Retransmissions in iperf3 vs retransmissions in wireshark

[johnerino123]

Why would tcpdump output (listening on client interface) show much less retransmissions than what iperf shows?

[davidBar-On]

iperf3 tcp retransmissions value is taken from the tcpi_total_retrans of tcp_info. tcp_info includes 3 TCP retransmission related fields: tcpi_retransmits, tcpi_retrans, tcpi_total_retrans. This comment explains that tcpi_retransmits counts only timeout based retransmissions, while tcpi_retrans counts also fast-retransmits, etc. I assume that tcpi_total_retrans counts the same retransmissions as tcpi_retrans (I didn't find where this is documented).

If tcpdump counts only the timeout based retransmissions, then it is expected to be less than tcpi_total_retrans which is reported by iper3.

[ddimopou]

Hello,

I face the same behavior as the one reported here (less retransmissions captured on the client interface than the number reported by iperf3). From https://github.com/esnet/iperf/issues/343, it seems that if the total number of retransmissions reported by iperf includes also the retransmission attempts, then the reason why iperf shows much more retransmissions is that something goes wrong before the packets are placed on the interface.

What could be a possible cause of having failures before packets reach the interface? (could it be CPU limits? / high IO?)

In my case I am testing a MPTCP scenario on Mininet, but I'm just curious to understand the reason behind this discrepancy.
The client sends traffic over two interfaces (10.0.0.2 & 10.0.1.2), and the server receives packets via a single interface (10.0.2.2).

I'd like to capture retransmissions in my tests, and I'm not sure which value is the valid one (i.e. the one coming from iperf or those measured on the wire). It seems that tcpdump will measure retransmission packets that made it to reach the interface and were actually re-transmitted.

Attaching the dump files captured on the client (cl.pcap) and the server side (serv.pcap), as well as the iperf3 client output in json format.

Filtering with "tcp.analysis.retransmission" in wireshark seems to count all types of retransmissions (also spurious and fast). The server side dump captures 49, the client 37 (also not sure why the client reports less than the server -- looks like the client couldn't capture its own retransmitted packets while these were actually sent and captured by the server end), and iperf reports 136.

However, it seems that the majority of retransmissions reported through iperf occur at the beginning of the test. Could this burst during i.e. tcp slow start phase justify some retransmission attempts' failures?

Please correct any of my assumptions above.
Your help to better understand this would be much appreciated.

Thanks,
Dimitris

Iperf3 output on client side: cl.txt