[OpenAFS] cache performance

[Warren.Yenson@morganstanley.com]

> On Tue, 2002-10-29 at 21:06, Lester Barrows wrote:
> > Whenever a file is accessed on the client, I believe it contacts the cache
> > manager to ensure that it hasn't changed. Perhaps the cache manager, rather
> > than the file server, would be the most authoritative place to collect this
> > information.

On 29 Oct 2002, Nathan Neulinger wrote:
> The cache manager is part of the client. So, yes, it is contacted.
>
> As long as a callback is still present with the server, there shouldn't
> be any communication with the file server.
>
> So, one possible solution would be a cache manager debug set (fs setset)
> that had a very minimal amount of logging generated - to where you could
> reasonably run fstrace regularly on clients. i.e. not a full bore -
> every access, just file opens.

The problem that Phil alludes to is not to get the most short term,
up-to-date information, but over the course of a day, to get the access
times of all volumes accessed by the client.

Since callbacks are 30 minutes for read-write files, and 2 hours (or
thereabouts) for read-only volumes, surely the client will contact the
server sometime during the 24 hour period we are interested in.  As the
callback expires, the client has to send a FetchStatus to renew / extend
the callback, which the server can record.  This is why Phil would like
that fact recorded in the server.

This is still more authoratative for our purposes as there number of
clients is in the order of thousands (to tens of thousands) but the number
of servers is in the low hundreds.  We also control the builds on the
servers much, much more tightly, and can ensure that the audits run and
complete successfully.