I don't see how this changes my answer.  It's still interesting.

If you find that some changes improve the accuracy of the split, please
report your results.

On Thu, Sep 01, 2016 at 11:53:37AM +0200, ibrahim menem wrote:
> Hi Ben,
> 
> I did more tests, using different subsets of fields(nw_src, nw_dst, nw_proto, 
> tcp_src, tcp_dst) 
> and against different traffic patterns with smaller IP ranges and varying 
> mainly on src ports.
> 
> The tests yield similar or slightly worse results.
> 
> Is it possible that I'm doing something wrong or it is the hash quality?
> 
> 
> > Date: Tue, 30 Aug 2016 08:05:15 -0700
> > From: b...@ovn.org
> > To: ibrahim...@hotmail.com
> > CC: discuss@openvswitch.org
> > Subject: Re: [ovs-discuss] Hash quality in select group with weighted 
> > buckets‏
> > 
> > On Tue, Aug 30, 2016 at 12:05:20PM +0200, ibrahim menem wrote:
> > > I'm testing "weighted select groups" behaviour on ovs2.5 stable, I'm 
> > > using this command:
> > > 
> > > ovs-ofctl -O OpenFlow15 add-group ovs0 "group_id=1,type=select,
> > > selection_method=hash,fields(nw_src,nw_dst,tcp_src,tcp_dst),
> > > bucket=bucket_id=1,actions=output:11,weight=$1,
> > > bucket=bucket_id=2,actions=output:12,weight=$2" 
> > > 
> > > $1 and $2 correspond to the weights column below:
> > > 
> > > Weights   Measured  Deviation 
> > > 
> > > 50/50      50/50         0
> > > 40/60      34/66         6  
> > > 30/70      22/78         8 
> > > 20/80      13/87         7 
> > > 10/90      5.6/94.4      4.4 
> > > 50/950     97/3          2 
> > > 
> > > 
> > > Traffic pattern (using pktgen_dpdk and only tcp packets):
> > > 
> > > dst.ip      2.0.0.1  ->  2.255.255.255  
> > > src.ip      1.0.0.1  ->  1.255.255.255
> > > dst.port    1        ->  65000
> > > src.port    1        ->  65000
> > > pkt.size    1500
> > > 
> > > 
> > > I
> > >  believed that given enough traffic diversity, weights would be closer 
> > > to the configured ones! I tested against narrower and more realistic 
> > > traffic patterns but still having similar results, any suggestions? 
> > 
> > It's interesting.
> > 
> > If you find that some changes improve the accuracy of the split, please
> > report your results.
>                                         
_______________________________________________
discuss mailing list
discuss@openvswitch.org
http://openvswitch.org/mailman/listinfo/discuss

Reply via email to