[OpenAFS-devel] optimizing afs caching (was Re: datagrams really arent big enough?

[Chas Williams]

In message <uocr94luakw.fsf@hurlame.pdl.cs.cmu.edu>,Nat Lanza writes:
>Keep in mind that intelligent caching decisions like this are still an
>active research topic; there are a bunch of things you can try, but
>it's hard to find a good policy that doesn't screw over some access
>patterns.

i fully agree.  we have particular needs that are going to be different
from others.  we like the local caching of smaller files, but for very
large file, afs caching scheme breaks down.  nfs performs 'ok' with
large files (even if you arent doing sequential access)  nothing is
cached locally you are limited to speed of the network.  since everyone
tells me that bandwidth is cheap now, i dont see this as a big problem.
so if i could switch to 'nfs mode' for certain files, that might be
somewhat beneficial.  its just not very practical to work with very
large files residing in afs.  this solution may or may not work.  i
dont have a better answer really.  making the cache faster might be
an answer, but that is really just making the step between the client
and server faster.  why not eliminate it for certain transactions?
i am not saying make the cache go away, its useful.  but for some
operations it just gets in the way.

>Also, exposing extra knobs to the user is pretty uniformly bad. Things
>should just work on their own, because in practice the user will never
>ever remember to tweak stuff appropriately for their workload, if they
>even know how to accurately characterize their workload.

we have a diverse user environment, and there are problems with providing
more control but deciding carte blanche what every user needs isnt the
right answer either.  you cant just tell people, the following option
exists.  you need to provide direction in its use.  this will handle
about 90% of the people.  the last 10% will need things down for them.
afs acl's are fairly complicated but most people seem to be able to
handle them with a bit of guidance.  users arent stupid, but they arent
going to be intuitive either.

>At most, any exposure of the caching system to the user should be on a
>hinting basis -- maybe allow the user to say "I will read this file

well i didnt really expect the user to have control of this.  since 
switching over to nfs mode would really depend mainly on the cache
size which is set by the admin in the first place.