Re: measured throughput variations

[chacron1 <eric chacron vz cit alcatel fr>]

Hi Stephen,

"Stephen C. Tweedie" wrote:

> Hi,
>
> On Thu, Jun 20, 2002 at 05:35:45PM +0200, chacron1 wrote:
>
> > I redo some test with 2.4.14 ext3 and raw device .
>
> Please try the current ext3, from -ac or ext3 cvs, to make sure you're
> not hitting something that's been fixed since 2.4.14.  2.4.14 is a
> very old, and known buggy kernel.
>

I have tried with 2.4.18-ac and i still get the variations in the measured
performances.

However i have maybe a begin of explanation for you.
I have used in parallel iostat to monitor the disk accesses.
For the same test program repeated several times i get very different
results
due to a very different amount of disk read.

Example: my program does a loop on {read(), write()} using 64 K data
length,
and writeback ext3 file mode opened with O_SYNC.
The first time i run it i get 22 ms mean time for read+write ( 2.9 MB/s
end-user throughput), and observed 6200 block write and 7200 block read
per second
( i guess iostat uses 512 B for block unit).
I redo the same test several times with the same level of result (
removing or
not the file before).

Then, a new session found very different results: 8 MB/s throughput
and iostat showed near 0 block read / s and 17000 block write /s !!

How is it possible that ext3 needs sometime to read as many data as it
writes and sometime it doesn't need to read anything ?

Eric