AMBER Archive (2002)

Subject: Gentoo vs. Redhat Linux, shared/distributed filesystem

From: Jianhui Wu (wujih_at_BRI.NRC.CA)
Date: Thu Aug 08 2002 - 14:44:53 CDT

Dear Amber users,

As I am in the process of 'configuring' a linux pc cluster (Dual Athlon
CPU, Gigabit ethernet). A lot of amber guys here run amber in linux
cluster for years. I would be grateful if some of you can give me some
help in the following issues:

(1) Which Linux OS distribution is more efficient for amber?

Initially, I selected RedHat but the vendor told me that Gentoo linux
would be better as it is optimized for 'your hardware' while Redhat is
optimized for an i386 (except kernel and glibc). Is that true that Gentoo
will give me 'significant performance gain'? I was also told that Gentoo
is not the true system V unix and harder to use. Did you compile and run
Amber in gentoo linux? If so, I would be grateful if you can share your
experience in term of the performance and management.

(2) Shared filesystem vs. distributed filesystem, which one is preferred?

If I get a large central 'disk' (RAID), export it and mounted it one each
node, this shared filesystem is easy to use but perhaps it is not
optimised for the performance as network I/O is intensive.

Altanatively, if I get 40 MB (for example) for each node, then I will have
480MB in total for 12 nodes. How do you 'aggregate' these HD into
'distributed' filesystem? What software do you use for this purpose? Is
openPBS compatible with this kind of filesystem?

(3) Master node do all the I/O

If the cluster is setup in such a way that all jobs are submitted through
the master node, then the network traffic will be very intensive for
master node as the 'master node will do all the I/O' in the case of MPI
version of sander. How do you address this problem? Perhaps I mix up the
concept of the 'master node for sander job' with master node for cluster
here? Perhaps the node I sumbit the sander job will be the master node of
that job?

Is it possible to have a filesystem setup like following:
I submit job1 at node1 using node2-4. Then, all I/O will be handled by
node1 in its local HD. Then, I submit job2 at node5 using node6-8, all I/O
will be handled by node5 via local HD. Then, probably open PBS can not be
used for this kind of management, right?

Thanks a lot for your help,

Jian Hui