Salsify – A New Architecture for Real-time Internet Video

My main advice for implementers would be, benchmark your end-to-end glass-to-glass video latency (including the time spent waiting for the next frame to be captured & encoded, and then the end-to-end latency through to the display of that frame) over varied/unpredictable networks. In my experience, implementers sometimes get caught up focusing too much on network-layer measurements (IP latency) and can end up missing what in my view is the bottom line: glass-to-glass video latency. You can use our mahimahi mm-delay/mm-link/mm-onoff/mm-loss tools (part of Debian/Ubuntu) and the included traces to model some bad networks. And then see how you do on the same traces we use in our paper. If you can make a plot like our Figure 8(a) and it all looks good, that seems like good progress in my book.

(1) Salsify is mostly about how you control the video encoder (e.g., a VP8/VP9/AV1/H.264/H.265 encoder) and transport protocol (e.g., RTP/UDP). H.264, RTP, and UDP themselves don't say anything about the control part, i.e., how to (1) estimate the network path's varying capacity, and then (2) how to adapt the desired frame size to match that capacity, and then (3) how to actually encode a frame of video to match the desired compressed frame size or bitrate.

If you have an unpredictable/variable network path, Salsify is probably better than the control strategies in systems with less tightly controlled encoders and transports, e.g. WebRTC.org/Chrome, Hangouts, Skype, or FaceTime. If the network is mostly constant or at least predictable, Salsify's not going to be helpful. So, bottom line is "maybe."

(2) Getting 20ms p95 glass-to-glass latency with compressed digital video is pretty difficult even under ideal circumstances, and basically impossible with a consumer webcam. At 60 Hz, just the interval between two frames is already 17 ms! So if you have one frame's worth of buffer between the camera/USB/encoder/sender and a frame's worth of buffer in the receiver/video card/display, you're already toast. You would really have to work hard to pipeline everything. Even the very best gaming monitors have latencies on the order of 10 milliseconds, and it's hard to buy a USB camera that even starts giving you any bits from the picture before the exposure is over. (And v4l2 doesn't return from VIDIOC_DQBUF until the frame is completely received, so you're probably looking at changing the kernel if you want to use UVC.) So <20ms I think is really hard. DJI just released an end-to-end custom-engineered system that claims 28ms latency (https://www.dji.com/fpv/info#specs).

If you look at Figure 8(a) from the Salsify paper, Chrome (using the WebRTC.org codebase) is getting per-frame latencies on the order of 600 milliseconds even when the network is totally constant and perfect, and Salsify's per-frame latencies are around 250 milliseconds but less consistent. Then you have to factor in the extra latency associated with waiting for the next frame -- if these systems are running at 10 fps (look at the extreme sparsity of the dots after the network hiccup, especially for WebRTC), there's 100 ms of extra glass-to-glass latency right there. Getting to <20ms @p95 is just another world from where these programs are today.

Salsify is mainly about the benefits you can get if you (a) extract the control loop out of the video codec, and make it expose a functional-style API, (b) use a Sprout-like congestion control algorithm that tries to follow evolving network capacity quickly and estimate "how many bytes can we send right now while trying to maintain a bound on end-to-end delay", and (c) have a single control loop that works every frame and has the choice of whether to send (1) a frame whose coded length is already known and is about equal to what we think the network can handle [it is very hard to get this from any existing video encoder on a single pass!], or (2) no frame at all.

You can certainly do this within the WebRTC protocol, but doing it within the WebRTC.org codebase is going to be a lot of work. :-( I think unfortunately the codebase has some pretty deeply ingrained assumptions about how the control flow is going to go. Inverting that (as we propose), I don't think is an easy incremental change.

Now, you might ask whether there are some incremental improvements you could make to WebRTC.org to get 80% of the benefits of Salsify without a major refactor. E.g., maybe you don't need the functional API if you just make some tweaks to the rate-control algorithm and congestion control. I don't think we know for sure though, but I suspect that for somebody already experienced with that codebase, probably yes, there are gains to be had. There's another question about whether the gains of Salsify (which are on a particular type of flaky cellular network) might come at the expense of costs on other types of (dependable?) networks. To be confident on that we'd probably need to try this stuff much more broadly and on real people (this is the motivation for Puffer).

One of the challenges here is that it's hard to know how big a compressed frame will be until you compress it, and that takes time. Ideally an encoder would be able to produce a compressed frame that's the exact size that can be accommodated by the network path, right away. In practice that's not so easy. Salsify gets around this by having the encoder encode two candidate coded frames (one bigger, one smaller) and then giving the application the option of sending either one (after encoding is finished and it knows their exact length and quality) or no frame at all. If the encoder could just do the job perfectly and instantly the first time, you wouldn't need all that.

Not really a breakthrough though, JPEG is just an old format. But maybe machine learning stuff will get the 4x video compression seen in the show.