6 Jun
2017
6 Jun
'17
9:35 a.m.
All,
Thank you very much for you inputs.
A fair point. The reason I had been focussing on packet loss vs.
congestion is because I had read somewhere that you could have different
transport characteristics depending on whether you were using unicast or
multi/broadcast; differences which might be in play if the problem were
congestion, but not if it were packet loss. Since (correct me if I'm
wrong) NetJack1 uses unicast, whereas NetJack2 uses broad/multicast,
trying NetJack1 would have offered a possible, and close-by, solution. I
was merely trying to tailor the problem to what seemed like the
lowest-hanging solutions. But this is moot now.
.-.
I'm currently rewriting my software and abstracting out the transport,
which I'll do myself instead of relying on Jack (Jack will still be
digesting the audio on the endpoint; just not transporting it there).
Once the framework is in place, I want to try with a simple TCP
transport and see what kind of reliability and latency I can achieve
with that. Perhaps it'll be workable, perhaps not. I can't really afford
the behaviour of the more common over-the-air music players which, as
Chris pointed out, will buffer quite heftily to achieve reliability.
Further on, I'd also like to try something like SCTP
(<https://en.wikipedia.org/wiki/Stream_Control_Transmission_Protocol>)
and see what gives -- although the Windoze support for that seems
sketchy, at best.
--
MC.
If the sound is always garbled, it never even starts
out with decent
quality, then you may be right, latency could be causing every cycle to
not have enough time. Do you get a lot of under-run notifications when
starting? If latency is the only problem you may be able to set a larger
number of network cycles and get around that.
You should be able to get some packet loss statistics from ifconfig, it
would be worth checking to see if there are lots of RX errors or RX
dropped notifications. I don't know how well that information is gathered
for UDP packets, so using iperf to generate traffic and then looking how
many dropped packets were detected would be good. You could probably get
an idea from iperf of the average and maximum latency as well.
In general real time audio transmission over Wi-Fi is going to be
difficult. I have no trouble streaming FLAC files to music players over
Wi-Fi, but those buffer over half a second of audio, and are using TCP so
the TCP layer will just send again any packets which get dropped. When
using jack the network connection has to be able to send packets reliably
with low latency, and no matter what you do Wi-Fi is still going to be a
shared medium, so every time another device anywhere in the local vicinity
transmits, your devices are going to have to listen to see if they need to
switch to a different channel, or pause to allow fair time for another
device. That other device might belong to your neighbor, so you can't
assume that just because you have only two devices connected to your
router that there will not be other interference.
I have run an analyser on my network, and I'm indeed getting a
horrendous (to the naive) amount of packet loss. Given that, it makes
perfect sense that jackd wouldn't be able to get coherent audio together.
>> So
your view is that packet loss is the cause? I had naively assumed it
>> might be due to latency or congestion, but packet loss would make a lot
>> of sense.
In realtime audio a package congestion IS a lost packagage.