Setting processor affinities might help, if you can afford the luxury of having one processor per decoder. Let’s imagine you have ten decoders running, then they will placed like this:

Now LiveThread (the thread that streams from cameras) stays at core index 0, all OpenGL operations and frame presenting at core index 1. Default value “-1” indicates that the thread is unbound and that the kernel can switch it from one core to another (normal behaviour). In Valkka, all threads can be bound to a certain processor core. Try for example the value 24 - you get each stream on the screen 24 times, without any performance degradation or the need to decode streams more than once. Replicate demonstrates how Valkka can dump the stream (that’s decoded only once) to multiple X windows. The bigger the buffering time, the more pre-reserved frames you’ll need and the more lag you get into your live streaming. This is necessary for de-jitter (among other things). If you’re planning to use 720p and 1080p cameras, reserve, say 200 frames for both.ĭecoded frames are being queued for “msbuftime” milliseconds. As you learned from the tutorial, in Valkka, frames are pre-reserved on the GPU.

Python ffmpeg rtsp