Computer Laboratory

A number of changes have happened at the same time with the processor bank machines.
  • The machines have had extra memory added (half doubling to 8GB, half quadrupling to 16GB)
  • The OS has been changed to Fedora Core 6 (FC6)
  • Condor has being upgraded to 6.8
  • The condor daemons no longer run as condor
  • All jobs should be submitted on condor-submit

Condor is being installed in a different place, so you need to make sure you have the condor programs on your PATH:
PATH=/opt/condor-6.8.3/bin:$PATH;export PATH

Under Condor 6.6 it was necessary to set CONDOR_CONFIG. Under 6.8 it isn't necessary as it will pick up a default. However, if you have previously used Condor 6.6 you will have it set to the wrong value which will override that default. So you need to make sure that CONDOR_CONFIG is unset.

A major bug in Condor 6.6 related to memory management - it just didn't work. By default Condor attempts to calculate how much memory each machine has. Because version 6.6 got it wrong it was configured not to do this and the amount of memory (actually just a very large number) was hardwired in. Condor 6.8 appears to get the memory calculations correct, so it is being allowed to do them itself. Thus it will know which of the machines will have 8GB and which 16GB. (As processor bank machines have two CPUs Condor divides the memory between them, possibly asymetrically, giving 4, 8 and 12 GB virtual machines). If it matters to your application then you can select the larger machines by making an explicit requirement. Give your requirements along with "rank = -memory" so that it uses the machine with the smallest memory which is sufficient - don't grab the largest machine.

Condor 6.8 has support for 64 bit machines. Thus cl-matlab (matlab is installed as cl-matlab on FC6 machines) should no longer have the -glnx86 flag, and it is no longer necessary to use -m32 -static-libgcc with gcc when compiling for the standard universe although it may be necessary to recompile binaries to run on the new pool. "getenv = True" is not needed for cl-matlab

A later version of Java is being used (currently 1.6, aka Java 6). The Java universe automatically used maxheap flags (-Xmx) under 6.6. These have not been set under 6.8 because of the improved memory management, but they may turn out to be necessary. (The test jobs I have are too small to test this properly). It uses the default (1GB) unless explicitly set. Also note that the command "java" (but NOT javac) is a 1.4.2 command - use /usr/java/default/bin/java if testing before use.

Condor_submit is a little more fussy about the location of log files, and may issue warnings about their location (WARNING: Log file blah.log is on NFS). The warning messages can be safely ignored.

Note that there are now only two submit machines: condor-submit0 and condor-submit1, with a shared name of condor-submit.