The new Jitterbuffer in Asterisk -------------------------------- Steve Kann The new jitterbuffer, PLC, and the IAX2-integration of the new jitterbuffer have been integrated into Asterisk. The jitterbuffer is generic and work is going on to implement it in SIP/RTP as well. Also, we've added a feature called "trunktimestamps", which adds individual timestamps to trunked frames within a trunk frame. Here's how to use this stuff: 1) The new jitterbuffer: you must add "jitterbuffer=yes" to either the global part of iax.conf, or to a peer or a user. (just like the old jitterbuffer). Also, you can set "maxjitterbuffer=n", which puts a hard-limit on the size of the jitterbuffer of "n milliseconds". It is not necessary to have the new jitterbuffer on both sides of a call; it works on the receive side only. 2) PLC: The new jitterbuffer detects packet loss. PLC is done to try to recreate these lost packets in the codec decoding stage, as the encoded audio is translated to slinear. PLC is also used to mask jitterbuffer growth. This facility is enabled by default in iLBC and speex, as it has no additional cost. This facility can be enabled in adpcm, alaw, g726, gsm, lpc10, and ulaw by setting genericplc => true in the [plc] section of codecs.conf. 3) Trunktimestamps: To use this, both sides must be using CVS-HEAD; setting "trunktimestamps=yes" in iax.conf will cause your box to send 16-bit timestamps for each trunked frame inside of a trunk frame. The other side must also support this functionality, or else, well, bad things will happen. If you don't use trunktimestamps, there's lots of ways the jitterbuffer can get confused because timestamps aren't necessarily sent through the trunk correctly. 4) Testing and monitoring: You can test the effectiveness of PLC and the new jitterbuffer's detection of loss by using the new CLI command "iax2 test losspct ". This will simulate n percent packet loss coming _in_ to chan_iax2. You should find that with PLC and the new JB, 10 percent packet loss should lead to just a tiny amount of distortion, while without PLC, it would lead to silent gaps in your audio. "iax2 show netstats" shows you statistics for each iax2 call you have up. The columns are "RTT" which is the round-trip time for the last PING, and then a bunch of s tats for both the local side (what you're receiving), and the remote side (what the other end is telling us they are seeing). The remote stats may not be complete if the remote end isn't using the new jitterbuffer. The stats shown are: * Jit: The jitter we have measured (milliseconds) * Del: The maximum delay imposed by the jitterbuffer (milliseconds) * Lost: The number of packets we've detected as lost. * %: The percentage of packets we've detected as lost recently. * Drop: The number of packets we've purposely dropped (to lower latency). * OOO: The number of packets we've received out-of-order * Kpkts: The number of packets we've received / 1000. Reporting problems ================== There's a couple of things that can make calls sound bad using the JB: 1) The JB and PLC can make your calls sound better, but they can't fix everything. If you lost 10 frames in a row, it can't possibly fix that. It really can't help much more than one or two consecutive frames. 2) Bad timestamps: If whatever is generating timestamps to be sent to you generates nonsensical timestamps, it can confuse the jitterbuffer. In particular, discontinuities in timestamps will really upset it: Things like timestamps sequences which go 0, 20, 40, 60, 80, 34000, 34020, 34040, 34060... It's going to think you've got about 34 seconds of jitter in this case, etc.. The right solution to this is to find out what's causing the sender to send us such nonsense, and fix that. But we should also figure out how to make the receiver more robust in cases like this. chan_iax2 will actually help fix this a bit if it's more than 3 seconds or so, but at some point we should try to think of a better way to detect this kind of thing and resynchronize. Different clock rates are handled very gracefully though; it will actually deal with a sender sending 20% faster or slower than you expect just fine. 3) Really strange network delays: If your network "pauses" for like 5 seconds, and then when it restarts, you are sent some packets that are 5 seconds old, we are going to see that as a lot of jitter. We already throw away up to the worst 20 frames like this, though, and the "maxjitterbuffer" parameter should put a limit on what we do in this case. If you do find bad behaviors, here's the information that will help to diagnose this: 1) Describe a) the source of the timestamps and frames: i.e. if they're coming from another chan_iax2 box, a bridged RTP-based channel, an IAX2 softphone, etc.. b) The network between, in brief (i.e. the internet, a local lan, etc). c) What is the problem you're seeing. 3) Take a look and see what iax2 show netstats is saying about the call, and if it makes sense. 4) a tcpdump of the frames, (or, tethereal output from), so we can see the timestamps and delivery times of the frames you're receiving. You can make such a tcpdump with: tcpdump -s 2048 -w /tmp/example.dump udp and port 4569 [and host ] Report bugs in the Asterisk bugtracker, http://bugs.digium.com. Read the bug guidelines before you post a bug. Have fun! -SteveK