Low Latency DASH Nginx RTMP

Question

I use arut nginx-rtmp-module (https://github.com/arut/nginx-rtmp-module) on the media server, then I tried to stream using FFmpeg to the dash application, then I test the stream by playing it using VLC.

And it waits around 30secs to start playing, and it plays from the beginning, not the current timestamp.

This is my current config on the RTMP block

rtmp {
    server {
        listen 1935;

        application live {
            live on;

           exec ffmpeg -re -i rtmp://localhost:1935/live/$name
              -c:a libfdk_aac -b:a 32k  -c:v libx264 -b:v 128K -f flv rtmp://localhost:1935/hls/$name_low
              -c:a libfdk_aac -b:a 64k  -c:v libx264 -b:v 256k -f flv rtmp://localhost:1935/hls/$name_mid
              -c:a libfdk_aac -b:a 128k -c:v libx264 -b:v 512K -f flv rtmp://localhost:1935/hls/$name_hi
              -c:a libfdk_aac -b:a 128k -c:v libx264 -b:v 512K -f flv rtmp://localhost:1935/dash/$name_dash;
        }

        application hls {
             live on;

             hls on;
             hls_path /tmp/hls;
             hls_nested on;

             hls_variant _low BANDWIDTH=160000;
             hls_variant _mid BANDWIDTH=320000;
             hls_variant _hi  BANDWIDTH=640000;
        }

        application dash {
            live on;

            dash on;
            dash_path /tmp/dash;
            dash_nested on;
        }
    }
}

This is the command I use for streaming

ffmpeg -re -i 2014\ SPRING.mp4 -c copy -f flv 
rtmp://52.221.221.163:1935/dash/spring

How can I reduce the delay, and make it play from the same timestamp as the streamer?

Can I achieve under 5s latency?

UPDATE

Tried to change the playlist length and fragment length, using this directive

dash_playlist_length 10s;
dash_fragment 2s;

But still got some latency problem, sometimes it's smaller than before, sometimes it's the same

Solution 1

Can I achieve under 5s latency?

No. DASH is a segmented protocol, meaning your media is chopped up into relatively large chunks. The player has to download some chunks before it can start playing them. Your encoder has to upload entire chunks before these chunks even appear in the manifest. This is the wrong tool for the job, and any attempts at reducing latency by cranking the chunk size down are adding massive overhead to your project. You're using the wrong tool for the job if latency is important to you.

How can I reduce the delay, and make it play from the same timestamp as the streamer?

You can't. Physics! It's impossible for you to play the same thing at the exact same time as it is being encoded. You're sending data over a packet-switched network, with many encoding/decoding steps in the way which all require a buffer as they work in chunks. The only way to playback what's coming in simultaneously is to go analog... at least there your only delay is the speed of light.

The best you can do is switch to a protocol designed for low latency, like WebRTC. Just be sure you understand the tradeoffs. Your codecs will be optimized for latency, not quality... so your quality will suffer. WebRTC over UDP (optional but common) means that some packets will get lost, causing your viewing experience to suffer. When you care about latency, it doesn't matter so much if you lose a chunk here or there, what matters is that you keep going. You can use WebRTC over TCP and keep your reliability at only a slight increase in latency.

Decide what really matters to you. In almost every case, it isn't actually low latency. You can't have it all ways. There are tradeoffs to every approach. You must decide what is best for your specific situation.

Solution 2

I also have the same issue with VLC media player. Most of the latency is by the client player, you can use the ffplayer with no buffer to check it.

ffplay -fflags nobuffer rtmp://192.168.1.66/myapp/live

Results of mine,

• latency of VLC: 6~7s
• latency of ffplay: 500ms

For more refer to comment of narlex of issue how to reduce latency on github

Solution 3

You may need to change the GOP size in ffmpeg command. The default GOP size for ffmpeg is 250 which means there will be a key frame every 250 frames. If your output is 25fps, then you will have a key frame every 10s in worst case (if scene cut detection is enabled, you may have shorter key frame interval).

For both HLS and DASH, the segments must start with a key frame. So you will have a lot of segments with 10s duration. You need to reduce the segment duration (the GOP size) in order to reduce latency.

Try to modify your ffmpeg command as follow to see if it helps

exec ffmpeg -re -i rtmp://localhost:1935/live/$name
          -c:a libfdk_aac -b:a 32k  -c:v libx264 -g 50 -b:v 128K -f flv rtmp://localhost:1935/hls/$name_low
          -c:a libfdk_aac -b:a 64k  -c:v libx264 -g 50 -b:v 256k -f flv rtmp://localhost:1935/hls/$name_mid
          -c:a libfdk_aac -b:a 128k -c:v libx264 -g 50 -b:v 512K -f flv rtmp://localhost:1935/hls/$name_hi
          -c:a libfdk_aac -b:a 128k -c:v libx264 -g 50 -b:v 512K -f flv rtmp://localhost:1935/dash/$name_dash;