Custom Query (981 matches)

What follows is rather lengthy and may only apply to Fedora-10-preview. Thanks in advance to anyone who chooses to wade thru it. This material perhaps only effects an extremely small user community.

The rev. boinc-5 has a long successful history here on Fedora 8. While preparing to upgrade the OS, boinc ran for months with working dir /db/BOINC. On this host, /db is a mount point for mysql, http, and other server data. After upgrading F8 to F10 and boinc-5 to boinc-6.2.15-1.20080818svn.fc10 the normal adjustment of BOINCDIR=/db/BOINC in /etc/sysconfig/boinc failed to assign the working dir. In addition frequent failures of boinc_cmd showing "Authorization failure: -155" and boincmanager made boinc quite unuseable.

As part of the troubleshooting, the SETI project was installed successfully while the working dir was set to /var/lib/boinc, but no work unit was queued for processing. Partially completed work for SETI and climate remained untouched under /db/BOINC.

To work around the script failure, the /etc/init.d/boinc-client script was revised to point to /db/BOINC without need for the sysconfig script -- this failed as well. As a "last ditch" effort, /var/lib/boinc was made into a symbolic link to /db/BOINC and this provided some limited success.

There is some indication that the problem is not with boinc software. The installer creates a "boinc" user with /var/lib/boinc as working dir -- this is what I saw here. After I assigned a pwd to boinc, tty login went normally. During troubleshooting the boinc user was helpful in monitoring the contents of the home directory. At the time of installing the symbolic link there was a major problem.

The procedure of installing the link was: log-off boinc user use another console as user root to do the following sequence: cd /var/lib ls -l | more mv boinc boinc.tmp ln -s /db/BOINC boinc ls -l | more ls boinc service boinc-client start After the previous sequence boinc-client failed to process the backlog from /db/BOINC queue.

Furthermore, when user boinc logged in again, the listing of home dir contents was identical to the former /var/lib/boinc contents and this is clearly wrong. After the sequence of commands: cd .. ls [This displayed bogus contents for /var/lib] cd .. ls [This displayed suitable contents for /var] cd lib/boinc ls [This displayed contents of the /db/BOINC directory as expected] The above series indicates [to me] a basic error in directory management.

The system today runs with selinux =disabled and iptables STOPPED, these settings will change soon. Boinc-client is processing the backlog from prior to the upgrade. Boinc_cmd usually works, but GUI boincmanager usually fails. Thus boinc is half healthy. My attempts to replicate and isolate errors are not entirely successful. However there seems to be an undocumented(?) dependency that requires "/bd/BOINC" assignment in /etc/passwd>boinc[working directory]. If in fact there is a requirement to assign the boinc user home directory identical to the boinc-client working directory, then perhaps this is overly restrictive?

In addition to the boinc software requirement(?) of the previous paragraph, today F10 seems to prevent adjustment to /etc/passwd from becoming immediately effective. I have not seen such delay in previous years. A system reboot will make the change effective, and sometimes waiting for 15(?) minutes will allow the old assignment to expire. I am taking this to the Fedora team for review.

Thus I have developed only an attitude that perhaps early F10-Preview has been a surprise. As for isolating a problem with boinc software, there is no "smoking gun". All critical reviews of this work are welcome, and suggestions will be tried here.

Currently BOINC 6.4.5 requires a "feeder" task to run the coprocessor. On a quad core this means BOINC defaults to 3 cpu tasks + 1 gpu task. This can be overridden to use all 4 cores via the <ncpus> in cc_config so you get 4+1 tasks running.

The issue comes when it runs out of gpu tasks it tries to run 5 cpu tasks, resulting in having to share a cpu with multiple science apps. Alter the scheduler to run the "feeder" task without it taking an entire core (ie share with a science app) and when out of gpu work resort back to the 4 cpu cores. If its a coprocessor "feeder" task then don't count it as a cpu task.

I get computations errors on a laptop using the GPU's. After about 10% work completion it will fail with an computation error. Running the latest video drivers. Windows Vista x64, BOINC 6.4.5 x64, 2 x 8800M GPU's. I only see this with SETI@Home workunits.