enable histogram latency reporting for isochronous workloads, add working...
A means to measure network responsiveness and throughput
Brought to you by:
rjmcmahon
Many latency sensitive applications are isochronous, some of these are constant bitrate. A useful (and commonly done) test is to generate isochronous, constant packet size, traffic in order to measure latency (e.g. pheist's iRTT test). This test is sometimes done in the presence of other workloads (e.g. multiple TCP connections) in order to test latency under load.
The bounceback test is a poor substitute when latency variability is high, since it makes more measurements when the latency is low and fewer measurements when the latency is high, thus biasing the resulting statistics toward the lower latency measurements.
Here is a run with UDP. Use --histograms on the server side. The T8 PDF is packets and F8 is frames. No working-load on this run. One can use the --ipg option too if that's needed between the 20 packet sends.
TCP run w/o a working load
Last edit: Robert McMahon 2023-11-28
Code fix for working load with tcp. A UDP flow with a TCP requires the use of pyflows or requires a separate server.
Run with TCP and working loads
Use the --working-load option on the server so the server side can start two threads, one for TCP and one for UDP