ADSM-L

Re: Dsmfmt on cx600 any thoughts?

2004-02-19 15:11:00
Subject: Re: Dsmfmt on cx600 any thoughts?
From: Richard Sims <rbs AT BU DOT EDU>
To: ADSM-L AT VM.MARIST DOT EDU
Date: Thu, 19 Feb 2004 15:10:32 -0500
>We have serious performance issue for dsmfmt command.
>
>We are getting 45MB/sec write speed on AIX4.3.3 box connected to CX600 ATA
>RAID5 4+1 disk group.But when we use dsmfmt it gives 9 to 16 MB/sec. Which
>is very sluggish to format 16 TB storage for storage pools.
>
>Does anybody have information on how dsmfmt works?
>
> or how to speed up so that it will give better I/O on SAN storage?

Realize a few things...

The 45MB/sec figure doesn't mean much without a lot of descriptive context.
For example, a timing test using ordinary, buffered I/O is rather meaningless
in that what's involved is principally the population of memory buffers
rather than doing actual I/O.  I have never seen any technical documentation
on the workings of the dsmfmt command, but it may be doing synchronous I/O,
which is necessarily slow.  Things are slower still if the command is going
back to check each block after writing for fidelity.  File system block size,
and file system fragmentation (in odd cases where the TSM "volume" is sharing
space in a file system) are other factors.  And then there is the RAID method:
as has been discussed on the List, RAID 5 is particularly poor at
sequential writing.  (See
http://www.pcguide.com/ref/hdd/perf/raid/levels/singleLevel5-c.html et al)

Dsmfmt is the kind of utility not intended to be a high-performance thing,
given that it's essentially intended to be used only for initialization, which
is rather rare.  You've received recommendations to use raw logical volumes
because it eliminates the need for formatting: that approach is of particular
value for disaster recovery.

   Richard Sims,  http://people.bu.edu/rbs