From: | Mike Mascari <mascarm(at)mascari(dot)com> |
---|---|
To: | Jeff Hoffmann <jeff(at)propertykey(dot)com> |
Cc: | "Ross J(dot) Reedstrom" <reedstrm(at)wallace(dot)ece(dot)rice(dot)edu>, pgsql-general(at)postgreSQL(dot)org |
Subject: | Re: [GENERAL] server hardware recommendations (the archives aredead) |
Date: | 1999-12-15 20:49:50 |
Message-ID: | 3857FEED.ED9EFCFC@mascari.com |
Views: | Raw Message | Whole Thread | Download mbox | Resend email |
Thread: | |
Lists: | pgsql-general |
Jeff Hoffmann wrote:
> "Ross J. Reedstrom" wrote:
> >
> > On Wed, Dec 15, 1999 at 11:27:36AM -0400, The Hermit Hacker wrote:
> > > On Wed, 15 Dec 1999, Jeff Hoffmann wrote:
> > >
> > > Most of my RAID tests are on Solaris+Disksuite...with good drives
> > > in the machine, my writes are something like 18MB/s to the drive, stripe'd
> > > and mirrored...I think reads worked out to be 19MB/s...(bad drives, same
> >
> > Ah, this would be a RAID 0+1 setup, then? Very different from Jeff's RAID
> > 5 configuration. I'd be willing to believe that software RAID 0+1 _could_
> > be faster than most hardware (it's just shuffling and dupping blocks
> > around to different drives, which could be done with clever pointer
> > twiddling) but calculating parity bits in hardware for RAID 5 had got
> > to be a win, doesn't it?
> >
>
> i would assume that this would be the case. for anybody who is going to
> spec a new machine for a database as small as 3-5G, RAID 0+1 has got to
> be the choice. i don't have a doubt that it'd be reasonably fast with
> software raid. anymore, it'd be hard to buy new disks that small to
> build a 0+1 for < 8G (4x4G will give you 8G). when you're on a budget
> with a backup server that needs 20+ drives (n+1 for raid5) vs. 40+
> drives (2N for 0+1), though, raid 5 is a good solution. doing it again,
> i'd go with a hardware controller since no one seems to be refuting my
> assumption that the raid5 daemon can suck up a lot of CPU when
> calculating parity, even with 2 fairly fast processors.
Of course, that's the real trick, isn't it? Hard drives are becoming so large, so
fast, it's difficult to determine the proper RAID solution with the supplied
budget. We wanted speed, not volume. So we wanted to build a software RAID 0+1
configuration as cheaply as possible with the fastest disks/controllers. We went
with a multi-channel Ultra-2 Fast Wide Differential contoller (80MB/s) and 80MB/s
LVD Cheetah drivers - the problem in building RAID 0+1 is that none of the drives
come smaller these days then 9G ($450US), so a minimal RAID 0+1 configuration
would be 4 drivers = 18G (2 for the stripe, 2 mirroring the stripe). That seems
like major overkill for a database that NEEDS SPEED, but may only grow to a couple
gig in size...It's too bad we couldn't buy 8 or 16 4G/2G 80MB/s at the
proportional prices.
Also, for what its worth, we've been running PostgreSQL on a dual 450Mhz SMP
running just RAID 1 for about a year now without problems under RedHat 5.2
(2.0.36), although in those earlier kernel versions you have to rebuild the kernel
with _SMP_ defined. It's pretty quick though...
Mike Mascari
From | Date | Subject | |
---|---|---|---|
Next Message | Mike Mascari | 1999-12-15 20:56:08 | Re: [GENERAL] server hardware recommendations (the archives aredead) |
Previous Message | Jeff Hoffmann | 1999-12-15 20:16:42 | Re: [GENERAL] server hardware recommendations (the archives aredead) |