Elliot Jaffe: EverLab Progress

Status

• www.everlab.org
• Running: evgsicsstg.sics.se
  • Installed AFS server

  TODO

  • Make openafs rpms part of the standard install package
  • Fix cache setup for openafs (/etc/openafs/cacheinfo and /etc/init.d/afs)
  • Create code to automatically mount /afs in each slice via mount --bind

  Steps to Make EverLab work

  • TAU
    • Done - Trial installation at TAU (B10)
    • Done - Get TAU servers moved to outside of firewall
  • Test AFS as a distributed filesystem under PlanetLab
    • Done - Build and install on FC4_4.
    • Done - Use mount --bind to put /afs on each slice
    • Done - Install Server on SICS storage node (18/12/05, node ready)
    • Done - test client on planet9
    • Add to root distribution
  • Implement PXE Boot for Clusters
    • Done - Build initrd
    • Done - Trial installation on evgsics13
      • FIXED - Cannot write to /tmp (its r/o because its from the initrd. ) Need to add "rw" the end of the INITRD field of the PXE Boot config file
      • Tested on local machine (complex-1). It works correctly if you use dhcp to find the network.
      • Done - Document process for use everywhere
    • Done - Implement CPU quotas
      • Done - Send mail to devel@planet-lab.org,
        Use the variable general_prop_share, which can also be called nm_cpu_share. "It lets you assign a weight to each slice, which is then used to assign CPU cycles in a weighted fair queue scheduler i.e., slices get CPU in proportion to their weight divided by the total active weights."
    • Done - Move Server to CS machine room
      • We will use complex-1 (2CPU Athalon 1.7Ghz) as everlab. The current everlab will be used as a backup
      • Done - Clean up complex-1
      • Done - Move to Internet and reload as FC4 everlab
      • Done - Debug the Ping of Death
        • The file /usr/local/planetlab/bin/pl-reboot which is called by plc_www/db/control/index.php is missing.
        • Checkout plc_scripts. It needs lots of tweaking for location and for passwords. Must write /usr/local/planetlab/bin/pl_poddoit by hand as a wrapper for rebooot.py.
      • Done - Implement Ganglia on EverLab nodes (in root partition?)
        • Put the ganglia-gmond-3.0.2-1.i386.rpm in /var/www/html/install-rpms/planetlab-v3
        • Run yum-arch in that directory
        • Edit the yumgroups.xml file in that directory and add a line that says
          '\ganglia-gmond\'
          within the PlanetLab group.
        • Create a gmond.conf file and put it /var/www/html/PlanetLabConf. Make each gmond deaf and have it send a udp result to the ganglia server.
        • Add the gmond.conf file to planetlab configuration files via the web interface. Have the update command be '/etc/rc.d/init.d/gmond restart'
        • Install gmond on your server and have it listen to the port listed in the client configuration
        • Check by telenting to the server at port 8649
      • How to remotely power-down/up the blades
      • Test Condor as a workload management system in a PlanetLab Slice
        • PlanetLab Central is working on this for a March deliverable.
        • Install master on www.everlab.org -- DONE
        • Create initscript -- DONE
        • Install el_condor_execute slice on each node. Leave one node on each cluster as a submission node -- DONE
        • Install el_condor, huji_condor, tau_condor, sics_condor and ucl_condor as submission slices on each cluster -- DONE
      • The mailing list options do not work (/planetlab/plc/scripts/*)

      Notes

      To add a new site, use the following command:
      plcmdline.py -u -r admin -p -c 'AdmAddSite("Hebrew University of Jerusalem", "HUJI", "huji")'

      Building a PLC from scratch

      • Install FC4 using the server option. Add the development tools and postgress.
      • Enable ssh, http, and https using the system-config-network script.
      • Update the system using "yum update"
      • Turn off selinux by editing /etc/selinux/config and setting to disabled
      • Follow the instructions from PlanetLab
      • When editing the config file, the package_source variable should be the default planet-lab site. This will be copied to the local host during the rpm_repo step.
      • Run /root/pl_box/update_rpm_repo.sh --createrepo
      • After a failure in bootmanager, comment out the curl line in /planetlab/bootmanager/support-scripts/buildnode.sh and edit the yum.conf file, switching the location boot.planet-lab.org with your new machine name. Restart setup_plc.sh

      OpenAFS Server notes

      • Use the file /usr/afs/local/NetInfo to restrict a multi-homed node to a subset of its available interfaces.

      Building an AFS client on FC4

      I needed to patch rx_vnet.c to include linux/vs_cvirt.h on LINUX26 in order to defeine vx_rmap_pid. I also needed to modify src/afsd/afs.rc.linux to match the locations of the kernel libraries and to correctly identify the processor type.
      • Download an afs build
        curl 'http://dl.openafs.org/dl/openafs/candidate/1.4.1-rc8/openafs-1.4.1-rc8-src.tar.gz' > openafs-1.4.1-rc8-src.tar.gz
      • Use the correct openafs.spec
      • yum install rpm-build kernel-devel kernel-smp-devel pam-devel ncurses-devel byacc flex gcc krb5-devel
      • cp openafs-1.4.1-rc8-src.tar.gz /usr/src/redhat/SOURCES
      • Create the files
        • /usr/src/redhat/SOURCES/openafs-CellServDB
        • /usr/src/redhat/SOURCES/openafs-ThisCell
        • /usr/src/redhat/SOURCES/openafs-CellAlias
        These files will be installed in the source RPMs. A simple default option is to get a CellServDB from grand.central.org and use "openafs.org" for ThisCell.
      • rpmbuild -ba openafs.spec
      • Install the openafs and openafs-client rpms from /usr/src/redhat/RPMS/i386