Scyld ClusterWare HPC: Installation Guide
<< PreviousNext >>

Changes to Configuration Files

Changes to Red Hat Configuration Files

An installation of Red Hat sets a default configuration optimized for a stand-alone server. Installing ClusterWare on a Red Hat installation changes some of these default configuration parameters to better support a cluster. The following sections describe the changes the ClusterWare installation automatically makes to the Red Hat configuration. Any of these may be reversed; however, reversing them may adversely affect the operation of the ClusterWare cluster.

  1. /etc/grub.conf has been modified.

    After ClusterWare has been installed, the default boot becomes the newest ClusterWare smp kernel.

  2. NFS Services default configuration has been modified.

    By default, Red Hat configures NFS to "off" for security reasons. However, most cluster applications require that at least the home directory of the master node be accessible to the compute nodes. NFS services on the master are set with the default to "on" for run levels 3, 4, and 5.

    The default out-of-box chkconfig for NFS on RHEL5 is as follows: 

    [root@scyld ~]# chkconfig --list nfs
nfs             0:off   1:off   2:off   3:off   4:off   5:off   6:off
    ClusterWare has changed the default to the following: 
    [root@scyld ~]# chkconfig --list nfs
nfs             0:off   1:off   2:off   3:on    4:on    5:on    6:off
    To get NFS to mount directories from the master to the compute nodes, the file /etc/exports needs one entry per line for each file system to export from the master to the compute nodes (the RHEL5 default is a blank/non-existent file). ClusterWare creates this file if it didn't already exist, and adds several new entries of the form:

    ExportedDirectoryPathname @cluster(accessMode,syncMode,no_root_squash)

    The export for /home from the master is configured with an accessMode of rw (read-write) and a syncMode of sync by default for data reliability reasons, and the non-/home directories are exported ro (read-only) for security reasons and async for performance reasons.

    See the ClusterWare Release Notes for details about which directories are added by Scyld.

  3. /etc/sysconfig/syslog has been modified.

    Compute nodes will forward messages to the master node's syslogd daemon, which places them in /var/log/messages. In order for this to function correctly, ClusterWare modifies the /etc/sysconfig/syslog file by adding the -r option to the SYSLOGD_OPTIONS line:

    SYSLOGD_OPTIONS="-m 0 -r"

<< PreviousHomeNext >>
Mapping Compute Node Partitions Possible Changes to ClusterWare Configuration Files