Re: RAID Newbie question...

"Joe Maloney" <jrpm_at_my-deja.com> wrote in message news:93ih5k$esg$1_at_nnrp1.deja.com...
> TJI, late in the discussion.
>
> Whether any raid 0, 1+0, 5, 7 or whatever will help you depends on
 the
> size of your database, raid vendor and transaction pattern.

Perhaps there's too much "marketing" about RAID. Don't think that RAID is somekind of universal performance wonder (and as mentioned below, at least originally it was intended for protection against a single-disk loss). Look first at your application, carefully identifying the problems to be solved (know the problem before you try to solve it).

>
> Basically, RAIDS are for security, not performance. Remember, they
 are
> Redundant Arrays of Inexpensive Drives, designed to handle drive
> failure.

> I know a lot of people talk about striping the data in a RAID5, but
 I
> believe they are confused. If you are retrieving a large row, it
 makes
> little (some, but little) difference if that row is in 4 pieces plus
> parity (original RAID5) or 1 piece (RAID 0 or RAID1). Breaking a
 table
> up so that it has extents spread accross several drives, striping
 and
> balancing the data that way at the logical level, can have
 significant
> effects.

Care to explain further?

That RAID 5 (4+1) can achieve as much as 4 times the IOPS of a single disk, I can't imagine how that could be considered a small difference. To achieve the same thing by manually "striping" across drives must be a admin. nightmare.

>
> My databases range from 100GB to 1.5TB. The discussion and
> consideration of RAIDx works very well on the smaller databases,
 which
> have local drive arrays and disk farms. The larger databases (and
 the
> newer small ones) are using EMC and related massive storage media.
 EMC
> uses RAID internally, but the server has 1 or two links to the EMC,
> admittedly fiberoptic or SCON, but only 1 or two links. The number
 of
> drives, logical or physical, is irrelevant because it is all going
> through a minimum number of i/o channels.

Why should host-based or remote storage matter? More disk is a way to increase IOPS. Those fiber channels can have sets of controllers performing in the 10k range, for that you would need a sh*tload of disks.

Regards
/Fad Received on Wed Jan 10 2001 - 20:46:05 CST