Bobsced Cluster

From Earlham CS Department
Jump to: navigation, search

Todo

  • 411 tools
    • fix ganglia to recognize broadcasts & update
  • Naming scheme
    • bs* vs compute-*-* vs c*-*
    • This is terrible it needs work
  • Updating bobsced0's RPM repo
    • yum-- free
    • up2date-- RHEL
    • "Aborting the rocks-update tool while the tool is downloading RPMs might produce corrupted RPM packages (SDSC Toolkit)" from pr_troubleshooting.doc
  • NIS map
    • /etc/passwd & /etc/group permissions
    • An architecture without a variable amount of delay before BobSCEd is updated would be nice.
  • What broke cluster-fork?
  • NAT & NFS
    • Mailing list
    • Mailing list
    • We should consider flattening the network, that is moving everything into the 159.28.234/24 subnet.
  • Stop the hourly cron from producing output on stdout unless there is an error
  • Setup and test Infiniband fabric
  • Why does hopper see an interface flap from bobsced0?
  • bosced0 wants to be the DNS server for the compute nodes

Howtos

Updating nodes to be kickstarted & adding new packages

  • bobsced0 can be updated by just installing rpms
  • Check for an RPM in: /state/partition1/home/install/rocks-dist/lan/x86_64/RedHat/RPMS/
  • Edit /home/install/site-profiles/4.1.1/nodes/extend-compute.xml
    • Add a package i.e. <package arch="x86_64">libgfortran</package>
  • Update the files that get loaded on kickstart:
    • cd /home/install
    • rocks-dist dist
  • Check the kickstart file
    • dbreport kickstart c0-0
  • If there were no errors, kickstart the node. i.e.:
    • shoot-node c0-0
  • Check the progress of a kickstart
    • ssh –p 2200 compute-x-x

Adding post install scripts to kickstart

  • Edit /home/install/site-profiles/4.1.1/nodes/extend-compute.xml
  • Add a <post arch="x86_64"> entry i.e.:
    • <post arch="x86_64">cp /cluster/ganglia/gmond.conf /etc/gmond.conf</post>

Using 411 tools

  • make -C /var/411 on bobsced0
    • Copies the files to /etc/411.d/ using 411put
    • Notifies client nodes to run 411get using ganglia
    • The files that are watched can be updated by changing the makefiles in /var/411/
  • cluster-fork /opt/rocks/bin/411get --all

cluster-fork

  • Used to run commands on all cluster nodes like the c3tools
    • Broken, see todo
    • Temporary fix: cluster-fork --nodes="compute-0-%d:0-14" <command>

disabling reinstall (kickstart) after hard reset

Generating an up-to-date machinefile

Ethernet:

cluster-fork /sbin/ifconfig -a | grep -1 Ethernet | awk '{printf("%s slots=4\n",$2)}' | cut -d : -f 2 > bs-eth-hosts

Infiniband:

cluster-fork /sbin/ifconfig -a | grep -1 UNSPEC | awk '{printf("%s slots=4\n",$2)}' | cut -d : -f 2 > bs-ib-hosts 

Notice that the only difference is the search field in the first grep command. UNSPEC here refers to Infiniband.

General Info

NIS Importing

  • /etc/cron.hourly/importNIS.sh
  • This comes from the rocks users guide & a mailing list thread.

http

  • /cluster/www/bobsced/

/cluster

  • Mounted using /etc/rc.local

/cluster/bobsced/etc/

  • What's in here? Things for client or bobsced0?

Known Error Messages

Please refer here if you encounter an error message on BobSCEd that you cannot handle.

RLIMIT_MEMLOCK

$./area_mpi
libibverbs: Warning: RLIMIT_MEMLOCK is 32768 bytes.
This will severely limit memory registrations.--------------------------------------------------------------------------
The OpenFabrics (openib) BTL failed to initialize while trying to
allocate some locked memory.  This typically can indicate that the
memlock limits are set too low.  For most HPC installations, the
memlock limits should be set to "unlimited". The failure occured
here:
 Local host:    bobsced0
 OMPI source:   btl_openib_component.c:1040
 Function:      ompi_free_list_init_ex_new()
 Device:        mthca0
 Memlock limit: 32768
You may need to consult with your system administrator to get this
problem fixed.  This FAQ entry on the Open MPI web site may also be
helpful:
   http://www.open-mpi.org/faq/?category=openfabrics#ib-locked-pages
--------------------------------------------------------------------------
--------------------------------------------------------------------------
WARNING: There was an error initializing an OpenFabrics device.
 Local host:   bobsced0
 Local device: mthca0
--------------------------------------------------------------------------

This error is specifically related to infiniband on BobSCEd. If you're using ethernet, specify as such by creating a file,

~/.openmpi/mca-params.conf

with the first line being

btl = ^openib

This will tell BobSCEd not to try infiniband, and will stop the error. If you are using Infiniband and find a solution to this error, please place it in the wiki.

References

Rocks Documentation

Troubleshooting Platform Open Cluster Stack (OCS) and Platform Lava

  • pr_troubleshooting.doc

411 Tools

RHEL