Subject: Re: Virtual Memory Subsystem
To: Jukka Marin <>
From: John S. Dyson <>
List: port-i386
Date: 11/26/1996 08:37:27
> > The merged cache helps alot of things, but frankly only helps a little
> > (IMO.)  If someone has an eccentric workload, I believe that they can
> > increase the size of the buffer cache manually...  I think that users
> > would like a reliable OS rather than extracting all the perf from a
> > system.  Merged cache mods are more likely to destablize NetBSD than
> > collapse fixes or pageout perf mods.
> Well, are these exclusive things? :)  StunOS has the merged cache and it's
> also pretty stable.  Well, I understand that making large and complex
> changes in the VM system _will_ make the OS unstable for some time.. and
> I do value reliability _very_ high (even more so after two recent lockups
> on my remote server).
They aren't really exclusive, but I am concerned for those that
might take responsibility for the merged cache changes.  They are extremely
non-trivial.  The efforts to do the collapse fix, pageout fix and the
merged cache are pretty much orthogonal though.  (Maybe a little interference
between pageout and collapse though.)

> But still, it would help with mmap() and make disk IO faster on machines
> with lots of RAM... without devoting 10 MB to disk buffers permanently.
> (Hey, even my Amigas (AmigaOS) run a dynamic disk cache software... ;-)
IMO (and it is really just my opinion) all modern U**X OSes should have
a merged cache.  However, I do want NetBSD to be stable (believe it or
not), and if someone takes the project on, I don't want to underestimate
it (the project) and set someone up for problems worse than they expect.

A collapse fix and pageout fix would certainly be good partial training for
working on the merged cache.