What is the difference between ultra-high-definition 8k and 4k? 4K/8K ultra-high-definition real-time processing and distribution technology

03

Distributed Encoding

After optimizing the above problems, the single-machine live broadcast system can achieve stable 8K 50FPS real-time encoding online. But the challenge we face may be to achieve an encoding speed of 120FPS. We proposed to support distributed transcoding in the live broadcast system. On-demand distributed transcoding should be familiar. After getting a file, cut the file into very small fragments, and send the fragments to multiple machines for processing. Each machine only processes a very short fragment, and finally the transcoded fragments are formed into a complete file.

Live broadcast distributed transcoding actually refers to the on-demand method. Usually, live broadcast transcoding completes all operations such as decoding, watermarking, filtering, encoding, etc. on a single machine, but we have modified the live broadcast process. It does not perform actual encoding, but pulls a live stream and cuts it into small pieces in real time according to the GOP dimension in the live broadcast scenario, sends it to the existing on-demand computing power node, and quickly transcodes it in the form of files. After the conversion, the live broadcast system recycles these small pieces and splices them into a real-time live stream for distribution. The overall idea is similar to on-demand distributed transcoding. From one machine performing all operations to n machines performing real-time transcoding together, the encoding speed and efficiency are greatly improved.

The other part is how to support real-time super-resolution? 4K real-time super-resolution can be achieved through video enhancement, AI enhancement algorithms and other operations, but it is currently difficult to support real-time super-resolution 8K. In this context, we use distributed enhancement capabilities to support super-resolution from 4K to 8K during live broadcasts. After decoding a video frame, the video frame will be compressed and sent to the downstream enhanced computing node in the frame dimension. Each computing node only performs super-resolution operations on a single frame. Through the massive GPU resources of the computing node, super-resolution enhancement from 4K to 8K is achieved in live broadcasts.

04

Network Optimization

As a cloud vendor, there are many scenarios for private cloud deployment, and the network and device environment in the private cloud environment is more complex.

For example, we can see from our monitoring that the entire live transcoding system will experience TCP slowdown every 1-2 hours. The reason may be that after the transcoding service we provide receives the streaming data packet, the ack message is sent from the virtual network card to the physical network card with a 3s delay, which should be instantaneous normally.

First, we suspected that it was a system load problem. When the CPU, memory, and bandwidth utilization were all good, we found that this problem occurred on some machines. The transmission process is TCP→IP layer→bond qdisc→ethX qdisc→ethX pcap. In the process of capturing packets from the virtual network card and the physical network card, we found that the reason for the slow speed was that the bond network card would delay 2-3 seconds to send the ack message. Through in-depth analysis with the netrace tool, we found that from the process of taking packets from qdisc to sending them to the driver, the driver status indicated that the message could not be sent. Finally, we confirmed that the network card driver of the proprietary environment equipment had an exception during large-volume transmission.

05

Distribution Optimization

Another problem encountered was that the customer's network environment was limited, as reflected in the fact that the internal network bandwidth was only a gigabit switch.

Under this condition, a more accurate load balancing algorithm is needed, and higher-performance APIs and system functions are used when sending UDP multicast packets. Here are a few tips:

1. The flow rate of UDP packets can be controlled. Because the encoding bit rate cannot be completely stable and fluctuates greatly, it can be achieved as long as the UDP packet sending rate is controlled below the network bandwidth limit.

2. Make full use of the two network cards of the switch. Configure a bond virtual network card and use this virtual network card for interaction to expand the original single-machine gigabit bandwidth.

After a series of optimizations, the real-time 8K transcoding system was also deployed in the customer's proprietary environment. The system deployed internally by CCTV.com supports real-time 8K encoding.

06

Summary and Outlook

Finally, let’s talk about the summary and outlook. Let’s talk about the transcoding service first:

1. Transcoding services must first optimize the encoder. The encoder optimization is divided into two major directions: first, how to improve the overall encoding parallelism and CPU resource utilization, and second, how to reduce the CPU computing power.

2. Different decoding optimization schemes are implemented for different decoders. For example, for AVS3, the conversion from NV12 to YUV is moved to the encoder kernel layer for operation; for H265, multi-TILE parallel encoding is used for acceleration

3. Solve the bottleneck of memory bandwidth. Optimize memory bandwidth by managing each operation, reducing all memory copies and memory bandwidth usage.

4. Improved stability of the transcoding link. This involves access to remote memory and local memory. It is necessary to plan which CPU each operation will run on, reduce cross-NUMA operations, and improve overall access efficiency.

Transcoding cluster:

1. Distributed transcoding supports up to 8K 120FPS transcoding, 4K to 8K super-resolution, etc. through multi-machine and parallel transcoding capabilities

2. For customer scenarios, we should pay more attention to possible TCP slowness, packet loss and other issues. Secondly, we should smooth the UDP packet sending algorithm and optimize the load balancing algorithm for distribution in the customer's restricted network environment.