Suckered ... again!

I decided to upgrade the nVidia driver on the Celeron/8600GT/Windoze machine to the latest available to be able to run GPUGrid and Einstein on this machine. So, I download the driver and install it, a reboot follows ... and no CUDA task starts up. They are all there with "Ready to Start" next to them but nothing is running. After a minute or two of messing around with no success, I decide that the driver I have installed hasn't installed properly. I uninstall and install an older driver that I have used before ... no luck. So I uninstall again and after all the reboots, go back to the original 182.50 version. Still no luck. I then upgrade to the latest BOINC client, still nothing. At this stage I resort to Google and find this little gem : "If every cpu is running in "high priority mode" the cuda task are switched to "waiting to run""

Aborted a QMC task that was busy and all is back to normal!

In the words of Homer Simpson ... "D'oh!"

One more thing...

I set about upgrading my home Ubuntu 64-bit PC in the manner I listed below and it took me some time to realize why SETI kept trying to download its stock CPU application. When I copied the files into the project directory, I hadn't checked the permissions and it turns out that they were only visible to 'root'.
I should have picked that up much sooner from the fact that the Boinc Manager wasn't picking up the "app_info.xml" file ...
Anyway, the Leadtek 9400GT is also crunching on CUDA2.2 now!

CUDA 2.2 VLAR Kill SETI application for Ubuntu??

Well, there is still no CUDA application for Linux from SETI and ever since I started with a nVidia card, I have been using the application from Crunch3r. There are two little issues with this application, the first is that it took around 100% of one of my CPU cores to run and secondly, VLAR (Very Low Angle Range) work units ran and took ages to complete.
While trolling the internet at lunch time, I came across a forum post on Crunch3r's site that mentions a Linux application with a VLAR Kill option. I have been using the Windoze equivalent and decided to pursue this.
A bit more reading revealed that the CUDA 2.2 version uses a lot less CPU time than the original application.

So, here is how I got it to work (there are probably easier/cleaner ways of doing it, but this is how I got to the end result).

First, you need the CUDA 2.2 libraries. I got them from the Nvidia site (http://www.nvidia.com/object/cuda_get.html). They don't list a CUDA 2.2 toolkit for 64-bit Ubuntu 9.04, so I download the one for Ubuntu 8.10.
This gives you a ".run" file which confused me a bit to start with. I got it installed with the following:

Navigate to the directory of the .run file. I had mine on the desktop so I used "cd ~/Desktop".
Then "chmod +x cudatoolkitxxx.run"
And "./cudatoolkitxxx.run"
I just used the default installation directories.

The next stage is to copy the required files from the CUDA Toolkit installation into the projects/setiathome.berkeley.edu directory. You will need:

libcudart.so
libcudart.so.2
libcudart.so.2.2
libcufft.so
libcufft.so.2
libcufft.so.2.2

The actual application comes from a link on Crunch3r's site (http://calbe.dw70.de/mb/viewtopic.php?f=9&t=116). I used setiathome-CUDA_2.2_6.08.x86_64_vlarkill.

Here is one part that I'm not sure about but it seems to be necessary. You need to copy the application to /usr/bin as well as projects/setiathome.berkeley.edu .This is a prime example of where there are better ways of doing things!

You will need an app_info.xml file in projects/setiathome.berkeley.edu to get it all to work. What I did here was take the included file from Crunch3r's normal CUDA app (http://calbe.dw70.de/linux64.html) and edit it, changing the name of the application to match setiathome-CUDA_2.2_6.08.x86_64_vlarkill. I left the names of the CUDA library files as they were.
Once the app_info.xml file was in the correct directory, I restarted the BOINC client and amazingly it all seems to work fine with a CPU usage rate of 2%. This means I have almost a whole additional core to run CPU work units. This is still using the 6.4.5 Boinc client

Credit to Crunch3r and sunu (on KWSN forums) for the info they have posted that got me through this!

Upgrading nVidia drivers in Ubuntu to ver 185.18.36

All seems to be going well with the Leadtek 9800GT so far, however GPUGrid work units started crashing at a good rate of knots this morning. Their home page carries the ominous phrase "The new Linux application is now based on CUDA2.2 and requires drivers 185.xx."
Up until now I have been using the 180.x driver that is in the Ubuntu repositories ... but alas, nothing newer than that available.

I resorted to using the following to get it to work (all entered into a terminal window):

sudo sh -c "echo 'deb http://ppa.launchpad.net/nvidia-vdpau/ppa/ubuntu jaunty main' >> /etc/apt/sources.list"
sudo sh -c "echo 'deb-src http://ppa.launchpad.net/nvidia-vdpau/ppa/ubuntu jaunty main' >> /etc/apt/sources.list"

Followed by (for GPG key):

sudo apt-key adv --keyserver keyserver.ubuntu.com --recv-keys CEC06767

And finally:

sudo apt-get update && sudo apt-get install nvidia-185-modaliases nvidia-glx-185

Everything seemed to work fine after a reboot, except for the Boinc client trashing all the previously downloaded SETI Cuda work units. It did, however download 3 new GPUGrid work units and seems to be crunching merrily.

So far, so good ...

The Leadtek 9800GT has been running for a good couple of hours now (it was turned off over the weekend however) and all seems to be good. It is doing a SETI work unit in about 40 minutes (actual wall time) and seems to have a steady temperature of 70'C which is pretty good as there is no case fan in this machine yet.
The Galaxy was also stable for the first couple of days before it started becoming unreliable, so I won't hold my breath just yet.

Leadtek 9800GT

Well, a Leadtek 9800GT arrived at the office via a courier yesterday afternoon. There are several differences between this card and the original Galaxy. The first and most obvious is that the Leadtek doesn't have plastic cowling that allows hot air to be expelled towards the DVI plugs and out of the case (this may present a problem in summer). The other difference that has me a bit confused is that the Leadtek doesn't require the 6-pin power cord that most the faster cards do. I'm wondering if this means it is a low power version.

I have elected to install it in my AMD X2 as it is connected to a DVI capable monitor. This machine runs Ubuntu and used to be home to the Galaxy 8600GT. The old 8600GT has been moved to the Windoze Celeron as it has a VGA plug.

The new card has been running overnight and has completed a couple of SETI work units as well as a GPUGrid work unit. I won't get too excited yet as the Galaxy only started presenting problems after a couple of weeks.

9800 continued...

Got an email from the computer store to say that they can't find anything wrong with the Galaxy card, however they will swap it for a Leadtek equivalent. I'm happy with that as the 9400GT Leadtek hasn't given a moments trouble since I put it in the same machine that the Galaxy 9800GT was in.
Hopefully I will have the replacement by the end of the week.

2.5 Million!

Well, reached the 2.5M credit mark this morning. My average credit has dropped to below 9k/day ... so the next mile-stone will be some way off. SETI@Home is short on work so the two slower CUDA cards aren't even earning their keep.
The 9800GT is on its way back to the supplier this morning via the post office. I'm only about 50% sure I will see it again...