[Ntop-misc] pfring_multichannel - packets are not evenly distributed

[Alfredo Cardigliano]

I agree with Miguel.
I just wanted to add that in order to solve/mitigate (but not limited to) this issue, we are developing a library (libzero) on top of DNA that is able to balance packets across threads or applications in zero-copy.

Regards
Alfredo

On Apr 16, 2012, at 5:17 PM, Miguel Quesada wrote:

> You have to take into account that the NIC is the one distributing the packets over to each queue (channel) when RSS is enabled (which is what I infer from your example)
> In the case of 82599-based cards the hash function used to choose what channel a packet is sent over is pretty much determined by the TCP/IP header of the packet. Refer to  http://www.intel.com/content/www/us/en/ethernet-controllers/82599-10-gbe-controller-datasheet.html  for more information.
> 
> Regards,
> -mq
> 
> On Mon, Apr 16, 2012 at 7:51 AM, Jaime Nebrera <jnebrera at eneotecnologia.com> wrote:
> Hi Emre,
> 
> Pfring clustering is limited to hash, thus if original traffic is not
> well balanced (nat, proxy, ...) load won't be balanced
> 
> Jaime Nebrera
> Enviado / Sent iPhone
> 
> El 16/04/2012, a las 15:46, Emre Güler <gueler at internet-sicherheit.de> escribió:
> 
> > I'm using PF_RING with ixgbe-3.7.17-DNA. I'm sending a pcap file with
> > "./pfsend -i dna0 -f /tmp/dump.pcap -n 1 -r 10". This is what
> > "./pfcount_multichannel -i dna0" says (I made it more readable):
> >
> > Absolute Stats: [channel=0][2220 pkts rcvd][0 pkts dropped]
> > Absolute Stats: [channel=1][521 pkts rcvd][0 pkts dropped]
> > Absolute Stats: [channel=2][514 pkts rcvd][0 pkts dropped]
> > Absolute Stats: [channel=3][326 pkts rcvd][0 pkts dropped]
> > Absolute Stats: [channel=4][1383179 pkts rcvd][0 pkts dropped]
> > Absolute Stats: [channel=5][319 pkts rcvd][0 pkts dropped]
> > Absolute Stats: [channel=6][430 pkts rcvd][0 pkts dropped]
> > Absolute Stats: [channel=7][4164 pkts rcvd][0 pkts dropped]
> > Absolute Stats: [channel=8][398 pkts rcvd][0 pkts dropped]
> > Absolute Stats: [channel=9][674886 pkts rcvd][0 pkts dropped]
> >
> > If you look at channel 4, you can see that it processed 1383179 packets,
> > while channel 5 only processed 319 (it's the same amount everytime I
> > repeat the test with the same pcap file).
> > Is there anything I can do (without changing the pcpap file) to make the
> > distribution of the packets more fair? Or is this something I have no
> > control over?
> > I thought that using pfring_set_cluster with cluster_round_robin is the
> > solution to my problem, but DNA doesn't seem to support set_cluster.
> >
> > Regards,
> > Emre
> > _______________________________________________
> > Ntop-misc mailing list
> > Ntop-misc at listgateway.unipi.it
> > http://listgateway.unipi.it/mailman/listinfo/ntop-misc
> _______________________________________________
> Ntop-misc mailing list
> Ntop-misc at listgateway.unipi.it
> http://listgateway.unipi.it/mailman/listinfo/ntop-misc
> 
> _______________________________________________
> Ntop-misc mailing list
> Ntop-misc at listgateway.unipi.it
> http://listgateway.unipi.it/mailman/listinfo/ntop-misc

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://listgateway.unipi.it/pipermail/ntop-misc/attachments/20120416/3c30e52a/attachment.htm>