Current tool provide statistics on latencies (max, histrograms, ...) but I can't see how to use this to investigate why the high latencies occur.
To do that, we usually use ftrace but, then we need to stop the ftrace buffer right when the high-latency is detected. The problem is, then, to detect high latencies on the fly and not after the facts in post-processing.
To do this, I can share 2 ideas:
Make the network stream bidirectional like a ping : the sender knows when it sent a frame, and can compute the latency with the timestamp from the corresponding received frame.
Put a precise timestamp in the SV frame. The receiver can then compute latency with its own receive timestamp.
Current tool provide statistics on latencies (max, histrograms, ...) but I can't see how to use this to investigate why the high latencies occur.
To do that, we usually use ftrace but, then we need to stop the ftrace buffer right when the high-latency is detected. The problem is, then, to detect high latencies on the fly and not after the facts in post-processing.
To do this, I can share 2 ideas: