lkml.org 
[lkml]   [2000]   [Nov]   [3]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectThrash reduction & RE: 2.4.0-test10 Sluggish After Load
Date

-----Original Message-----
From: Christoph Rohland [mailto:cr@sap.com]
Sent: Friday, November 03, 2000 7:54 AM
To: Rik van Riel
Cc: Jonathan George; 'matthew@mattshouse.com';
'linux-kernel@vger.kernel.org'
Subject: Re: 2.4.0-test10 Sluggish After Load


Hi Rik,

>On Wed, 1 Nov 2000, Rik van Riel wrote:
>> The 2.4 VM is basically pretty good when you're not
>> thrashing and has efficient management of the VM until
>> your working set reaches the size of physical memory.
>>
>> But once you hit the thrashing point, the VM falls
>> flat on its face. This is a nasty surprise to many
>> people and I am working on (trivial) thrashing control,
>> but it's not there yet (and not all that important).
>
>I looked into this argument a little bit further:
>In my usual stress tests 12 processes select a random memory object
>out of 15 to mmap() or shmat() it and then access it serially. Each
>segment is 666000000 bytes and I have 8GB of memory. So at one time
>there are at most 666000000*12 bytes = 7.45GB memory attached and in
>use. So I do not see that the machine qualifies as thrashing. Of
>course the memory pressure is very high all the time since we have to
>swap out unused segments.
>
>But the current VM does not behave good at all on that load.
>
>Greetings
> Christoph

I wonder how much of that memory is actually being used by your processes.
My guess is that it's not the whole thing (unless you are running on a 64bit
architecture).

--Jonathan--

P.S. ------------------ THRASH REDUCTION THOUGHTS -----------------

I have given some thought to the issue of thrash recovery, and I have come
up with a simple (conceptually) way to implement a fair thrash reduction
algorithm. However, since I know a minimal amount about the low level
implementation of the Linux scheduler and VM beyond behavioral
characteristics I would love to have some feedback since it should at least
clear up some of my questions about kernel internals.

My idea:
If THIS process is paged in for consecutive scheduler slices
Then
Allow THIS process to run for XX% of the time
it took to swap THIS process in.

Where:
THIS -- is the current runnable process.
(multiple processes for SMP)
XX% -- is the thrash recovery bias.
(from 10-150% is probably reasonable)
time it took to swap --
-- combined system swap time from fault.
(until execution for this scheduler slice)

What do you think?
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
Please read the FAQ at http://www.tux.org/lkml/

\
 
 \ /
  Last update: 2005-03-22 12:45    [W:0.046 / U:3.552 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site