Strping or fragmenting
Jon
I'm about to migrate a 30-ish Gb system onto a new machine. Currently, the
database is using Informix mirroring, which I'm perfectly happy with. What
I'm not happy with is the layout. The tables are in various dbspaces, most
of which have had many chunks added over the years. My plan is to fragment
the larger tables by round robin across several disks and controllers. The
real problem lies with the indexes. These are currently all in one dbspace
which is several Gb and made up of loads of chunks. Now some of the tables
have over 20 million rows so this is clearly not good.
So - what would be the best strategy for the indexes? I don't think any of
the larger ones are good candidates for expression based fragmentation, so
I'm tempted to go for disk striping. I'm concerned about what would happen
if a disk died, though. At the moment, this is pretty easy to recover from
in that I just replace the disk, then use onspaces to recover the affected
chunks. How would this be done if I was using striping?
The system is AIX with SSA disks.
All suggestions greatly appreciated!
Thanks,
Jon.
Art S. Kagel
Don't use raw stripes there is no recovery. The best scheme is to use
RAID 1+0 or striped mirrored pairs. You would create a number of
mirrored pairs using hardware or OS mirroring then create a large
stripe set from 10 (or more) mirrored pairs. This will give you one
HUGE logical drive with mirror safety and recovery at the single drive
level. Then just subdivide the large logical drive into 2GB or smaller
virtual disks and these are your Informix chunks. (BTW make the stripe
block size 16K (the normal stripe size is 64K or larger which works
well for filesystems but not for databases).
The alternative, make two large stripe sets and subdivide each then
with either OS or Informix mirroring mirror parts of one stripe to the
other. This is bad news if one drive on one stripe set goes one whole
side of every mirror will be down and the entire drive array will be
busy for hours rebuilding when the offending drive is replaced.
This second scheme is known as RAID 0+1 which performs about the same
as RAID 1+0 (also called RAID 10) described above but RAID 1+0 recover
only involves the single drive and mirrored pair that has a down member
in recovery all other mirrored pairs behave and perform normally. Also
RAID 0+1 once a single drive is down any drive failure on the other
side of the mirror will destroy all data. With RAID 1+0 you can lose
up to 1/2 of your drives with no dataloss so long as both sides of the
same mirrored pair are not lost before recovery and since recovery is
faster in RAID 1+0 the risk is proportionately smaller of even that
happening.
(You can improve the odds against data loss even further by making sure
no two drives of the same mirrored pair are from the same manufacturing
lot, in case a drive lot is bad. Happened to us! Lost 100 drives from
the same IBM drive lot on 20 machines over the course of a single week
because of a bad lot so don't poo poo this one! We had 300 drives from
that lot and 15 failed before we could determine the cause. We lost
the others while waiting for replacements to be rebuild by the RAID
firmware and that was RAID 5, against my wishes I'll add, which put us
at extreme risk!)
Art S. Kagel
bmc...@my-dejanews.com
> Don't use raw stripes there is no recovery. The best scheme is to use
> RAID 1+0 or striped mirrored pairs.
Will RAID 10 perform as well as many 2 gig mirrored pairs with an intelligent
fragmentation strategy? How does RAID 10 effect parallel processing? Or
would you use a few RAID 10s?
Brian McGinity
-----------== Posted via Deja News, The Discussion Network ==----------
http://www.dejanews.com/ Search, Read, Discuss, or Start Your Own