[OmniOS-discuss] illumos bug #2869 duplicate packets with vnics over aggrs

Richard Elling richard.elling at richardelling.com
Sun Jul 22 01:09:04 EDT 2012


You need to set the window size and ignore measurements of the startup.

Also, FWIW, I've not found L4 to be particularly useful. Link aggr is just lame.

Finally, dladm show-aggr -s, is your friend.
 -- Richard


On Jul 20, 2012, at 9:18 PM, Denis Cheong <denis at denisandyuki.net> wrote:

> Thanks Theo, I will see what I can pull.  I think grabbing just the aggr module is the best option at the moment.
> 
> The guest is Gentoo (Kernel 3.2.12).
> 
> Here's the performance bottleneck I'm seeing ..
> 
> gazoo is OmniOS host with 2 x igb in L4 aggr with main lan on VLAN 300
> slate is Gentoo KVM on OmniOS with 1 x virtio vnic on aggr VLAN 300
> myth is Gentoo phys with 1 x nvidia single gige with main lan on VLAN 300
> 
> Issue:
> gazoo <> myth = good
> anything <> slate = bad
> 
> Examples:
> 
> myth ~ # iperf -c gazoo
> ------------------------------------------------------------
> Client connecting to gazoo, TCP port 5001
> TCP window size: 22.5 KByte (default)
> ------------------------------------------------------------
> [  3] local 192.168.245.136 port 60917 connected with 192.168.245.134 port 5001
> [ ID] Interval       Transfer     Bandwidth
> [  3]  0.0-10.0 sec  1.09 GBytes   932 Mbits/sec
> 
> root at gazoo ~ # iperf -c myth
> ------------------------------------------------------------
> Client connecting to myth, TCP port 5001
> TCP window size: 48.0 KByte (default)
> ------------------------------------------------------------
> [  4] local 192.168.245.134 port 36001 connected with 192.168.245.136 port 5001
> [ ID] Interval       Transfer     Bandwidth
> [  4]  0.0-10.0 sec  1.05 GBytes    899 Mbits/sec
> 
> root at gazoo ~ # iperf -c slate
> ------------------------------------------------------------
> Client connecting to slate, TCP port 5001
> TCP window size: 48.0 KByte (default)
> ------------------------------------------------------------
> [  4] local 192.168.245.134 port 51500 connected with 192.168.245.152 port 5001
> [ ID] Interval       Transfer     Bandwidth
> [  4]  0.0-10.0 sec  26.5 MBytes  22.2 Mbits/sec
> 
> 
> myth ~ # iperf -c slate
> ------------------------------------------------------------
> Client connecting to slate, TCP port 5001
> TCP window size: 22.5 KByte (default)
> ------------------------------------------------------------
> [  3] local 192.168.245.136 port 56461 connected with 192.168.245.152 port 5001
> [ ID] Interval       Transfer     Bandwidth
> [  3]  0.0-10.2 sec  91.8 MBytes  75.3 Mbits/sec
> 
> 
> 
> On Mon, Jul 9, 2012 at 10:53 AM, Theo Schlossnagle <jesus at omniti.com> wrote:
> Yes. If you switch to the bloody repo, you should be able to do a pkg
> update and get a new kernel.
> 
> The release publisher is: http://pkg.omniti.com/omnios/release/
> Bloody's publisher is: http://pkg.omniti.com/omnios/bloody/
> 
> pkg unset-publisher omnios
> pkg set-publisher -g http://pkg.omniti.com/omnios/bloody/ omnios
> pkg update
> 
> This should give you a new BE running bloody to reboot into.
> 
> Alternatively, you could just snag the aggr module from bloody and
> replace the one you have.
> 
> There are a couple of severe performance pathologies regarding
> networking under qemu/kvm, can you say what the guest is?
> 
> On Sun, Jul 8, 2012 at 8:24 PM, Denis Cheong <denis at denisandyuki.net> wrote:
> > Hi Theo
> >
> > Is there an easy way that I can test your proposed fix to this issue?
> >
> > I am having major network performance bottlenecks in KVM and I suspect that
> > it has something to do with the duplicate packets, so would like to test it
> > out.
> 
> 
> --
> Theo Schlossnagle
> 
> http://omniti.com/is/theo-schlossnagle
> 
> 
> _______________________________________________
> OmniOS-discuss mailing list
> OmniOS-discuss at lists.omniti.com
> http://lists.omniti.com/mailman/listinfo/omnios-discuss
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://omniosce.org/ml-archive/attachments/20120721/0e869154/attachment.html>


More information about the OmniOS-discuss mailing list