On Sat, 11 Dec 1999 collver@mad.scientist.com wrote:

> I don't believe you can out of the box, but there are hacks floating
> around to do it.
> 
> http://www.math1.rwth-aachen.de/~heine/nfs-swap/

I actually used to follow this quite closely when I still had Linux
machines to maintain.  NFS is great for an unstable development box,
especially when you have a primitive filesystem that doesn't guarantee
metadata consistency.

AFAICT, those hacks are for Linux 2.0.x, and will not work with Linux
2.2.x and later versions of Linux 2.1.x.  These are the versions after
Olaf Kirsch's knfsd patches were integrated.  knfsd allows a server to
accept multiple outstanding NFS commands at one time, rather than making
them all line up synchronously (like our nfsd functionality).  knfsd also
introduced a newer, faster NFS client filesystem, but the new nfsfs didn't
allow creation of UNIX-domain sockets last time I checked, meaning no
running X or syslogd.  You could make a non-swappable ramdisk out of the
buffer cache, of course.

Linux has supported nfsiod/biod-type concurrency for a while, but AFAIK
none of the unintegrated, unmaintained swap-over-NFS code that showed up
over the years ever did this. Linux swap-over-NFS, as far as it exists at
all, can only issue one NFS op at a time.  This is probably why they keep
making the mysterious claim that their archaic RVD-ish Network Block
Device is ``faster'' than NFS.

Can our NFS swap code issue more than one operation at a time, like
nfsiod/biod does for regular file access?

-- 
Miles Nordin / v:1-888-857-2723 fax:+1 530 579-8680
555 Bryant Street PMB 182 / Palo Alto, CA 94301-1700 / US