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 ? 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
Skellefte?
Sweden
On Mon, 4 Feb 2002 23:24:11 +0100
Roger Larsson <[email protected]> (by way of Roger Larsson <[email protected]>) 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
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 ? 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
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 ? 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
Skellefte?
Sweden