Re: Out of VM idea

Re: Out of VM idea

[Zlatko Calusic]

Rik van Riel <H.H.vanRiel@phys.uu.nl> writes:

> On Tue, 28 Apr 1998, George wrote:
> > On Tue, 28 Apr 1998, Rik van Riel wrote:
> > 
> > >Following some observations from Michael Remski (sent
> > >to me by private e-mail), I've come to the conlusion
> > >that we really should do something about out-of-VM
> > >situations.
> > 
> > At the moment, (2.1.98), I can lock my 64 MB machine up with a 'make
> > MAKE='make -j20' zImage'.
> > 
> > At the time of memory death:
> > * It has 4 megabytes of free pages.
> > * It has 6 megabytes of buffer memory.
> > * But it dies because it has 0 swap left.
> > 
> > Those hard limits on memory how much memory to not grab should definitely
> > go. 
> 
> You can tune the buffermem & pagecache amount of memory
> in /proc/sys/vm/{buffermem,pagecache}.

Every time before he starts compiling, and then return to old values
when he's finished?

IMNSHO, kernel should be autotuning.

> But why your system has 4 MB of free memory I really
> don't know...

mm/page_alloc.c (in free_memory_available()):

	/*
	 * If we have more than about 6% of all memory free,
	 * consider it to be good enough for anything.
	 * It may not be, due to fragmentation, but we
	 * don't want to keep on forever trying to find
	 * free unfragmented memory.
	 */
	if (nr_free_pages > num_physpages >> 4)
		return nr+1;

With 64MB of memory, last 4MB are almost never used!!!

MM in last kernels is not very good.

Except Stephens great improvements of the swapping system, where he
did a really good job, I believe we did a step backward with recent
changes.

Regards,
-- 
Posted by Zlatko Calusic           E-mail: <Zlatko.Calusic@CARNet.hr>
---------------------------------------------------------------------
	  (A)bort, (R)etry, (P)retend this never happened...

Re: Out of VM idea

[Rik van Riel]

On 29 Apr 1998, Zlatko Calusic wrote:

> > You can tune the buffermem & pagecache amount of memory
> > in /proc/sys/vm/{buffermem,pagecache}.
> 
> Every time before he starts compiling, and then return to old values
> when he's finished?
> 
> IMNSHO, kernel should be autotuning.

How do you propose we should do this? The round-robin
deallocation and on-demand allocation of buffer/user
pages are somewhat auto-tuning.
Maybe we should age the page cache & buffermem pages
to achieve a more LRU-like discarding scheme (the
buffer pages are thrown out randomly at the moment).

> > But why your system has 4 MB of free memory I really
> > don't know...

> 	if (nr_free_pages > num_physpages >> 4)
> 		return nr+1;
> 
> With 64MB of memory, last 4MB are almost never used!!!

I believe George said something about my patch, with
which the number should be lower.
Anyway, the freepages number should be sysctl tunable,
together with kswapd agressiveness and clustering
size.

> MM in last kernels is not very good.

True, but maybe Linus will integrate my patch, which
makes the kernel behave somewhat more predictable, and
which has a builtin low/high watermark so thrashing is
reduced.

Rik.
+-------------------------------------------+--------------------------+
| Linux: - LinuxHQ MM-patches page          | Scouting       webmaster |
|        - kswapd ask-him & complain-to guy | Vries    cubscout leader |
|     http://www.phys.uu.nl/~riel/          | <H.H.vanRiel@phys.uu.nl> |
+-------------------------------------------+--------------------------+