[OpenAFS] switch from memcache to diskcache

[Derek Atkins]

First, I don't understand -- if your machines are diskless then
how does a "disk cache" save startup time?  You're not saving data
to any disk -- you HAVE no disk.

Based on your design, the only way to shut AFS down cleanly is to
umount /etc, /usr, and so forth, _then_ shutdown AFS from within the
root ramdisk (the same way it was started), and then complete the
shutdown.  All you need to do is "unmount /afs" after all references
are cleared.

Note that AFS was never designed or meant to be a root file system, so
you shouldn't complain when it doesn't work ;)

Also note that you -can- have "dataless" workstations that have disks
-- systems that are centrally administered and maintained, with a
centralized image...  If you do that then all this goes away.  Yes, it
adds aboout $100 to the cost of your machines, but frankly it saves
your network.  There is no way your network could deal with a restart
avalanche of 1000 diskless machines!

-derek

Adrian Knoth <adi@drcomp.erfurt.thur.de> writes:

> On Thu, Oct 31, 2002 at 01:32:05PM -0500, Derek Atkins wrote:
>=20
> > I dont understand your file system layout.=20=20
>=20
> The kernel creates an initial ramdisk (minix) and loads the AFS-modules.
> A cache-partition is active. Afterwards, afsd is startet and the whole=20
> system from the /afs-tree is bind-mounted over the existing /sbin, /bin
> and so on. Then init is replaced by the standard init and the system
> boots.
>=20
> > If you are running diskless (frankly a silly thing to do in this day=20
> > and age of $100 80-GB disks),
>=20
> It is not silly. We're running this setup on 15 clients at university.
> It will be extended to hundred or more clients in the near future.
>=20
> By doing this we have absolutely no administrative efforts/costs, because
> a centralized installation is used for all clients. Because of some
> modifications there is even no need for host-specific /etc.
>=20
> > then why don't you always run AFS with a memcache?=20=20
>=20
> This would be good if we have more RAM. And a disk-cache is nice for
> reducing the network-bottleneck, even between reboots.
>=20
> > Are you saying that you are:
> >         1) running diskless, and
> >         2) running your _ROOT FILESYSTEM_ out of AFS???
>=20
> A virtual root-fs, yes. / is ramdisk, but in fact all entries below /bin,
> /etc, /usr and so is imported via AFS.
>=20
>=20
> --=20
> mail: adi@thur.de  	http://adi.thur.de	PGP: v2-key via keyserver
>=20
> Wer die H=E4nde in den Scho=DF legt, mu=DF noch lange nicht unt=E4tig sein

--=20
       Derek Atkins, SB '93 MIT EE, SM '95 MIT Media Laboratory
       Member, MIT Student Information Processing Board  (SIPB)
       URL: http://web.mit.edu/warlord/    PP-ASEL-IA     N1NWH
       warlord@MIT.EDU                        PGP key available