Re: [PATCH net-next-2.6] inetpeer: lower false sharing effect

[Eric Dumazet <eric.dumazet@gmail.com>]

Le vendredi 10 juin 2011 à 10:05 -0700, Tim Chen a écrit :
> On Fri, 2011-06-10 at 06:31 +0200, Eric Dumazet wrote:
> 
> > 
> > Thanks Tim
> > 
> > I have some questions for further optimizations.
> > 
> > 1) How many different destinations are used in your stress load ?
> > 2) Could you provide a distribution of the size of packet lengthes ?
> >    Or maybe the average length would be OK
> > 
> > 
> > 
> 
> Actually I have one load generator and one server connected to each
> other via a 10Gb link.
> 
> The server is a 40 core 4 socket Westmere-EX machine and the load
> generator is a 12 core 2 socket Westmere-EP machine.
> 
> There are 40 memcached daemons on the server each bound to a cpu core
> and listening on a distinctive UDP port.  The load generator has 40
> threads, with each thread sending memcache request to a particular UDP
> port.
> 
> 
> The load generator's memcache request packet has a UDP payload of 25
> bytes.  The response packet from the daemon has a UDP payload of 13
> bytes.
> 
> The UPD packets on the load generator and server are distributed across
> 16 Tx-Rx queues by hashing on the UDP ports (with slight modification of
> the hash flags of ixgbe).

Excellent, thanks for all these details.

I had the idea some weeks ago to add a fast path to udp_sendmsg() for
small messages [doing the user->kernel copy before RCU route lookup],
that should fit your workload (and other typical UDP workloads)

I'll try to cook patches in next days.

Thanks