threads and VMS [Re: dce/rpc "client" api]

[Jeremy Allison]

Bob Mastors wrote:
> 
> A single NT client running netbench will have multiple outstanding
> CIFS commands to a server at many points in time. In other words
> the NT client will not always wait for a CIFS command to complete
> before issuing another one.

That's fine - such things get queued in the TCP layer
in the current smbd design.

> To take advantage of this you could keep the samba model of one
> smbd per connection but redesign smbd to use posix threads.
> This won't make the security context issues any different than
> in the current smbd design.

True - but it does complicate the smbd architecture considerably.
All smbd code would need to be fixed to be thread safe. This is
not a trivial change.

> This may provide a performance boost but it will be very dependent
> on where the bottlenecks in the entire system are.

Currently (on Linux 2.2.x) they're in the TCP stack.
 
> So using this information you could design smbd to use posix threads
> that would be efficient for running benchmarks but maybe not
> so efficient for running real-world loads.

Now I'm not averse to benchmark hacks - but that would take
the biscuit :-) !

> Anyway it sure is a lot easier solving these problems in a kernel environment.

Very true,

	Jeremy.

-- 
--------------------------------------------------------
Buying an operating system without source is like buying
a self-assembly Space Shuttle with no instructions.
--------------------------------------------------------