RARE user and assistance email list

["mc36" <>]

and to introduce the other party, the guy in cc did some measures and 
compared them to vpp...
in multi-session iperf, both codebases performed equally around 2.2 gbps 
whereas on single-session
iperfing, p4dpdk was about 20% below vpp, somehow... my suggestions were to 
play with the tunables,
(over the regular nic->socket, mem->socket bindings in qemu-kvm)
and seemingly the burst of packets did ~ 10% increase, the rest were 
unmeasurable...
thanks,
cs


On 2/9/22 08:22, mc36 wrote:
> hi,
> so there is an other guy (cc-ed) who started recently playing with the
> p4dpdk... i bounced the today's mailing to the  ...
> he already did 6gbps in a vmware env and recently moved to a new
> environemnt... please consider discussing your findings on-list, maybe
> you could help you each others.... i'm personally not a big expert of
> fine-tuning this stuff, and more interested in developing the code,
> but seemingly you two are interested in the same topic, right now... :)
> thanks,
> cs
>
>
> On 2/9/22 06:24, mc36 wrote:
> > On 2/9/22 06:22, mc36 wrote:
> > > hi,
> > >
> > > On 2/9/22 06:20, mc36 wrote:
> > > > hi,
> > > >
> > > > On 2/9/22 05:59, mc36 wrote:
> > > > > > Maybe vpp runs workers in some specific way, it can run on cores
> > > > > > listed in isolcpus=1-4 kernel.
> > > > > to do that, 0 1 2     1 3 4     would do the trick...
> > > > > but as your previous mail, maybe 0 1 1     1 2 2     ? :)
> > > >
> > > > so i use the dpdk builtin from their examples to spwan the forwarding loop...
> > > > https://github.com/mc36/freeRouter/blob/c6c22bada52209759197736aaceed80ffe063878/misc/native/p4dpdk.h#L475
> > > > could you verify please if it pins itself to the isolcpus?
> > > >
> > > > i personally don't do such because my vcpus easily forward
> > > > at 5gbps from an af_packet interface to my ixgbe,
> > > > (it's expensive because of the user-kernel switching)
> > > > and does the 10gbps between the two ports of the nic...
> > > > but it's native...
> > > > so my questions:
> > > > -you said kvm, but is it on a multi-socket server?
> > > > -is the qemu taskset on the host too, to the nic's socket?
> > > > -is the nic->socket on the same numa node on the host?
> > > > -is the ram->socket on the same numa node on the host?
> > > ^^^^^ so all these for the qemu, on the host....
> >
> > just asking because all these are taken into consideration when allocating 
> > the mbufs in p4dpdk,
> > and all the <port> <rxcore> <txcore> is all about that... (mostly...:)
> >
> > regards,
> > cs


-=-=-=-=-=-=-=-=-=-=-=-
Groups.io Links: You receive all messages sent to this group.
View/Reply Online (#153): https://groups.io/g/freertr/message/153
Mute This Topic: https://groups.io/mt/89016789/6413194
Group Owner: 
Unsubscribe: https://groups.io/g/freertr/unsub []
-=-=-=-=-=-=-=-=-=-=-=-