What is the new algorithm 'Fugu' that greatly improves the efficiency of streaming?

Many people are refraining from going out due to the new coronavirus, and the effect is

exposed in the form that Netflix, a major video streaming distribution service, is down for one hour . As a result, streaming distribution companies have been forced to reduce bit rates and resolutions to maintain services. Meanwhile, it is reported that an algorithm has appeared that significantly improves the communication efficiency without degrading the image quality of the image.

Learning in situ: a randomized experiment in video streaming | USENIX

Computer scientists create a 'laboratory' to improve streaming video | Stanford School of Engineering

The new communication algorithm 'Fugu' was developed by the research group of

Francis Yang et al. Who are studying computer science at Stanford University. Mr. Yang built the streaming distribution service ' Puffer ' provided by Stanford University, and developed the algorithm by machine learning based on the data obtained there. 'Puffer' is a word that means blowfish in English.

The current popular streaming distribution uses a technology called buffer-based algorithm (BBA), which aims to play video seamlessly. This is to check how much video is stored in the buffer , which is the area of the PC that temporarily stores data, and adjust the quality of the video accordingly.

For example, if the BBA determines that 'there are only 5 seconds worth of video stored on the PC,' then the BBA sends a low quality video to the PC to keep up. On the other hand, if it turns out that '15 seconds are available', the highest quality video, which takes a long time to communicate, is sent. With this method, BBA has been widely used for streaming distribution services for a long time because it can achieve the highest possible image quality while maintaining smooth video playback.

While BBA is simple, yet unsophisticated, many researchers have sought to find more efficient communication algorithms. However, most of these studies were based on machine learning done in virtual environments, and there was a problem that they might not fit well in the real Internet environment.

So Yang et al. Actually launched Puffer, which is an original free streaming distribution service, and distributed video for a total of 38.6 years to 63,508 users. We have trained a

deep neural network with supervised learning using data obtained in a more realistic environment. The algorithm created in this way, Fugu, can predict the data transmission time in advance and perform more efficient congestion control than BBA which only refers to the current buffer status. .

When Yang et al. Actually examined the effects of five algorithms including Fugu in Puffe, Fugu showed excellent performance in terms of image quality and resolution, and the short duration of the image. Also, in a test that delivers video to viewers using a random algorithm, viewers who saw the video delivered via Puffer enjoyed the video for 5 to 9% longer than other algorithms. is.

Keith Winstein, co-author of the paper, said, 'We have discovered how machine learning can overcome the difference between reality and simulation. Given that this will solve many problems, It's very exciting. '

In addition, in the paper, Yang et al., 'This research shows that in order to create a learning algorithm that demonstrates robust performance on the real Internet, real-time training is performed using data from an actual streaming distribution environment. We also found that we needed an algorithm that had a sophisticated structure and simple enough to withstand such training, '' he said.Training in a practical environment is essential for developing a practical algorithm. I said that.

in Web Service,   Video, Posted by log1l_ks