[1054] in linux-scsi channel archive
Re: news on dual buslogics with MD
daemon@ATHENA.MIT.EDU (Leonard N. Zubkoff)
Tue Dec 3 01:29:51 1996
Date: Mon, 2 Dec 1996 22:24:45 -0800
From: "Leonard N. Zubkoff" <lnz@dandelion.com>
To: rsr@lab.net
CC: linux-scsi@vger.rutgers.edu
In-reply-to: <Pine.LNX.3.95.961202142846.28694A-100000@Citadel.Lab.NET>
(message from Ryan Smith-Roberts on Mon, 2 Dec 1996 16:09:10 -0800
(PST))
Date: Mon, 2 Dec 1996 16:09:10 -0800 (PST)
From: Ryan Smith-Roberts <rsr@lab.net>
I realize this is somewhat of a FAQ, but there is no linux-scsi FAQ and no
archives, so unfortunately we are forced to ask the same silly questions
over and over.
I have been asked to set up the be-all end-all news server. A rather
difficult request, but the SCSI subsystem I was thinking of putting in
involved a veryfast 1GB Seagate root/swap disk and 3x 4GB Barracuda news
disks, MDed together on two BusLogic 958 controllers, stuffed into a dual
P200 box.
Well, I'll be happy to comment, but do be aware that I've never personally run
a news server.
First, why a dual processor machine? As far as I'm aware, news is generally
I/O bound rather than CPU bound, so I don't see what the extra processor is
buying you. Also, Linux' SMP support is best at compute bound tasks since only
one processor can be in the kernel at a time. Given the high I/O load on a
news server, I'd suggest you get a single P6-200 rather than a dual P5-200.
Make sure you have enough memory that the machine never pages.
I would appreciate any feedback anyone, especially Leonard, can provide
about the possible caveats:
- MD has in the past been unstable with SMP; is anyone running SMP MD
news servers?
I was the one who reported a problem with MD and SMP together which turns out
to be the P6 Local APIC Spurious Interrupt bug instead. Once the workaround
patch was back in place for that bug, there were no further problems on my dual
P6-200. No one else reported any problems, though several people did report
running MD and SMP together without problems if I recall correctly.
- Are 958 controllers bus-mastering? Will they play well together in the
same machine?
They most certainly are bus-mastering. The maximum number of BusLogic
MultiMaster controllers I've run in the same machine at the same time is six.
I've had three PCI and three EISA running together. I've also tested five PCI
controllers all running together. I expect you can run as many BT-958's
together as you can fit in slots.
- Is the FlashPoint DW controller supported?
Yes, but the driver is in beta test. In my opinion, any mission critical
machine like a news server deserves the more proven MultiMaster support at this
time. The FlashPoint support is working fine for many people (and not-so-fine
for a few others), but it's definitely not yet as robust as the MultiMaster
support.
- Will MD work across two controllers / a FlashPoint DW?
MD could care less where the striped partitions are. Someone recently reported
building an MD stripe including two SCSI and one IDE drives. It worked, though
I gather the performance was not quite what one might hope for.
And now some generic SCSI questions...
If one is going to put two or more reasonably fast 4GB drives on one
chain, is F/W SCSI II enough to keep them both hauling data at peak
efficiency, or is Ultra Wide SCSI necessary?
The real question is what does the I/O load look like? From what I've learned
talking to people who do run news servers, the I/O load on a news server is
largely random access, not sequential, and the size of individual transfers is
not that large. For sequential I/O the SCSI bus bandwidth might be an issue,
but for random access there's just no way. Sequential I/O tends to be measured
in MB/second, whereas the interesting measurement for random I/O is
operations/second.
Pursuing this further, would an appropriate architecture be one or two
controllers using Ultra Wide disks, or say three/four controllers using
Fast SCSI II disks, with the cheaper disks offset by the cost of more
controllers and more complex cabling?
If this were a video server, this would be a more interesting/difficult
problem. For news, I really doubt you'll see much difference between Fast SCSI
and Ultra SCSI transfer speeds on the same physical disk. A single BT-958
won't be taxed in random I/O throughput by even four wide disks, though
naturally there will be a slightly higher latency. I'd be real curious to see
actual news server measurements comparing one and two controllers. The biggest
reason I don't tend to run more than four drives per controller is due to SCSI
cable length restrictions, and that my favorite external disk tower holds four
drives.
In short, I don't see any evidence that more than two controllers would be
necessary for this type of I/O load.
Also, in the experience of more-experienced news admins than any of us,
are multiple smaller drives better, WRT seek time, or should one purchase
fewer larger drives? As news is far from mission-critical, excessive
reliability is not a particular concern.
For a random access I/O load, the key performance factor is having more
spindles, so long as you don't compromise the individual speed too much. It's
always a good idea to use identical drives or at least drives of similar
performance in a stripe. What's best to use at any time depends on the current
technology. Right now the next generation of 7200 rpm Ultra SCSI drives are
just barely arriving, but I don't know that I'd trust them yet. Fast 2GB
drives are often only slightly cheaper than the 4GB models, so you may well be
better off staying with the 4GB ones but getting more of them.
Well, those are some slightly rambling thoughts on the subject.
Leonard
