Linux server hardware

["Johnson, Paul" <Paul.Johnson@marconi.com>]

We're thinking of getting a new Linux server.  This thing is going to have a
couple of hundred gigs of disk space initially, and over its lifetime we
will probably want to push that up to a TB or so.  Its got to serve this at
a fairly high rate to a small number of clients, probably a dozen or so
initially, growing to a few dozen in time (assuming we get the budget).
Most of these clients are just going to be doing ordinary stuff, but a few
will be crunching big data, reading and writing tens of gigabytes as fast as
reasonably possible.  Currently this is mostly sequential access, but its a
fair bet that one of these days we are going to find ourselves with a mongo
SQL DB of some sort.  (This is a research lab, so future requirements are a
bit unpredictable).

Budget is constrained, as you would expect.  Anything over #5,000 ($8,000)
initial cost is pushing it.  This also has to go through a rather
conservative purchasing department so proposals to buy something cheap from
eBay will be blocked.

Given the above, we want a sensible balance of storage, performance, cost
and availability.  So, some questions:

1: The big debate here is IDE RAID vs SCSI RAID.  SCSI RAID is a known
quantity, IDE RAID rather less so but it looks a lot more cost effective.
We know about the Promise cards, but not how they fly in practice.  Does
anyone have any real experience?  I've found some benchmarks on the Web, but
nothing that directly compares a good SCSI RAID with a good IDE RAID.  What
topology of cards and drives will work best?  I know about the slowness of
writing to RAID-5, so we were thinking of RAID 0+1: data striping for
performance plus dual redundancy to allow for failures.  Can IDE and SCSI
both do hot swapping?

2: I've read about the LVM, which is included in SuSE.  We run Red Hat here,
but I really don't like the idea of retrofitting the LVM on to the Red Hat
kernel.  Is it worth the trouble of having a server with a different
distribution?  It certainly looks attractive.  Our current server has One
Big Root Partition, so a rogue process could kill it by filling the disk.
OTOH I know that we won't get the partition sizes right first time.  Or we
could just wait for Red Hat to incorporate the LVM.

3: I've seen Ethernet cards with multiple connectors.  Are these intended
for routers, or is this for higher bandwidth stuff?  Would it make sense to
hook a high bandwidth client directly into one of these ports in order to
leave the others free for everyone else?  Or should we just run 4 lines from
the server to the switch in order to have 400Mbits?

4: How many CPUs should we get on the mobo, and how much RAM?  Most traffic
will be served through NFS, but we also run Samba for Windows people.  The
heavy duty stuff should go through NFS though.

BTW I haven't forgotten backup.  We'll use the SunStor 4000 SCSI tape
hanging off the back of our current server until the backup partitioning
gets too silly.  We use Amanda to do our backups, so breaking up a big
partition across several tapes is fairly straightforwards.

Thanks for any information on any of these questions.

Paul.
_______________________________________________
LinuxManagers mailing list - http://www.linuxmanagers.org
submissions: LinuxManagers@linuxmanagers.org
subscribe/unsubscribe: http://www.linuxmanagers.org/mailman/listinfo/linuxmanagers