Installation guide

Release Notes

7. Fixes a bug where cluster startup would leave temporary ﬁles in /tmp/. These are now properly deleted.

8. The beoserv tftp server, which executes on the master node, now only listens on the private cluster interface. Previously

it listened on all interfaces for tftp requests. Additionally, previously tftp requests only retrieved ﬁles that resided in the

/var/beowulf/boot/ directory. Now it treats a requested ﬁlename as being a pathname relative to that base directory,

i.e., the ﬁle may reside in a subdirectory of /var/beowulf/boot/.

9. The beoserv daemon now automatically removes duplicate MAC addresses from ﬁle

/etc/beowulf/unknown_addresses.

10. The master node’s /etc/ofed/dat.conf is now copied to each node as /etc/dat.conf where various MPI imple-

mentations (e.g., HP-MPI included with Fluent 12 and some versions of Intel MPI) expect to ﬁnd it.

11. The /etc/beowulf/config prestage directive now supports prestaging any master node ﬁle to compute nodes at

cluster startup. Previously, prestage was limited to ﬁles that reside in one of the libraries directories.

12. Introduces cleaner support for the Inﬁniband RDMA Protocol (SRP) with a new startup script,

/etc/beowulf/init.d/20srp. To use SRP, you must install the optional srptools rpm from the base distribution,

enable the 20srp script (e.g., using /sbin/beochkconﬁg), and reboot the cluster nodes.

13. OpenMPI is upgraded to version 1.4.1.

14. TORQUE is upgraded to version 2.3.10.

15. When starting Beowulf services (/etc/init.d/beowulf), Scyld ClusterWare now automatically increases some sys-

tem resource parameters to better handle the demands of small- to medium-sized clusters:

• Increase the number of available pids to a minimum of 98,304. See the Section called Optionally increase the max

number of processID values for more information.

• Increase the number of nfsd threads to at least one thread per compute node, with a lowerbound of eight (the Red

Hat default) and an upperbound of 64. See the Section called Optionally increase the number of nfsd threads for more

information.

• Increase the ARP cache capacity from the default threshold values of 128, 512, and 1024 to new values of 512, 2048,

and 4096, respectively, and increase the gc_interval from 30 seconds to 240 seconds. See man 7 arp for more details.

New in Scyld ClusterWare 4.8.0

1. The initial CW4.8.0 release included a kernel that was based upon RHEL4 2.6.9-89.0.15. The current

CW4.8.0 yum repository contains a newer kernel that is based upon RHEL4 2.6.9-89.0.18. See

https://rhn.redhat.com/errata/RHSA-2009-1024.html, https://rhn.redhat.com/errata/RHSA-2009-1132.html,

https://rhn.redhat.com/errata/RHSA-2009-1211.html, https://rhn.redhat.com/errata/RHSA-2009-1223.html,

https://rhn.redhat.com/errata/RHSA-2009-1438.html, and https://rhn.redhat.com/errata/RHSA-2009-1522.html for

details about Red Hat kernel changes between CW4.3.1 and 2.6.9-89.0.15. See https://rhn.redhat.com/errata/RHSA-

2009-1541.html and https://rhn.redhat.com/errata/RHSA-2009-1671.html for details about subsequent changes through

2.6.9-89.0.18.

2. The initial CW4.8.0 release included the Scyld Integrated Management Framework (IMF) with some enhancements

that were only available as separately licensed modules, versus the unrestricted full Scyld IMF that was bundled into

CW4.3.1 and called ClusterAdmin. The latest CW4.8.0 yum repository once again contains the fully functional Scyld

IMF and is distributed under the Scyld ClusterWare license.