SUMMARY: Disc sync speed up ?

From: Peter Bartal <bartik_at_gmx.net>
Date: Wed, 22 Jul 1998 15:31:57 +0100

Sorry for the late summary. Thanks for the answers from

Kym Schwarz
Ed Murphy
Mullin Stephen
Alan Rollow
Tom Webster
Allen Mark

    I'd like to add some more info about the background of what happened.
The SAP people did an client export and some of the tables are over 1 GB, so
my swap became full. At the morning I got to work I coudn't log in. From
anywhere. Even the console was blocked. That was the reason for the
unplanned shutdown and it led to the corruption. We do use LSM and do have
UPS.

Here are the two answers which I think cover the problem best:

Digital UNIX's LSM has a facility called "block change logging" (which
requires a license PAK, I believe.) BCL's purpose is to dramatically
speed disk syncs after a service interruption. With BCL, the disk
management system knows precisely which blocks on a logical volume
changed and so it only copies those blocks during a sync operation.
(This is much faster than the default sync behavior which is to copy the
entire volume from a mirror.)

The downside to BCL is that the log buffer is extremely small, and the
LSM subsystem will delay subsequent I/O requests if the buffer is full
and some pending I/O operation is in progress. This obviously would
significantly impact any kind of OLTP type system, as well as any other
disk subsystem under a load.

The reports I've heard are that BCL will be significantly improved in
Digital UNIX 5. Unfortunately, for now, though, I think the answer to
your question is "No."

Regards,

Mark

--
Mark Allen -- 415-995-3726 -- Mark.Allen_at_pbdir.com
Senior UNIX System Administrator -- Pager/PCS 510-599-7893
Pacific Bell Directory -- San Francisco, California
I'm not sure what you mean by syncing.  If you are refering to an fsck
being run on UFS volumes that were not shutdown cleanly, the only way to
avoid fscks is to switch to a journaled filesystem like AdvFS.
If you are talking about the automatic resync that LSM mirrors do on a
restart, which can take a quite a while if it was an unplanned shutdown
and the plexes do differ; use hardware mirroring rather than software.
I know it is an extra cost, and you can't do things like mirror across
controllers, but it seems to work better in general.  When we restart
our systems with LSM mirrored boot disks we have between 15-45min of
high CPU usage while it tries to make sure the mirrors are in sync.  I'd
generally prefer that this is left to a (semi?)intelligent RAID controller
whenever possible.
Then again, a UPS is a pretty inexpensive investment, esp. on a database
server.  APC has a version of their unattended shutdown software that is
supposed to work under DU, and Best's software (included in the price of
the UPS) is also supposed to work under DU.  Please think about it, between
(1) brown-out and black-out protection, (2) surge supression (better than
those cheap power strips), and depending on the model (3) clean filtered
power (which may increase the lifespan of your equipment) it isn't a bad
investment.
Tom
--
+-----------------------------------+---------------------------------+
| Tom Webster                       |  "Funny, I've never seen it     |
| SysAdmin MDA-SSD ISS-IS-HB-S&O    |   do THAT before...."           |
| webster_at_ssdpdc.lgb.cal.boeing.com |   - Any user support person     |
+-----------------------------------+---------------------------------+
|      Unless clearly stated otherwise, all opinions are my own.      |
+---------------------------------------------------------------------+

Received on Wed Jul 22 1998 - 13:36:13 NZST