new class / geom_cache / request for comments

[R. B. Riddick]

--- Eric Anderson <anderson at centtech.com> wrote:
> R. B. Riddick wrote:
> > --- Eric Anderson <anderson at centtech.com> wrote:
> > > Just curious, how is this different/better than the regular buffer
> > > cache?
> > Do you mean this cache of those file systems?
> > If yes: The answer should be: "It just makes everything worse."
> > (there is not even some read ahead...)
> 
> Not sure what I'm missing here, please clue me in.  I guess I am indeed 
> talking about the cache normally used for filesystems.  If I read a 
> 500MB file in, first time it comes from disk, second time it comes from 
> memory cache, correct?  How is geom_cache different?
>
Yes, I think so...
But geom_cache is just useful, when file system's buffer cache cannot help.

E. g.:
A degraded RAID5 on 4 consumers (3 good plus 1 failed).
When we want to get a data block, that resides on the failed consumer, we have
to read all corresponding blocks (2+1) in order to rebuild the missing block.
When we do a sequential read, we would have to read the consumers, that hold
the data blocks twice (2 x 2).
So the geom_cache could help here (2+1 real reads plus 2 from the cache), if
the provider is not too busy.

> > It is just useful, if you dont have any other caches (e. g. a ufs on a
> > geom_raid5 (I think I should have it tomorrow... :-) ) on some geom_cache
> > providers)...
> 
> I suppose I just need to play with it to completely understand..
> 
:-)
Maybe it is useless... and I dont see it... and nobody dares to say it...
I saw, that the CVS tree does not have a sys/geom/cache directory...

-Arne


__________________________________________________
Do You Yahoo!?
Tired of spam?  Yahoo! Mail has the best spam protection around 
http://mail.yahoo.com