[OpenAFS] Re: "hard mounts"

Phil.Moore@msdw.com Phil.Moore@msdw.com
Mon, 19 Mar 2001 09:40:42 -0500 (EST)

>>>>> "Lyle" == Lyle Seaman <lws@spinnakernet.com> writes:

Phil> We depend on this here at MSDW, and without it, wouldn't be running
Phil> AFS.  See the long winded story below if you don't understand why this
Phil> is so important.

NOTE: I left out the long winded story.  Its actually a pretty good
techno disaster story, but I don't know if people here are terribly
interested in lame attempts at geek literature....

Lyle> I think I can explain this concisely.  MSDW has (had?) an
Lyle> essentially data-less configuration.  98% of the client OS
Lyle> executables actually resided in AFS Read-Only volumes,
Lyle> especially including the X server and some other critical
Lyle> things.

That's basically correct.  Given the poor performance of the AFS
client on large MP machines, we're kept most heavily accessed RW data
our of AFS, and used NFS.

This is one area where NFS has leap frogged AFS in a big way.  MP fast
client and server NFS implementations are commonly available, and the
RW performance blows AFS away.

AFS' fortes are RO redundancy, WAN performance, and administration,
but NOT RW performance.

Lyle> In the event of a network outage, the client would sequentially
Lyle> time-out on each of the three or four replicas, and then X would
Lyle> crash, along with all the applications.  Recovery basically
Lyle> involved rebooting, which turned a four-minute outage into a
Lyle> 15-minute outage.

Lyle> Worse, if the network outage affected a hundred clients, and then a
Lyle> hundred clients were forced to reboot more-or-less simultaneously, they
Lyle> would hit the file servers pretty hard, turning a 15-minute outage into a
Lyle> half-hour outage.

Lyle> (how's that?  about right? )

Prety much, yeah.