Inaccuracy with bandwidth shaping

[LeoCosseron]

Hello,

We used Kollaps to measure the TCP throughput between two nodes, by using the flent tool (which is based on netperf). We perform the measurement directly in one of the containers.

It seems that with Kollaps, the maximum throughput we can achieve is around 95.5%-96% of the link capacity, while on bare-metal the maximum throughput is only around 94.2% of the link capacity.

We believe this is because Kollaps does not take into account the 24 bytes overhead of the Ethernet protocol (8 bytes for the preamble, 4 bytes for CRC, 12 bytes for inter-packet gap).

I attached the Kollaps topology we used (2 containers on the same machine, 100Mbps link bandwidth), and the results file, in an archive.

The command we used to launch flent is flent tcp_upload -p totals -l 30 -H n1 --socket-stats kollaps_bandwidth.zip

[sebastiaoamaro]

Thank you for the issue, we will take a look into this.