Hi folks,
Chris Mason wrote:
>
> On Friday, November 09, 2001 03:34:31 PM -0500 "Philip R. Auld"
> <[EMAIL PROTECTED]> wrote:
>
> > Looks like the highmem change does buy a little bit, but not enough.
>
> Ok, it probably doesn't make much sense to go further without a better idea
> of what the benchmark is doing, especially if it is o_sync/fsync heavy.
>
> -chris
I found some time to poke into aim. The workload I've been using
simulates a number of simultaneous users on a shared system.
There are 3 jobs it does where it works on various files with O_SYNC.
I wouldn't say in anyway that is was o_sync heavy because it's a small number
of the total jobs in the work load. However, without that things were much
better.
First, here is ext3 on the initial workload:
Tasks jobs/min jti jobs/min/task real cpu
500 142.34 92 0.2847 21286.96 7244.69
(and reiserfs and ext2 on the same test as a refresher)
Tasks jobs/min jti jobs/min/task real cpu
500 74.33 93 0.1487 39150.29 9955.42
Tasks jobs/min jti jobs/min/task real cpu
500 1170.59 90 2.3412 2485.93 2787.82
Here are reiserfs and ext2 without the O_SYNC part of the workload.
reiserfs:
Tasks jobs/min jti jobs/min/task real cpu
500 1336.69 99 2.6734 2109.69 4195.49
Tasks jobs/min jti jobs/min/task real cpu
500 1615.73 98 3.2315 1745.34 3469.24
We see much better performance without O_SYNC jobs. With it
the journaling really hurts. So that seems to be the issue
here.
Thanks for the input.
Have a good weekend,
Phil
------------------------------------------------------
Philip R. Auld, Ph.D. Technical Staff
Egenera Corp. [EMAIL PROTECTED]
165 Forest St, Marlboro, MA 01752 (508)786-9444
