Re: questions about disk configurations

On Mon, 9 Dec 2002, Josh Berkus wrote:

>
> Scott,
>
> > Actually, first I'd try one big RAID 5 and see how it runs. THEN I'd
> > spend time mucking around with different configs if that wasn't fast
> > enough. If you need x performance and get 10x with a RAID 5 then move on
> > to more interesting problems.
>
> Depends on how much time you have to spend re-installing. IMHO, RAID 5 is
> slower that straight disks for Postgres, especially with large numbers of
> writes. This may not be true for $1000 RAID controllers, but I have yet to
> use one.

Even the fastest RAID 5 boxes aren't superfast, but a RAID5 of 15k drives
with a lot of drive in it does OK, since it can 1: spread small writes
around on many different drives (i.e. if you have 12 drives, and a lot of
small writes, a lot of them will be on different drives.) as well as
spreading out random reads, while providing good large reads, i.e.
sequential scans.

The key to good RAID 5 is to throw as many drives as you possibly can at a
problem, preferably across several SCSI interfaces. Or FC-AL.

> I have a box with a low-end RAID 5 controller, and it drives like a single IDE
> drive on large UPDATE queries. Slower, somethimes.

Many low end RAID 5 controllers are pretty slow. The adaptec AIC133
series (I think that's the right number) are total dogs. The older AMI
Mega raids were fast for their day, but any decent 350 MHz machine with a
dual channed SymBIOS card will outrun it at RAID 5.

> > Not in my experience. I'd estimate my test box with dual 18 Gig UW scsis
> > runs about 1.5 to 1.8 times faster with the two drives in a RAID1 as if
> > a single one is used. Bonnie confirms this. single drive can read about
> > 25 Megs a second, a pair in a RAID1 reads at about 48 Megs a second.
>
> This is Linux software RAID?

Yep. The kernel level drivers are quite fast in my experience, but they
don't seem to give any improvement when layered (i.e. 1+0 or 0+1) over
whatever is the slowest of the two layers. I.e. setting up a RAID5 of
RAID0s results in almost the exact same performance as if you'd just setup
the same number of drives under RAID 5 as you had mirror sets in RAID0.
Since this is the case, you get better performance just going to RAID 5
with twice the disks and twice (-1n) the space.