Subject: Re: Building an alternate backing store.
To: None <email@example.com>
From: gabriel rosenkoetter <firstname.lastname@example.org>
Date: 07/14/2000 09:14:52
On Fri, Jul 14, 2000 at 09:10:30PM +1000, Darren Reed wrote:
> In some email I received from Roger Brooks, sie wrote:
> > Surely this is not too different with what happens when a system is
> > booted diskless with an NFS-mounted swap file?
> In addition to that, how about this:
> - create a swap file on your networked host
> - mmap it in
> - mlock that memory in
> - export that file via nfs as a swap file
> or will that not do what I expect ?
The points you're both missing is that if NFS is used at any time a)
you have NFS latency (noticeable), and b) you're STILL writing to
disk, just not the local one (mlock'ed, mmap'ed swap is an interesting
approach, but a touch clumsy, IMHO, in its scaleability).
I want neither of those.
As Erik Fair pointed out, there could well be latency issues in my
implementation, but that's a question to ask after it's implemented.
The principle is that memory (even paged memory) is something that
should not be run out to a file system unless absolutely necessary, so
if there's a another machine within a cluster not using all of its
physical RAM, let's put some pages in the space it's not using.
Please, folks, none of this exists yet and I have a lot of reading to
do before a lot of writing, but I assure you I *am* thinking things
through (my poor initial comprehension of uvm_fault() not being a
particularly good example).
Clearly, Andrew Sporner and I have some discussing to do.
~ g r @ eclipsed.net