Why does iperf3 measure asymmetric throughput?

[NullHypothesis]

I'm working on a project that builds on top of gvisor-tap-vsock. I reproduced your iperf3 throughput measurements and noticed that server-to-client traffic is about 3x faster than client-to-server traffic—at least in our more complex test setup.

The README says "[...] depending on which side the test is performed [...]", so this is not news to you. I wonder however why that is? What exactly contributes to the throughput being asymmetric?

[cfergeau]

Guillaume who ran these tests and added this to the README is no longer active on gvisor-tap-vsock, so I'm afraid this knowledge was lost :-/

[NullHypothesis]

Guillaume who ran these tests and added this to the README is no longer active on gvisor-tap-vsock, so I'm afraid this knowledge was lost :-/

Having studied the code for a while, I can now answer my own question. The send and receive path for network packets differs in non-trivial ways. The same is true for the tools cmd/gvproxy and cmd/vm. I haven't measured the exact overhead but I'm fairly confident that this explains the difference. See this PR for some additional context.