cluster packages/distributions roundup
shewa at inel.gov
Fri Mar 15 14:32:54 PST 2002
I have decided to review some of the Linux cluster/ssi software that has
been produced, and I
would like feedback on my initial list of candidates as well as the
judging criteria. In the case
of commercial distributions, I plan on evaluating the basic or trial
versions. Also, I welcome
anybody who wants to help me evaluate any of these.
Background info: I am a computer scientist (BA) at the INEEL. I have
built and administered a
couple of medium size clusters between school and work, using both Scyld
and Mandrake Linux.
Suggested candidate list is (in no particular order):
Although the current version of ROCK includes the MPP projects's
software, I will
probably wait until version 1.7 is released in a couple months.
A significant collaboration including SGI, Intel, Dell, MSC.Linux, NCSA,
IBM, and others.
Makes use of the System Installation Suite (http://www.sisuite.org),
which grew out of
VA's SystemImager and IBM's LUI projects.
ClumpOS - a mini Linux/Mosix distribution
It isn't a full distribution since it is dependent on a head node
having Mosix (soon
OpenMosix) already installed, but it looks like a great solution for
A Bproc based distribution produced by Los Alamos National Laboratory.
is doing some very interesting work in many areas including LinuxBIOS, a
language with checkpointing built into the compiler, and fault tolerant MPI.
Warewulf is brand new, and it is mostly a slave node builder right now.
NPACI Rocks Cluster Software
Scalable Cluster Environment
Developed by Kasetsart University in Thailand. A very complete looking
The PC Cluster Consortium develops SCore and the Omni OpenMP compiler.
They say that
their OpenMP compiler can actually compile programs to run on a cluster
with a software distributed shared memory system called SCASH.
Single System Image Cluster for Linux
This project was started by Compaq and it inherited some of their
Unixware code, but it also borrows
from Mosix. It isn't ready yet, but it looks like they have virtual
clusters up and running with
User Mode Linux.
Turbolinux with Enfuzion
Suggested evaluation criteria:
-interconnects (I can only test fast ethernet currently)
-scalapack, blas, etc. (atlas)
-batch system/grid access
+license/cost of code
+type of SSI
I will also be taking into account the guidelines created by the NACSE** .
Thank you for any feedback/help.
More information about the Beowulf