[prev in list] [next in list] [prev in thread] [next in thread] 

List:       linux-megaraid-devel
Subject:    Re: Atrocious RAID performance on PE2400
From:       "Michael K. Johnson" <johnsonm () redhat ! com>
Date:       2001-07-10 16:49:01
[Download RAW message or body]

Matt_Domsch@dell.com writes:
>Indeed.  Slightly slower on block writes, slightly faster on rewrites.

And it's more faster (I've always wanted to say that in a grammatically
correct way, and now I have my chance) on rewrites than it is slower on
block writes:

>  512           2G 12916  99 27012  15  7818   6  8231  69 30846  14 431.3
> 1024           2G 12912  99 31712  17  7757   6  8806  74 30470  14 424.6
> 4096           2G 12917  99 32016  18  7580   6  7888  66 30612  14 418.7

~9% slowdown on block writes, ~11% speedup on block rewrites going from
4096 to 1024.

By comparison, ~18% slowdown on block writes and ~7% slowdown going from
1024 to 512.

I think 1024 is the sweet spot here.  I've heard hints that the midlayer
might be happier with 1024 than 4096 anyway, so it sounds to me like a
good balance.

I also wonder if throughput gains in expense of latency when going
with larger numbers; do you have any tests that test latency?

>All of these tests were done with a 6 disk RAID 10.  I'm curious how
>different the numbers would be with RAID 5 (i.e. how big a chunk does the
>XOR engine work best with?).  A project for later this week I guess.

I'm looking forward to the results of this testing.

>There isn't a way (presently) to have different per-disk max sectors or
>other values.  They're per-controller.  If it turns out that some values
>work best at one RAID level, and other numbers at other RAID levels, it may
>be appropriate to move the decision down to the per-disk level.

I think other kinds of performance tuning are more likely to produce
useful improvements, and this seems like a level of tuning that is
more invasive than is appropriate without involving linux-kernel.
Frankly, having a target-mode scsi emulator in the driver is probably
responsible for more performance degradation than you'll ever fix
by doing per-logical-disk max_sectors tuning.

michaelkjohnson

 "He that composes himself is wiser than he that composes a book."
 Linux Application Development                     -- Ben Franklin
 http://people.redhat.com/johnsonm/lad/


[prev in list] [next in list] [prev in thread] [next in thread]