Hi Dom,
UDP is without any flow/congestion control. That is, there is nothing to push
back on the sender when it over drives the receiver. Increasing the number of
rx descriptors probably helps a bit but unless the rx nic is faster, I don’t
know if there’s anything else that could avoid the drops.
I’m saying that because one connection should be able to do more than 10Gbps.
But to be sure, does “sh session verbose” indicate that your rx fifo is full?
Regards,
Florin
> On Feb 19, 2020, at 9:30 AM, dchons via Lists.Fd.Io
> <dchons=akouto....@lists.fd.io> wrote:
>
> Hello,
>
> I've been trying to do some basic performance testing on 20.01 using the
> vpp_echo application, and while I'm getting the expected performance with
> TCP, I'm not quite able to achieve what I would expect with UDP. The NICs are
> 10G X520, and on TCP I get around 9.5 Gbps, but with UDP I get about 6.5 Gbps
> with about 30% packet loss.
>
> The commands I use are:
> Server: ./vpp_echo socket-name /tmp/vpp-api.sock uri udp://10.0.0.71/9999
> fifo-size 1000000 uni RX=50Gb TX=0 stats 1 sclose=Y rx-buf 1400 tx-buf 0
> mq-size 100000
> Client: ./vpp_echo socket-name /tmp/vpp-api.sock client uri
> udp://10.0.0.71/9999 fifo-size 1000000 uni TX=50Gb RX=0 stats 1 sclose=Y
> tx-buf 1400 rx-buf 0
>
> (For TCP tests the commands are pretty much the same, except for the URI
> which is tcp://...)
>
> I have a couple of hints but not sure how to make the necessary tweaks to
> improve performance. On the receiver side, vpp# sh hardware-interfaces shows:
> rx offload avail: vlan-strip ipv4-cksum udp-cksum tcp-cksum tcp-lro
> macsec-strip vlan-filter vlan-extend jumbo-frame
> scatter
> security keep-crc
> rx offload active: ipv4-cksum jumbo-frame scatter
>
> I'm thinking that udp-cksum not being active is an issue, is this something
> that I need to explicitly enable somehow? I do have the following in
> startup.conf:
> dpdk {
> dev 0000:05:00.0{
> num-rx-desc 1024
> num-tx-desc 1024
> tso on
> }
> enable-tcp-udp-checksum
> }
>
> My other clue is this (again on the receiver side):
> vpp# sh interface
> Name Idx State MTU (L3/IP4/IP6/MPLS)
> Counter Count
> TenGigabitEthernet5/0/0 1 up 9000/0/0/0 rx
> packets 25107326
> rx bytes
> 36136837440
> tx
> packets 1
> tx bytes
> 60
> drops
> 44
> ip4
> 25107281
> rx-miss
> 11599259
>
> Any tips on what might be causing the rx-miss, or things I should tune to
> improve this for UDP?
>
> Thank you!
> Dom
>
-=-=-=-=-=-=-=-=-=-=-=-
Links: You receive all messages sent to this group.
View/Reply Online (#15466): https://lists.fd.io/g/vpp-dev/message/15466
Mute This Topic: https://lists.fd.io/mt/71401293/21656
Group Owner: vpp-dev+ow...@lists.fd.io
Unsubscribe: https://lists.fd.io/g/vpp-dev/unsub [arch...@mail-archive.com]
-=-=-=-=-=-=-=-=-=-=-=-
