Open Stack

On 02/25/2015 05:52 AM, Miguel Ángel Ajo wrote:
> I’m writing a plan/script to benchmark OVS+OF(CT) vs
> OVS+LB+iptables+ipsets,
> so we can make sure there’s a real difference before jumping into any
> OpenFlow security group filters when we have connection tracking in OVS.
>
> The plan is to keep all of it in a single multicore host, and make
> all the measures within it, to make sure we just measure the
> difference due to the software layers.
>
> Suggestions or ideas on what to measure are welcome, there’s an initial
> draft here:
>
> https://github.com/mangelajo/ovs-experiments/tree/master/ovs-ct

Conditions to be benchmarked

     Initial connection establishment time
     Max throughput on the same CPU

Large MTUs and stateless offloads can mask a multitude of path-length 
sins.  And there is a great deal more to performance than Mbit/s. While 
some of that may be covered by the first item via the likes of say 
netperf TCP_CRR or TCP_CC testing, I would suggest that in addition to a 
focus on Mbit/s (which I assume is the focus of the second item) there 
is something for packet per second performance.  Something like netperf 
TCP_RR and perhaps aggregate TCP_RR or UDP_RR testing.

Doesn't have to be netperf, that is simply the hammer I wield :)

What follows may be a bit of perfect being the enemy of the good, or 
mission creep...

On the same CPU would certainly simplify things, but it will almost 
certainly exhibit different processor data cache behaviour than actually 
going through a physical network with a multi-core system.  Physical 
NICs will possibly (probably?) have RSS going, which may cause cache 
lines to be pulled around.  The way packets will be buffered will differ 
as well.  Etc etc.  How well the different solutions scale with cores is 
definitely a difference of interest between the two sofware layers.

rick