> On 13 Apr 2017, at 16:05, Ernst, Eric <eric.er...@intel.com> wrote:
>
> Thanks for the quick reply, Maciek.
Np.
>
> Just to be double sure, your jumbo frame results should be ~1.3Mpps/16 _G_bps,
> not _H_, right? I assume you aren't using humble-burst measurements.
More like “Heroic-" :)
Fat fingers on my side - sorry - it should be:
s/Hbps/Gbps/
>
> I looked through some of the links, and understand that generally performance
> should scale with frequency/number of threads, though mileage may vary when
> dealing with VMs. With this in mind, are the numbers you're quoting dealing
> with
> just a single worker thread on that single core?
Yes, single VPP worker thread on single core.
Generally we expect VPP throughput to scale linearly with increasing worker
threads
and cores.
-Maciek
>
> Thanks,
> Eric
>
> On Thu, Apr 13, 2017 at 11:52:26AM +0000, Maciek Konstantynowicz (mkonstan)
> wrote:
>> +csit-dev
>>
>> Eric,
>>
>> The way we test vpp vhostuser in CSIT is by generating packet streams
>> from external traffic generator per VM topology shown at this link:
>>
>>
>> https://docs.fd.io/csit/rls1701/report/vpp_performance_tests/overview.html#tested-physical-topologies
>>
>> VM vhostuser testing methodology is described here:
>>
>>
>> https://docs.fd.io/csit/rls1701/report/vpp_performance_tests/overview.html#methodology-kvm-vm-vhost
>> https://wiki.fd.io/view/CSIT/csit-perf-env-tuning-ubuntu1604
>>
>> Reported measured non-drop-rate throughput, on machines in FD.io labs,
>> is at the level of 3Mpps for 64B and and 1.3Mpps/16Hbps for 1518B
>> Ethernet frames, for VPP data plane thread (worker thread) running on a
>> single physical core.
>>
>> Note that this is an external throughput i.e. from traffic generator
>> perspective. To get the actual VPP pkt forwarding capacity, numbers need
>> to be multiplied by two, as each packet is switched by VPP twice in each
>> direction: NIC-to-VM and VM-to-NIC.
>>
>> Here pointers to throughput graphs and detailed results listings:
>>
>>
>> https://docs.fd.io/csit/rls1701/report/vpp_performance_tests/packet_throughput_graphs/vm_vhost.html#ndr-throughput
>>
>> https://docs.fd.io/csit/rls1701/report/vpp_performance_tests/packet_throughput_graphs/vm_vhost.html#pdr-throughput
>>
>> https://docs.fd.io/csit/rls1701/report/detailed_test_results/vpp_performance_results/vpp_performance_results.html#vm-vhost-connections
>>
>> In case relevant, FD.io CSIT performance test environment is described here:
>>
>>
>> https://docs.fd.io/csit/rls1701/report/vpp_performance_tests/test_environment.html
>>
>> For VM tests, Qemu is configured with tso4 and tso6 off:
>>
>> https://git.fd.io/csit/tree/resources/libraries/python/QemuUtils.py
>>
>> hth
>>
>> -Maciek
>>
>>> On 13 Apr 2017, at 05:26, Ernst, Eric <eric.er...@intel.com> wrote:
>>>
>>> Hey,
>>>
>>> I've been reading through the various use case examples at wiki.fd.io, and
>>> after reading through
>>> https://wiki.fd.io/view/VPP/Tutorial_Routing_and_Switching,
>>> I came up with a recipe for testing:
>>>
>>> VM <- vpp vhost-user - l2 bridge domain - vpp vhost-user -> VM
>>>
>>> For reference, I describe the setup at:
>>> https://gist.github.com/egernst/5982ae6f0590cd83330faafacc3fd545
>>>
>>> After verifying connectivity, I used iperf3 to get baseline bandwith
>>> numbers.
>>>
>>> I am seeing on the order of ~45 Mbits/sec in this configuration, using
>>> default
>>> VPP config options, on a DP Xeon system running on the same socket. I was
>>> surprised
>>> by this, so ran a similar test using vhost/veth, connecting two namespaces
>>> also
>>> through a l2 bridge domain. In this case I saw ~2 Gbits/sec. Better, but
>>> still surprising.
>>>
>>> Is there something (obvious) wrong in my base setup (ie, using a l2
>>> bridge-domain?)?
>>> I think it'd be useful to have a "best" example on the site for VM
>>> connectivity,
>>> but first want to make sure I'm down the right path before
>>> submitting/contributing.
>>>
>>> While I don't think it'd account for this amount of degradataion, I'm
>>> curious if
>>> there is TSO support in VPP?
>>>
>>> Finally, @ ONS Ed gave a great overview of VPP, including sample benchmark
>>> numbers.
>>> Are there similar results for vhost-user enabled VMs?
>>>
>>> Thanks,
>>> Eric
>>> _______________________________________________
>>> vpp-dev mailing list
>>> vpp-dev@lists.fd.io
>>> https://lists.fd.io/mailman/listinfo/vpp-dev
>>
_______________________________________________
vpp-dev mailing list
vpp-dev@lists.fd.io
https://lists.fd.io/mailman/listinfo/vpp-dev
