On 11/29/2018 1:05 PM, Templin (US), Fred L wrote:

> iperf3 is a real Internet application in the same way that ping, traceroute,
> tcpdump, etc. are real applications. It is a well-known tool that network
> engineers use on a daily basis and demonstrates that UDP performance
> is highly correlated with UDP datagram size (i.e., even for sizes that
> exceed the PMTU).

AFAIK the main difference between large and small UDP packets is doing
the fragmentation/reassembly in the application instead of in the
kernel's UDP stack. For an 8K packet, that means 6 socket calls in the
application case, versus 1 in the kernel case. That is indeed some
overhead. It is not a big deal for medium speed application like video,
but i can see how it will get in the way of running QUIC at several
Gbps, for instance. But then, the overhead could be trivially eliminated
with API improvements, such as passing several packets in a single call.
We can ponder why we have not seen such improvements yet, the main
explanation being lack of demand. I fully expect that this will change
if QUIC gets widely used. In fact, it would be good if the fragmentation
draft discussed this API issue.

-- Christian Huitema



_______________________________________________
Int-area mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/int-area

Reply via email to