Improving sending queue sleepto() accuracy

[maxsharabayko]

The sleeping time for the SRT sending queue is 10 ms, and it is discrete.
That means, if a thread wants to sleep for 5 ms, it will sleep for 10 ms.
If a thread wants to sleep for 15 ms, it will sleep for 10 ms, and then again for another 10 ms, that will make a total sleep time of 20 ms.
This behavior is incorrect and unnecessary.

Fixes #673.

TODO

Accuracy of the new sleepto()

• Windows 10 test sleepto function accuracy
• Linux test sleepto function accuracy
• MacOS test sleepto function accuracy
• Other platforms (iOS, Android, etc.)

Accuracy on Windows

• Check accuracy of CreateWaitableTimer
• Check whether increasing thread/process priority increasing the timing accuracy.

Live mode sending accuracy

• Windows 10 sender accuracy measurements (based on pcapng)
• Linux sender accuracy measurements (based on pcapng)
• MacOS sender accuracy measurements (based on pcapng)

File mode sending accuracy

• Sending

Do we need a maximum sleeping time of 10 ms?

In pseudocode, the sleepto function works like this:

CTimer::sleepto(nexttime) {
    m_ullSchedTime = nexttime;
    rdtsc(now);
    while (now < m_ullSchedTime) {
        condTimedWaitUS(&m_TickCond, &m_TickLock, (m_ullSchedTime- now));
        rdtsc(now);
    }
}

CTimer::sleepto(nexttime) waits on a CV m_TickCond.

CTimer::tick() function

The function CTimer::tick() signals this CV, and it will break the wait to force time checking. The tick() function is called from the receiving thread in CRcvQueue::worker_RetrieveUnit() before every operation of reading data from the socket. This is probably wrong, but requires further review.

CTimer::interrupt() function

In case the thread has to be forced break, there is another function CTimer::interrupt(), that changes the m_ullSchedTime to he current time and calls tick() to break any waits.

Conclusion

This means that maximum waiting time of 10 ms is not necessary.

Do we need a minimum allowed sleeping time?

As shown in #637, the sleepto() function has certain inaccuracy.

A final matrix to be placed here.

CTimer::sleepto() accuracy (proposed implementation) in CV-based waiting mode. Each column shows the actual sleeping time in microseconds for each platform. Measured using unit test CTimer.SleeptoAccuracy.

Time, μs	Win 10	CentOS 7	MacBook Pro	iOS	Android
1	64	49	1
5	209	58	10
10	354	62	17
50	1181	99	68
100	1536	150	140
250	1559	315	389
500	1469	582	738
1000	1593	1080	1346
5000	5506	5110	6118
10000	10717	10112	11963

[maxsharabayko]

Wireshark captures were recorded during sending sessions. Based on the capture files, the time intervals between packets can be calculated. Based on the inter-packet times measurements, mean (average) sending rate, standard deviation and oscilation measurements are calculated on three intervals: 1 sec, 100 ms and 10 ms.
The following results show improvements in sending accuracy on Linux, mostly at high bitrates (above 100 Mbps).

Sender: CentOS 7

500 Mbps

Approximate target inter packet time 24 μs.

Alg	1 sec			100 ms			10 ms
	mean	std	S	mean	std	S	mean	std	S
default	518	40,9	0.000789	517	25,9	0.050163	517	35,6	0.068952
PR	518	31,4	0.000605	517	28,2	0.054498	517	28,8	0.055761

Default sleepto() at 100 ms resolution

PR sleepto() at 100 ms resolution

Default sleepto() at 10 ms resolution

PR sleepto() at 10 ms resolution

100 Mbps

Approximate target inter packet time 120 μs.

Alg	1 sec			100 ms			10 ms
	mean	std	S	mean	std	S	mean	std	S
default	103	1,59	0,015416	103	5,04	0,048903	103	6,28	0,060979
PR	103	1,61	0,015611	103	5,10	0,049462	103	5,79	0,056257

1 Mbps

Approximate target inter packet time 12 ms.

Alg	1 sec			100 ms			10 ms
	mean	std	S	mean	std	S	mean	std	S
default	1,02	0,002	0,00215	1,02	0,006	0,05894	1,02	0,004	0,42051
PR	1,02	0,002	0,00215	1,02	0,006	0,05894	1,02	0,004	0,42051