Subject: Re: RAID warning - dangerous?
To: NetBSD Users <firstname.lastname@example.org>
From: Chuck Yerkes <email@example.com>
Date: 09/21/2002 08:31:53
Quoting Wojciech Puchar (firstname.lastname@example.org):
> > One of the things I REALLY like about my Baydel RAID units
> > (besides the tremendous performance that happens to not be
> > priced like the Clariions they compete with) is that I can
> > get stats from the box.
> > And I can use that information to adjust stripe size.
> here it's RAID1 case - so there's nothing to tune.
Missed that. Don't find a lot of use for software RAID1,
it means that my CPU has to write everything twice.
Promise IDE cards cover that for me (adaptec in SCSI-land).
> > RAID storing email data performs better with different
> > settings that the box backing a Sybase database.
> using RAID5 for databases in generally very bad idea - unless it's mostly
> readonly database.
I love generalizations like this. Here's more:
"Computers make errors, so you should never use them for critical things."
"Nobody needs more than 640k"
"Nobody would need a computer in their home."
So let me change yours to:
"In my experience, RAID 5 is too slow for database use."
In the past (1995) RAID processors were generally general
purpose CPUs and slow ones to keep the price points down. In
an era of ASICs and CPUs that are excessively faster than disk
transfers, parity (or even irony :) calculations don't delay
writes. Moreover, using a battery backed mirrored bank of RAM
to buffer the data coming in from the bus means that the RAID
controller can take advantage of a little backup and write the
data out in a sorted manner - don't write to sector 5 then
9000 then 20 then 1050, write to 5 20 1050 9000. Keep the
heads from bouncing too much. Obviously theres more complexity
involved what with services reads.
In email, a mail file is deleted moments after it's created.
But it MUST (per 822) be written and fsync'd to permanent
storage. The enormous advantage I find with a write buffer
is that, often, those files never actually hit disk. Oh, but
for those that do and stay, having 10GB or more means I don't
worry about overflowing an SSD. Most journaling file systems,
such as Veritas or Sun's current FFS just journal metadata rather
than actual files. It's a help, but not as good as a good
RAID's write buffer.