packet loss in Qbv Demo 3 Scenario 3

[duiduiba]

Hi,
I run the demo in Qbv Demo 3 Scenario 3, and the plotted graph image is perfect. But when reading the generated file latencies.dat which was generated by plot.sh, I found some latencies with 999975 or 1000002, or even bigger, such as 16499976. I think it may be caused by packet loss, and is it normal？

Thanks!

[vincentxw]

Depending on the OS or hardware you're running on, yes. The application is trying to send a packet periodically (e.g. every 500us). However, the OS's scheduler might pre-empt the application and perform some operations that take longer than 500us, resulting in a delay in the next packet. These should be the occasional "misses" or outliers you observe.

Our project/apps were tested in a PREEMPT_RT-enabled kernel and was optimized through some BIOS settings and scripts to ensure our application has sufficient priority to run at a interval with no occassional spikes/misses. The specific version it was verified on is documented here:

[duiduiba]

Hi,
Thanks.

I use the command "tc -stats qdisc show dev ens3" and get info below. It seems the kernel drops the pkt in etf_dequeue_timesortedlist.
qdisc etf 80ec: dev ens3 parent 100:1 clockid TAI delta 5000000 offload on deadline_mode off
Sent 49693632 bytes 776463 pkt (dropped 125, overlimits 125 requeues 0)

In fact, I use the RT kernel with kernel version 5.2.21-rt13, and I suspect the hrtimer of ETF is not so accurately in ns.

[duiduiba]

Hi,
I add -E option when running the app, and print the time gap between now and tstamp in the pkt in function process_socket_error_queue of app, and get the diff almost 10ms. I enlarge the waketx_delay of the app to 19ms and delta of the etf qdisc to 20ms, and it won't happen again.