James Mastros wrote:
>
> Hey all.
> I was looking through prescan, I put in a little monprint for instructions
> being prescanned, figuring that speeding up the prescan is probably the
> easiest way for me to wring out some speed, I noted that some instructions
> are prescanned up to 160 times before even finishing Uncompressing Linux.
>
> It seems to me that caching the metadata on about three pages would be
> pretty much optimal -- from the log, it looks like it tends to alternate
> between two pages fairly often, and sometimes three.
>
> I expected to find that this effected largely the BIOS, but was proven wrong
> -- the most scanned parts of the BIOS are C0000:E6-FE at 24 scans.
>
> And also, it doesn't look like a large change in methodology. I'm not
> certian, however, about doing the mappings properly -- the meta cache is a
> nexus structure, if I'm reading the code right.
Quite interesting. Are writes to the code pages flushing them,
causing the 160 iterations? One of the changes I have in mind
is to only flush code meta data when prescanned regions are
stepped on. When writes hit unscanned regions, there's no
poing.
-Kevin
--
=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-
Kevin Lawton [EMAIL PROTECTED]
MandrakeSoft, Inc. Plex86 developer
http://www.linux-mandrake.com/ http://www.plex86.org/
