Re: Knob turning on mtest01 for 2.4.13-pre5aa1

[Andrea Arcangeli <andrea@suse.de>]

On Sun, Oct 21, 2001 at 11:48:05PM -0400, rwhron@earthlink.net wrote:
> Kernel	2.4.13-pre5aa1
> 
> Goal:	Measure the affect of changing new vm parameters.
> 
> Test:	Run 2 iterations each of LTP tests mtest01 and mmap001.
> 	mtest01 files 80% of virtual memory and writes to each page.
> 	mmap001 mmaps and writes to 100000 pages.
> 	listen to long playing (50+ minutes) mp3 sampled at 128k.
> 	page-cluster=2 for all tests.  (best value so far for 
> 	non-skipping mp3).
> 
> There was only a second or two of mp3 skipping throughout the entire test.
> 
> Pleasantly, one set of values came up on top for both tests, but the
> overall variance isn't that large.  There could be several good values 
> for these parameters.
> 
> mtest01 (2 iterations) with various settings of 2.4.13-pre5aa1 knobs:
> 
> 105 seconds on mtest01 vm_balance_ratio = 16   vm_mapped_ratio = 1   vm_scan_ratio = 16

Cool. this just proofs the vm_mapped_ratio logic is not worthwhile (I
had similar results here so this just confirms).  So I'm killing it
enterely (Linus was completly right that it wans't worthwhile). I'm also
changing a bit the semantics of vm_balance_ratio (similar to pre3aa1)
and I'm lowering it down due the slight change of semantics, plus I'm
including the PG_launder (that resembles the PG_wait_for_IO logic in
pre3aa1) and slightly modified BH_wait_IO logic from Linus. Hopefully
the end result will be positive.

Thanks very much for this great feedback! :)

Andrea