libraries, without include, without anything. Finally run ldconfig to update the cache.
--- Quote from: BeemerBiker on 15 Feb 2010, 01:45:04 pm ---Do you know why the app_info has the two cuda llibs listed? Boinc attempted to download them and of course they failed.
--- End quote ---
They need to be there, like the names of the other executables.
You'll also need to raise the priority of the cuda processes. If you don't know how to do this, use the script I attach below (you'll need to install the packages procps and bsdutils).
[attachment deleted by admin]
BeemerBiker:
Thanks SUNU
I looked at the resource monitor and tried adjusting the priority from 10 to 0 but that does not work as well as allocating 3 of 4 cpu's to boinc. ie: the time to complete visiibly counts down faster when a full cpu is reserved. SETI CUDA seems to use only %4 of the cpu. I can see where setting the priority higher can make a difference. Einstein uses a full cpu and gpugrid is almost as bad. On those projects, the time to comlplete can rise from 12 to 24 houirs or more especially for gpugrid. Reserving a cpu for them seems to be the best fix. Collatz could possibly use the priority chainging scheme.
IMHO Boinc should set the priority of the seti task when it dispatches it. One would think that a low cpu task like seti cuda could have its piriority increased while a cpu hog like einstein's cuda could stay at a nice 19.
I am quite pleased with the outcome of this SETI CUDA program. Here is a comparision of two system. I have a 9800gtx+ on vista 64 with a pair of opterons running at 2.2ghz. The second system is the linux x64 cuda with GTS-250 and the opteron pair run only a 2ghz. I tried matching the seti cuda work unit as best as I could, the vlar for the linux64 is .448 and the vlar for the vista64 is .382 . The linux system completed its work unit in about 1/2 the time that the "faster" vista64 system
Linux CUDA time
Run time 1452.497664
CPU time 139.17
http://setiathome.berkeley.edu/result.php?resultid=1516909565
Vista CUDA time
Run time 2430.243868
CPU time 564.9063
http://setiathome.berkeley.edu/result.php?resultid=1512627579
I recall one of the 3rd party boinc apps (Raistmer's?) could set the priority
sunu:
Those angle ranges are not comparable. The sad truth is that the linux cuda client is slower that the windows one. The cuda client in linux is based on a "first generation" code, while in windows there have been small tweaks in the code and several releases of the optimized client.
Navigation
[0] Message Index
[*] Previous page
Go to full version