RE: HP vs Informix mirroring
-----Original Message-----
From: Mark Collins [mailto:mcollins@us.dhl.com]
Sent: Thursday, March 19, 1998 3:43 PM
To: Crisp, Kevin; informix-list@iiug.org
Subject: RE: HP vs Informix mirroring
> Funny, but we had exactly the same set up: RAID-5, HP/UX 10.20 and
Informix
> 7.23UC1, but performance decreased when we went to a (Informix) software
> mirroring scheme (Raid 1/0? - I can never remember). I'm fairly convinced
> this was because it has a capacious and expensive 3rd party disk
controller
> and cache. My gut feeling would be to avoid RAID - 5 like the plague -
> mainly due to the inability to use table fragmentation.
RAID 5 does incur overhead, no argument there. Regardless, it does make
sense for SOME
sites. The overhead applies primarily to disk writes (inserts and updates),
so for data
warehouses or other systems with primarily read access it may be
appropriate. Agreed, RAID
0/1 will run faster, but the additional hardware costs for full mirroring
may outweigh the
improved write performance, for some people.
As to the fragmentation issue, with a truly large database, you probably
have multiple RAID
arrays, and you can fragment across multiple arrays, rather than multiple
disks.
I think this is a bit of a generalisation. about 200Gb can be fitted
on one RAID array, okay a 'truly large database' is probably terabytes, but
more than 50Gb still impresses me I'm afraid ;-)
I have made a small promise to myself to never willingly use RAID-5
or ANSI mode again.
> One interesting thing that came up on this particular machine is that in
> 7.23+, on an HP/UX 10.20 box, with an ANSI database, performance dies, by
a
> factor of many THOUSANDS, when you have more than about 4 million records
in
> the database. Apparently, Informix loses all grip of query paths, and
starts
> doing sequential scans all over the place.
Bug Ref. 65369 & 67022
We did run the update stats, using the migration manual stuff - no
effect.
I assume you ran SET EXPLAINs before and after, and that is how you know
it's now doing
sequential scans. I know it's a basic question, but have you updated
statistics according
to the release notes? Is there any way you can create a test database of
similar size but
with non-ANSI logging, to see if the performance drop is limited to ANSI
databases?
Since it is being fixed in 7.24.UC5, we haven't worried too much
about it.
> Also, with post 7.23x ANSI databases, dbschema crashes the engine when
run.
Only ANSI databases? What did Informix support say?
They said it's a bug. I don't have the number to hand.
> AND (will this list never end?)there is a bug in HP/UX 10.20 that crashes
> (i.e. reboots) the whole box after a few minutes of usage in 7.23.uc4+.
Is this a known bug, with an Informix bug number? I'm running 7.23.UC6 on
HP-UX 10.20. How
many minutes is "a few minutes of usage"? Mine has been staying up all day
for weeks. We
do bring it down each morning, but I also have several test databases with
similar setups
that have been running for 3 weeks. "onstat -" gives "INFORMIX-OnLine
Version 7.23.UC6 --
On-Line -- Up 21 days 00:52:55 -- 11352 Kbytes". I'm not doubting that you
have experienced
the problems you mention, I'm just wanting to see if there's an Informix bug
that may affect
me. One possible difference is logging mode - you mention ANSI logging
several times
earlier, and I don't use that mode. Were the databases on the rebooting
machines also ANSI?
I'm afraids I don't have the original tech support fax, that
Informix sent
but I can confirm that it is a bug with HP/UX rather than Informix,
and
we reproduced it several times on two different boxes (D220s). There
are
various possibilities as to why you are not suffering it namely:
1. It only effects boxes with ANSI databases on them
2. You have the appropriate patch from HP (do you apply
their patch
bundles regularly
3. It only effects D200s, and you ain't got one
> Ahhhh Informix.
>
> Kevin Crisp
My apologies for the struture of this message, haven't got to the bottom of
my new Email editor yet!
Kevin Crisp