Skip to content
Discussion options

You must be logged in to vote

So in addition to the points above...

We've discovered with 100GE and faster network adapters, achieving high through can also depend on the CPU and I/O topology of the host. Pinning the iperf3 process and thread(s) to CPU cores located "close" to the NIC can make a significant difference. iperf3's -A flag (to set CPU affinity) isn't quite expressive enough to handle this well, but (at least on Linux) we've had good luck with using numactl -N for this. Note that this is a consideration for "actual" high-performance data transfer applications as well.

(There is no intrinsic limit within iperf3 that would limit the speed.)

Replies: 1 comment 1 reply

Comment options

You must be logged in to vote
1 reply
@bmah888
Comment options

Answer selected by SteveSMC
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Category
Q&A
Labels
None yet
3 participants
Converted from issue

This discussion was converted from issue #1968 on November 17, 2025 18:52.