Re: slow access to files

[Diego Roversi]

On Tue, Nov 06, 2001 at 10:04:34PM +0100, Niels Möller wrote:

> > I think that storeio is the only possibility to put a cache mechanism in
> > user space. But I see some drawbacks:
> > - memory user for caching can be paged out 
> 
> If that happens, you're doing it wrong, I think. 

Of course. The problem is that I don't know exactly what is the right way...
:)

> You want to use a
> special pager for the cache. When the kernel tells the pager to page
> out a page in your cache (my understanding of the details on how this
> happens is very vague), you don't want to write it to the swap, you
> can just throw it away, and reread it from the original location when
> it is accessed the next time (ok, that's for read-only access, but
> that's the simplest case).

Uhm, it seems an interesting solution but I'm a bit
perplexed because cache and the pager have to be strongly coupled.

> On the other hand, AFAIK, ext2fs *is* already mmapping the entire
> device (that's the source of the current limit on partition size), so
> I don't know why we're having this problem. Perhaps the general
> mmapping mechanisms can be improved to cache better?

I think so. Beside of the limit of device size, mmaping the device is the
more efficient and elegant solution. For example in this way phisic memory can 
be
fairly used among process and caching data, instead of having a fixed size
cache in storeio.

I'll investigate in gnumach virtual memory...

-- 
Saluti / Regards

Diego Roversi | diegor at maganet.net
              | diegor at tiscalinet.it