Subject: Re: Compressed cache system [Re: Google Summer of Code project ideas]
To: Matt Fleming <mattjfleming@googlemail.com>
From: Hubert Feyrer <hubert@feyrer.de>
List: tech-kern
Date: 04/22/2006 04:12:28
[CC:d tech-kern@ as that's probably a lot more appropriate]


On Fri, 21 Apr 2006, Matt Fleming wrote:
> I was thinking of implementing this (a compressed cache for UVM) as a 
> patchset for UVM, as I feel that a kernel option to turn this on and off 
> would be a little 'icky' (for lack of a better word), _especially_ for people 
> who don't care about this option.

Why is it "icky"?

I'd rather call a patchset that needs to get downloaded and applied a lot 
more "icky" than something that's in-tree, where it's most likely to also 
get maintained a lot better than some out-of-tree patch.

Of course judging that would require some more details, ie. what parts of 
UVM etc. you'd need to modify. I guess using some "compressed pages" thing 
as backing store should fit rather neatly in, and not be too "icky" for
an option or (pseudo)device.


> Also, a question I've been thinking about, some compressed cache systems only 
> store pages that are backed by files and not  swap-backed pages. What is the 
> communities view on this? The research from 
> http://linuxcompressed.sourceforge.net/ shows that good results can still be 
> achieved even if the compressed cache stores swap-backed pages.

As far as I understand your intended project, it is to add an extra layer 
of backing store for memory, i.e. instead of writing data straight to 
swap, you first move it to "compressed swap (memory?)" which then has swap 
as backing store.

What kind of storage "swap" exactly is, if a file or a partition, 
shouldn't matter at all.


  - Hubert