From: Judith Lebzelter <judith@osdl.org>
To: Suparna Bhattacharya <suparna@in.ibm.com>
Cc: linux-aio@kvack.org, linux-osdl@osdl.org
Subject: [linux-osdl] Re: OSDL aio-stress result on buffered reads for recent AIO patch set
Date: Wed, 04 Aug 2004 11:54:10 -0700

On Thu, 29 Jul 2004, Suparna Bhattacharya wrote:

> Hello Judith,
>
> Thanks for reporting this. Sorry, I was travelling and couldn't
> respond earlier.
>
> Could you try a run with a single file and let me know if you
> still see inconsistencies ? The AIO readahead congestion message
> might indicate saturation of the queue during readaheads, making
> the code fallback to a slow path, but I'm not entirely sure.

Hello Suparna;

This took a while to get to them due to our STP release last week, but I
finally did run the tests on your aio patch set again with one 4Gbyte
file.

aio-stress Buffered Reads
stp2-000            295650      295649      295646
sequential reads:  43.89 MB/s  43.88 MB/s  44.07 MB/s
random reads       32.39 MB/s  32.25 MB/s  32.21 MB/s

stp2-003           295654       295645
sequential reads:  46.73 MB/s  1.23 MB/s
random reads       32.12 MB/s  32.14 MB/s


Using only one file, the results are more consistant, except for
sequential reads in test 295645 on host stp2-003, which took inexpectedly
long.

I used these settings:

file size 4096MB, record size 64KB, depth 64, ios per iteration 8
max io_submit 8, buffer alignment set to 4KB
threads 1 files 1 contexts 1 context offset 2MB verification off

The host has 1G RAM.  The filesystem is ext2.

I also still saw the 'AIO readahead congestion' message for at least one
run.

>
> Your options are reasonable from a stress testing perspective,
> though typically for performance testing it makes more sense to
> have individual aio-stress runs for each of the modes -o1 to -o3,
> explicitly, (typically both the non-cached and cached cases) so
> that results aren't affected by the history of earlier accesses
> or readahead state on the file descriptor.

I just changed this and am testing it with some more runs.  I am
running a few with the same settings as that last data set to do a
compare.

Judith

>
> Regards
> Suparna
>
> On Thu, Jul 22, 2004 at 03:12:33PM -0700, Judith Lebzelter wrote:
> > Hello;
> >
> > I have run aio-stress on the OSDL conduit, which includes the recent
> > aio patch set from Suparna Bhattacharya on linux-2.6.7.  I have found
> > inconsistant throughput for buffered sequential and random reads.
> >
> > On 2-CPU/1G/ext2 run with 4 X 1G files:
> >              2.6.7^     aio-patch*  aio-patch** aio-patch***
> >                            run 1       run 2      run 3
> > Seq Read    24.36 MB/s  45.83 MB/s   28.27 MB/s   1.25 MB/s
> > Ran Read     7.14 MB/s   1.33 MB/s    4.08 MB/s  13.71 MB/s
> >
> > Raw data:
> > ^http://khack.osdl.org/stp/295153/results/runlog.buffer.txt
> > ^^http://khack.osdl.org/stp/295109/results/runlog.buffer.txt
> > *http://khack.osdl.org/stp/295135/results/runlog.buffer.txt
> > **http://khack.osdl.org/stp/295132/results/runlog.buffer.txt
> > ***http://khack.osdl.org/stp/295110/results/runlog.buffer.txt
> >
> >
> > For the patched kernel run 295110, the Sequential read throughput
> > was very inconsistant across files.  For run 295135 the Random
> > reads were inconsistant across files.  In general runs do not
> > seem reproducible.
> >
> > However, for the linux-2.6.7 the throughput is generally consistant
> > across files for a single run, and reproducible across seperate
> > runs.
> >
> > I see similar inconsistancy on the 1CPU host and the 4CPU host:
> >     http://khack.osdl.org/stp/295107/results/runlog.buffer.txt
> >     http://khack.osdl.org/stp/295113/results/runlog.buffer.txt
> >
> >
> > I am seeing a message on the console 'AIO readahead congestion' over and
> > over again while the test is running for at least some runs.
> >
> > Are my options realistic?  Or any other suggestions??  I plan to run
> > a few more tests on other filesystem types and with oprofile.
> >
> > Regards;
> >
> > Judith Lebzelter
> > OSDL
> >
> >
> > --
> > To unsubscribe, send a message with 'unsubscribe linux-aio' in
> > the body to majordomo@kvack.org.  For more info on Linux AIO,
> > see: http://www.kvack.org/aio/
> > Don't email: <a href=mailto:"aart@kvack.org">aart@kvack.org</a>
>
>