UDP throughput issue

[borkra]

Have anybody tried to implement UDP throughput test using sendmmsg/recvmmsg on Linux?

Because currently iperf UDP performance test, tests more how fast particular CPU performs context switch from userspace to kernel and back (at least on a single core CPUs), than network throughput.

During TCP test Iperf sends 128K bytes buffer in a single call, so 1 userspace - kernel transition per 128K
During UDP test Iperf sends 1470 bytes bytes buffer in a single call, so nearly 100 - userspace - kernel transition per 128K

I believe using sendmmsg/recvmmsg would allow for UDP test to have performance comparable to TCP.

[borkra]

More like enhancement request, then a question.

[bmah888]

iperf3 isn't about trying to compare the performance between TCP and UDP in terms of socket calls. UDP normally requires one system call per packet sent, whereas TCP allows for some fragmentation...that's just the way the socket API works.

If someone's interested in implementing this, they should consider that sendmmsg and recvmmsg are Linux-specific (at least according to the manual page entries on these two system calls), and iperf3 runs on a variety of other platforms as well.

[borkra]

Few things:

• Yes it's not trying compare performance in terms of socket calls but on some platform it end up being just that especially on slower arms. Thus the problem.

• UDP support fragmentation too up to with packet size 64k, I agree it rearely used.

• Safe to say rear tcp user sends 128K of data in one call, so both benchmark is somewhat artificial. I would just prefer them to be artifical the same way.

[gegles]

Any progress on this? @borkra , I too would be very interested in seeing this happen.

[davidBar-On]

I have sent the pull request #1034 with a proposed implementation to add sendmmsg/recvmmsg support for UDP (not supported when reading from a file or writing to a file). An option --send-recvmmsg was added to enable the feature. When enabled, burst number of packets (defined by the -b option) are all sent by one sendmmsg call.

In my machine (same one for the sender and receiver) improvement is up to about 15% and this is achieved when burst size is about 50.