math at velocet.ca
Thu Feb 28 13:45:36 PST 2002
On Thu, Feb 28, 2002 at 04:07:25PM -0500, Robert G. Brown's all...
> Dear Liststers,
> I'd like to request comments on a couple of dual Athlon motherboards.
> We are considering both the Tyan Tiger 2466N (760 MPX) and the MSI K7D
> Master (MS-6501) (also 760 MPX). Our local vendor "supports" MSI
> motherboards (which just means that we deal with them rather than Tyan
> in the event of a return, but which makes it reasonable to use the MSI
> all things being equal). We are going with 760 MPX to get the 64/66 PCI
> slots, of course -- we actually have a small stack of 2460 Tigers which
> are not totally painless but which we've more or less tamed.
> Any experiences yet, good or bad, with either motherboard? The vendor
> is probably going to loan us an MSI-based dual to test, but there's
> nothing like the experience of somebody actually running a cluster if
> there is anybody out there already doing so.
We have both. Minimal experience yet in comparison, only had the 2466 around
for a week or so. Some cards that dont work in 2460 work in the 2466 (such
as the addtron GBE cards - just a POS card we tested for the hell of it).
66Mhz bus is sweet and works well. The fact that the thing has an onboard
vortex and takes regular DDR ram (ecc in our case) now that ram is pricey
again is great. Saved us a bunch per board (~$100). The NIC might not
be as great as an intel EEpro but it works great for us (much better
than a lowly rtl 8139). On board PXE is da bomb, and of course serial
console exists on both boards is nice for saving on videocards (and power
since AGP just burns hot).
Im extremely happy with both so far (tho the 66Mhz is really the best part of
the 2466), but i havent used both for a month intensively yet.
> I'd also like comments on RAID alternatives. We have a group who needs
> about 500 GB of RAID. We just got a Promise UltraTrak100 TX8 (IDE-SCSI)
> RAID chassis that advertised decent itself as OS-independent plug and
> play -- attach to SCSI bus and go. The first unit we were shipped
> didn't work under any OS. The second we were shipped we got the vendor
> (Megahaus) to verify function before shipping and it does "work", but it
> returns unbelieveably poor performance at RAID 5 -- a (very) few MB/sec
> -- under bonnie. From this we learned (among many things:-) that
> vendors often quote performance numbers on a RAID from its RAID 0
> configuration, which would kind of funny if it weren't for the murderous
> impulses it creates when you learn that their numbers are some sort of
> cruel joke under RAID 5.
Raid 5 is costly. As I understand it (not very well) there's 3 writes for
RAID 5. Reading can be slow, but not nearly as bad. Obviously you dont want
to scratch to raid 5 ;)
We've used software raid for as long as its been out for freebsd (longer than
linux). In the last years and my partner Dave Gilbert has submitted patches to
freebsd for it. We've had great performance out of it, and only headaches with
hardware raid (AMI MegaRAID) , as the list was recently talking about.
> a) Another IDE-RAID enclosure, perhaps from a better manufacturer.
> However, at this point we're more than a bit concerned about the gap
> between vendor performance claims and reality. There are vendors that
> assert 100 MB/sec read times, but we are concerned that they mean "at
> RAID 0" which is useless to us. We need real-world loaded numbers at
> RAID 5 (e.g. multiple instances of bonnie). Folks we know locally who
> have e.g. zero-d chassis report real world throughput more like 20
> MB/sec RW, but their boxes are a year or two old and may not reflect
> current rates. 20 MB/sec is pretty much the LOWEST rate we could
> tolerate in this application under multithreaded load, and we'd like
> something better. Any enclosure/controllers out there that give good-to
> excellent performance that you'd care to recommend?
can you avoid raid 5 writing for any part of the job? You can also do
raid 10 (1+0) which is often faster (depending on the configuration).
With software raid you get to control what you're doing, hardware raid
vendors may not give these options.
> b) md-raid, either ide or scsi, on a straight linux server. We know
> that this works remarkably well. We run md raid in the departmental
> server (scsi, with a stack of 36 GB disks in RAID 5) and get excellent
> performance -- ~40 MB/sec write throughput and even better for read.
> Unfortunately large SCSI disks are still excessively expensive and we
> don't have the budget to reach 500 GB with SCSI disks for this cluster.
> IDE is cheap and easy, but we would like a bit of assurance that linux
> won't have (e.g. DMA) problems when dealing with 6-8 ide controllers on
> one bus. Is anyone doing this? Good, bad experiences, hardware
> recommendations or gotchas all welcome.
Thats why we end up going with scsi. The drive limit on ATA is a bit
> c) SCSI RAID. Definitely works, definitely high performance, but also
> the most expensive and again, we won't be able to afford to reach our
> design spec with the money allocated to this ($5-6K total).
> If we have to fall back to SCSI we will and will live with a smaller
> RAID than we had hoped, but we'd very much like to first find out if
> IDE-based RAID solutions (RAID 5 on ~500GB total disk) with >20 MB sec
> worst case write rates under heavy load exist.
Raid 10 an option? Eats more disk but is faster usually.
> Robert G. Brown http://www.phy.duke.edu/~rgb/
> Duke University Dept. of Physics, Box 90305
> Durham, N.C. 27708-0305
> Phone: 1-919-660-2567 Fax: 919-660-2525 email:rgb at phy.duke.edu
> Beowulf mailing list, Beowulf at beowulf.org
> To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
Ken Chase, math at velocet.ca * Velocet Communications Inc. * Toronto, CANADA
More information about the Beowulf