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

[Andrew Tridgell]

> this might explain why samba's performance ramps up from a factor of 2
> worse than NT with small numbers of netbench clients, crossing over at
> equal performance at around 12 clients, and overtakes by a factor of 2 at
> large numbers of clients.

I think you might have got this backwards Luke. The results I've seen
show Linux/Samba beating NT for small numbers of clients and losing
for larger numbers (at least with a 2.2 kernel). See for example
http://www.mindcraft.com/whitepapers/nts4rhlinux-filethru.gif

I really am _very_ skeptical about threading/non-threading having any
impact on netbench results. I'm sure it would be possible to carefully
construct a benchmark where threading matters, but it would have to be
pretty strange.

Does anybody have any non-anecdotal evidence that the lack of
intra-smbd threading is hurting our performance?

Cheers, Tridge