Re: SCSI disk: still the way to go?

From: Chris Browne <cbbrowne(at)acm(dot)org>
To: pgsql-general(at)postgresql(dot)org
Subject: Re: SCSI disk: still the way to go?
Date: 2006-05-31 17:50:04
Message-ID: 60ejyat777.fsf@dba2.int.libertyrms.com
Views: Raw Message | Whole Thread | Download mbox | Resend email
Thread:
Lists: pgsql-general

jd(at)commandprompt(dot)com ("Joshua D. Drake") writes:
>> When we were trying to do DW stuff on Linux + Opteron + FibreChannel +
>> EMC DiskArray, we too frequently found filesystems keeling over. It
>> was neither cheap nor reliable.
>
> When is When? Not trying to start a flame war but I am curious as to
> your specifications to make this stuff work. Was it kernel 2.4 or 2.6?
> It 2.6, which? What filesystem are we talking about?
>
> Are we talking the last 12 months? Or earlier then that?

The project ended last fall, roughly speaking.

And that was with kernel 2.6; 2.4 was a complete non-starter as far as
Opteron was concerned.

If memory serves, 2.6.13 was about the best option, but it turned out
to be pretty easy to toast filesystems.

When Josh presented last year at OSCON
<http://conferences.oreillynet.com/cs/os2005/view/e_sess/6574>, he had
a "sidebar" where he discussed the contortions of kernel versioning he
had to go through in order to get PostgreSQL to play reasonably well
with Opteron + Disk Array; it seemed quite similar to our experience,
particularly in that he had to pick very specific kernel versions in
order to get a modicum of stability.

The trouble seems to be that what with the vast amounts of hacking
Gitting into the Linux kernel, somewhere in between [FibreChannel
Drivers | SCSI processing layer | VFS | FileSystems], things aren't
anywhere near completely stable on AMD64.

There's not one place to pin down: it's somewhere in the interfacing
between all of these "layers."

If you take out any of the "exotic" parts, things get better:

- Opteron introduces 64 bittedness, and changes memory addressing over
"plain old Intel."

- "Everyone" runs ATA, so funky FibreChannel is exotic enough that it
doesn't get used enough to get easily debugged.

But for real high performance, you *want* 64 bits, and FibreChannel
interfaces. And Linux just isn't ready for that. Nor is *BSD, I
expect, for that matter, but they're more straightforward about
documenting what *isn't* expected to work.
--
(format nil "~S(at)~S" "cbbrowne" "acm.org")
http://cbbrowne.com/info/wp.html
If anyone ever markets a really well-documented Unix that doesn't
require babysitting by a phalanx of provincial Unix clones, there'll
be a lot of unemployable, twinky-braindamaged misfits out deservedly
pounding the pavements.

In response to

Browse pgsql-general by date

  From Date Subject
Next Message Ramakrishna Reddy 2006-05-31 17:57:47 Re: A quick question on CONTRIB package
Previous Message Scott Marlowe 2006-05-31 17:35:21 Re: A quick question on CONTRIB package