How does iperf3 calculate udp throughput

[cyanideangel2000]

Hello Team,

Quick query, is the iperf3 UDP throughput a reliable metric?
I am using iperf3 version 3.16

I ran the following command to pump packets over UDP between two hosts
./iperf3 -c 10.136.57.72 -p 2291 -t 60 -i 5 -P 8 -Z -u -b 30G

When I see the output i notice:
Connecting to host 10.136.57.72, port 2291
[ 5] local 10.136.16.114 port 51916 connected to 10.136.57.72 port 2291
[ 7] local 10.136.16.114 port 49101 connected to 10.136.57.72 port 2291
[ 9] local 10.136.16.114 port 40756 connected to 10.136.57.72 port 2291
[ 11] local 10.136.16.114 port 37425 connected to 10.136.57.72 port 2291
[ 13] local 10.136.16.114 port 53046 connected to 10.136.57.72 port 2291
[ 15] local 10.136.16.114 port 48840 connected to 10.136.57.72 port 2291
[ 17] local 10.136.16.114 port 35327 connected to 10.136.57.72 port 2291
[ 19] local 10.136.16.114 port 39682 connected to 10.136.57.72 port 2291
[ ID] Interval Transfer Bitrate Total Datagrams
[ 5] 0.00-5.00 sec 1.54 GBytes 2.65 Gbits/sec 1142085
[ 7] 0.00-5.00 sec 1.18 GBytes 2.03 Gbits/sec 875029
[ 9] 0.00-5.00 sec 1.60 GBytes 2.75 Gbits/sec 1186688
[ 11] 0.00-5.00 sec 1.54 GBytes 2.64 Gbits/sec 1140354
[ 13] 0.00-5.00 sec 1.46 GBytes 2.52 Gbits/sec 1086011
[ 15] 0.00-5.00 sec 1.07 GBytes 1.84 Gbits/sec 793978
[ 17] 0.00-5.00 sec 1.39 GBytes 2.38 Gbits/sec 1028043
[ 19] 0.00-5.00 sec 1.36 GBytes 2.34 Gbits/sec 1008826
[SUM] 0.00-5.00 sec 11.1 GBytes 19.1 Gbits/sec 8261014

---

[ 5] 5.00-10.00 sec 1.52 GBytes 2.62 Gbits/sec 1129406
[ 7] 5.00-10.00 sec 1.16 GBytes 2.00 Gbits/sec 862803
[ 9] 5.00-10.00 sec 1.79 GBytes 3.08 Gbits/sec 1330759
[ 11] 5.00-10.00 sec 1.52 GBytes 2.61 Gbits/sec 1126821
[ 13] 5.00-10.00 sec 1.49 GBytes 2.55 Gbits/sec 1102738
[ 15] 5.00-10.00 sec 1.03 GBytes 1.76 Gbits/sec 760682
[ 17] 5.00-10.00 sec 1.37 GBytes 2.35 Gbits/sec 1014828
[ 19] 5.00-10.00 sec 1.36 GBytes 2.33 Gbits/sec 1006596
[SUM] 5.00-10.00 sec 11.2 GBytes 19.3 Gbits/sec 8334633

=====
[SUM] 0.00-60.00 sec 135 GBytes 19.4 Gbits/sec 0.000 ms 0/100291189 (0%) sender
[SUM] 0.00-60.00 sec 98.2 GBytes 14.1 Gbits/sec 0.007 ms 27446878/100288639 (27%) receiver

Mine is a single connection upload test, however I see throughput being reported for both sender and receiver.
How do we read the throughput numbers for sender and receiver.
Does it mean the sender could pump at the rate of 19.4Gbps out of the node, however receiver due to 27% loss was receiving at the rate of 14Gbps?

Thank you.

[davidBar-On]

Hi @cyanideangel2000,

Does it mean the sender could pump at the rate of 19.4Gbps out of the node, however receiver due to 27% loss was receiving at the rate of 14Gbps?

In general this is the case. The client shows the sent throughput and the server shows the received throughput, which s affected by the network throughput.

If you want to better understand the performance you may try using different packet sizes (-l option). For UDP the default is usually 1460 bytes. Larger sizes (e.g. 8K, 32K) may improve throughput, although this is not for sure because of the added fragmentation. (Note by the way that -Z is applicable only for TCP).

Also, set -b to just above the expected throughput per stream. E.g. -b 3G per the results you sent. This is because trying to send redundant messages overload the CPU / buffers / etc. and reduces performance.

Another thing to try is changing the number of streams (-P), as if several streams run on the same processor, that may result in lower throughput.