New VM Testcase (2.4.18pre7 SWAPS) (2.4.17-rmap12b OK)

New VM Testcase (2.4.18pre7 SWAPS) (2.4.17-rmap12b OK)

[Roger Larsson]

[-- Attachment #1: Type: text/plain, Size: 667 bytes --]

When examining Karlsbakk problem I got into one quite different myself.

I have a 256MB UP PII 933 MHz.
When running the included program with an option of 200
(serving 200 clients with streaming data a 10MB... on first run
it creates the data, from /dev/urandom - overkill from /dev/null is ok!)

ddteset.sh 200
[testcase initially written by Roy Sigurd Karlsbakk, he does not get
into this - but he has more RAM]

the 2.4.18pre7 goes into deep swap after awhile .
It is impossible to start a new login, et.c. finally
the dd processes begins to be OOM killed... not nice...

the 2.4.17-rmap12b handles this MUCH nicer!

/RogerL

--
Roger Larsson
Skelleftea
Sweden



[-- Attachment #2: ddtest.sh --]
[-- Type: application/x-shellscript, Size: 890 bytes --]

#!/bin/bash
MAX=100
BS=$((1024 * 2048))
COUNT=10
FPATH=/tmp

case $# in
        0)
                i=$COUNT
                ;;

        1)
                i=$1
                ;;
        *)
                printf "Error:\nSyntax: $0 [ numfiles ]\n"
                exit
                ;;
esac

if [[ $i -lt 1 ]]; then
        printf "Can't read $i files\n"
        exit
fi

if test \! -f `printf "%s/file%04d.mp0" $FPATH $i`; then
        c=$i
        echo "Writing $c files..."

        while [[ $c -gt 0 ]]; do
                file=`printf "%s/file%04d.mp0" $FPATH $c`
                touch $file
                dd if=/dev/urandom of=$file bs=$BS count=$COUNT
                c=$(( $c - 1 ))
        done
fi

printf "Reading $i files..\n"

while [[ $i -gt 0 ]]; do
        file=`printf "%s/file%04d.mp0" $FPATH $i`
        dd if=$file of=/dev/null bs=$BS &
        i=$(( $i - 1 ))
done

Re: New VM Testcase (2.4.18pre7 SWAPS) (2.4.17-rmap12b OK)

[Stephan von Krawczynski]

On Mon, 4 Feb 2002 23:24:11 +0100
Roger Larsson <roger.larsson@norran.net> (by way of Roger Larsson <roger.larsson@norran.net>) wrote:

> When examining Karlsbakk problem I got into one quite different myself.
> 
> [...]
> 
> the 2.4.18pre7 goes into deep swap after awhile .
> It is impossible to start a new login, et.c. finally
> the dd processes begins to be OOM killed... not nice...
> 
> the 2.4.17-rmap12b handles this MUCH nicer!

What about -aa ?

Regards,
Stephan

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/

Re: New VM Testcase (2.4.18pre7 SWAPS) (2.4.17-rmap12b OK)

[Ed Tomlinson]

On February 4, 2002 05:24 pm, Roger Larsson wrote:
> When examining Karlsbakk problem I got into one quite different myself.
>
> I have a 256MB UP PII 933 MHz.
> When running the included program with an option of 200
> (serving 200 clients with streaming data a 10MB... on first run
> it creates the data, from /dev/urandom - overkill from /dev/null is ok!)
>
> ddteset.sh 200
> [testcase initially written by Roy Sigurd Karlsbakk, he does not get
> into this - but he has more RAM]
>
> the 2.4.18pre7 goes into deep swap after awhile .
> It is impossible to start a new login, et.c. finally
> the dd processes begins to be OOM killed... not nice...
>
> the 2.4.17-rmap12b handles this MUCH nicer!

Roger what happens if you add my patch that allows the shrink
functions to return the number of pages they free?

(patch was posted to lklm sunday, copy sent privatly to Roger)

Ed Tomlinson
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/

Retest: Re: New VM Testcase (2.4.18pre7 SWAPS) (2.4.17-rmap12b OK)

[Roger Larsson]

On Tuesdayen den 5 February 2002 01.36, Ed Tomlinson wrote:
> On February 4, 2002 05:24 pm, Roger Larsson wrote:
> > When examining Karlsbakk problem I got into one quite different myself.
> >
> > I have a 256MB UP PII 933 MHz.
> > When running the included program with an option of 200
> > (serving 200 clients with streaming data a 10MB... on first run
> > it creates the data, from /dev/urandom - overkill from /dev/null is ok!)
> >
> > ddteset.sh 200
> > [testcase initially written by Roy Sigurd Karlsbakk, he does not get
> > into this - but he has more RAM]
> >

I rerun the testcases, this time the 2.4.18-pre8 did not go as deep - and 
survived. But had worse performance - then I remembered... I had added
file readahead. Retested again - it still survived...

echo "file_readahead:255"> /proc/ide/hda/settings

The results from all tested kernels standard and with readahead follows...

2.4.18pre7
	start: with 'bi' at 12000
	after awhile 'bi' hovers at 7000-9000 (sporadic swap outs)
	after yet some time it starts to swap in too - but without performance loss

2.4.18pre7 with file_readahead:255
	start: bi at 15000-18000
	after awhile heavy swap out: 600(!)-10000
	after yet some time, now with swap in too: 1000-6000

2.4.18pre7 w. Ed Tomlinsons patch
	start: bi at 12000
	stays at: 9000-12000
	more swapout causes throughput loss: 5000-9000
	but finally stabilizes at 7000-10000

2.4.18pre7 w. Ed Tomlinsons patch and file_readahead
	start: bi at 15000-23000
	...
	rather soon ends up in both swapping in and out...
	(about the same throughput at 2.4.18pre7)

2.4.17rmap12c
	during the whole testrun: bi at 10000-12000
	exept for some short dips downto at most 8000

2.4.17rmap12c with file_readahead
	during whole testrun: bi at 20000-23000
	short dips downto 16000 (once 9000)

This should be a picture, but... some other day...

/RogerL

-- 
Roger Larsson
Skelleftea
Sweden
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/