Commercial CDN returns to its original intention: focus on details and maximize distribution efficiency

On October 20, Live Video Stackcon 2017 was held in Beijing. Technical elites in the audio and video industry gathered together to discuss the technical evolution and development of audio and video in 2017 and future trends. CDN, as an important part of audio and video content distribution, naturally cannot be absent. Akamai, Wangsu, Yunfan Accelerator, etc. all attended the event. Yunfan Accelerator co-founder and CTO Fu Kai was invited to share his technology titled "CDN in the Big Video Era" at the conference.

[[207307]]

2017 Audio and Video Conference - Yunfan Acceleration Fukai Speech Site

Is CDN's original intention still there?

As we all know, CDN came into being with the development of the Internet in the 1990s, solving the situation of "the world is waiting". This is the original meaning of CDN and the original intention of the birth of the CDN industry - to serve Internet companies and be a strong support behind the Internet. Today, the video industry has experienced explosive growth in the update and iteration of technologies such as live broadcast, VR, and 4k. CDN, as a rigid demand in the video field, has entered a fierce competition in this environment. The blue ocean market has gradually turned into a red ocean. All companies have tried their best to break through, and the price of CDN has been reduced again and again, but is the original intention behind this still there? What is the ideal CDN service for video companies?

[[207308]]

Fu Kai, Co-founder and CTO of Yunfan Accelerator

Fu Kai, a senior CDN expert with 15 years of experience, has served as an architect and CTO in many Party A companies such as Tudou and Yinyuetai. In this sharing, he combined his entrepreneurial experience of Party A company and Yunfan Accelerator to introduce how CDN companies can provide technical services to video companies from both sides, returning to the original intention.

The troubles of the explosive growth of video companies: file storage and user experience

Video companies have also been experiencing twists and turns during their explosive growth. Although China's overall video technology level has improved a lot, according to extensive market research, there are still two most important "hard nut to crack" problems in the industry: file storage and user experience. These two issues are also a matter of life and death for video companies.

First of all, file storage is the most basic and critical module of the video website system capacity. With the development of UGC, for large video websites, a large amount of file data will be uploaded from the user end to the website every day, and then transcoded into different clarity levels. At this time, a video will become multiple copies, such as HD, SD, and smooth. This poses an extremely severe challenge to the storage space of the website server, and the storage space required for videos is rapidly increasing.

Generally speaking, the file processing process involves many levels, and the classification and definition of files are particularly important. The following is the file processing process of a large video website:

In terms of storage, Fu Kai said that the files uploaded by users will be transcoded and reviewed to generate different levels of clarity. Then, the transcoded video files with different levels of clarity will be stored in at least three computer rooms such as A/B/C. This will undoubtedly increase the burden on the storage and space occupancy, but it is necessary for the sake of availability. When any of the computer rooms A, B or C is abnormal, the other two computer rooms can work normally. When three copies are stored, it can be guaranteed that at least two computer rooms can work normally when there are abnormalities.

Fu Kai mentioned that in terms of user experience, after the uploaded files are accessed several times by users, the big data analysis system will define the files as unpopular files and hot files based on these access records. The popular files will then be pushed to the edge nodes for nearby calls, while the unpopular files are still stored in the upper-level servers and are directly retrieved from the upper level when called. In addition, the website editors themselves will also mark a file as hot as needed.

The above-mentioned integrated storage experience solution seems to be the best, but in fact, there are some problems: there may be waste of edge storage space, some files may not be accessible, and performance issues of small operators or specific regions. For example, in the list of newly uploaded files on all websites, how many files are pushed to the edge? Moreover, file access itself has regional characteristics. There will definitely be differences in viewing habits between netizens in Beijing and netizens in Shanghai. Small operators are completely different, and the definition of hot and cold files is calculated by algorithms and is universal. Just like the average value cannot show the personalized characteristics of the region, it will cause a large number of upper-level nodes to visit.

Commercial CDN returns to its original intention: focus on details and maximize distribution efficiency

Fu Kai introduced in detail how commercial CDN acceleration handles this aspect: First of all, the existence of commercial CDN is to supplement and support the content distribution of super-large video websites, and the operating principle is " exchange space for traffic, and exchange back-source for space ". Therefore, compared with video websites, it has a more distinct hierarchy in architecture.

In the CDN architecture, users access the nearest edge node, and the files in the storage space of the node serve users. The process is to exchange space for traffic, but due to the limited storage space of the edge server, some cold spot files will be retrieved by returning to the source. As a supplement to the content distribution of super-large video enterprises, commercial CDN can supplement the access to files with regional characteristics. In addition, the upper layer still uses its own, which further ensures the security of the source station.

Customized switches, EdgeControl for logic processing, Cache software for storage management

A single CDN node mainly solves the problems of storage and speed. Yunfan Accelerator has customized an ospf switch for batch management and batch configuration of computer rooms, which can shield downed machines at any time.

In the actual user access process, each customer has different characteristics and the logic (demand) is also very strange. Therefore, the edge node is divided into two parts, EdgeControl is used for logical processing, and Cache is used for storage. Through 7 layers of hash, each file is stored only once on the node. Because only one copy is stored, it is easy to be abnormal when there are a large number of concurrent hot spots. Therefore, when hot requests are sent, Yunfan will use hot spot migration. Yunfan acceleration technology uses queues, that is, URLs are queued. When the queue value of the URL reaches a certain number, it will be immediately stored in all edges to improve processing efficiency.

When Cache processes file storage, it does not store the entire file, but divides the file into 1M storage blocks and then stores it in our file system. Therefore, the block storage of the entire hard disk is physically stored in blocks on the hard disk, but when the user's HTTP request comes, the response sent is a logically complete file. Therefore, these physical blocks are arranged to be evenly distributed in the hard disk, which brings the following benefits;

1. When there are a large number of IOs on a file, it can ensure that all IOs are consistent;

2. When one of the hard disks fails, you only need to go to the upper layer to retrieve the files on this hard disk and put them on other disks. This can effectively avoid downtime and effectively improve the hard disk data processing efficiency. It can withstand even the "Four Returnees" announcing their love affairs at the same time.

3. Even if a hard disk crashes, it will not affect the file transmission. Yunfan Accelerator can retrieve the storage modules distributed in the crashed hard disk by returning to the source and evenly distribute them in other hard disks.

4. On-demand storage and on-demand return to the source, with high drag efficiency. For example, when a user is watching Wolf Warrior 2, according to viewing habits, the opening and ending credits are likely to be skipped, so when storing, the opening and ending credits will be stored in the upper layer and taken down because there is no access, while the main part is stored at the edge.

5. When deleting files, because the files are divided into several modules and spread across various hard disks, they will not occupy IO when deleted and only the files will be deleted from the memory records.

In addition, Fu Kai emphasized a detail of Yunfan Accelerator in file storage - directory refresh, which also made a qualitative breakthrough. Basically, once the file path is modified in storage, it can take effect immediately on the entire network, which is leading in the industry. In addition, Yunfan Accelerator has made further optimizations in business language based on the business it is exposed to. Customers can customize the business language according to their own habits, which directly greatly reduces the steps of repetition in the program. Operation and maintenance do not need to submit work orders, and can directly locate the problem based on the business language code.

How does CDN help video companies improve user experience?

CDN was originally created to solve the problem of "the world is waiting", so what the industry needs to solve is distribution efficiency, shortening the time required for content to reach the audience as much as possible.

Here we have to mention intelligent scheduling, and the Yunfan acceleration scheduling system is divided into two parts: traffic scheduling and edge scheduling.

Generally speaking, scheduling is global scheduling, with a scheduling center to coordinate all scheduling requests. However, when actually handling problems, it is found that when this type of scheduling receives a large number of requests, it is easy to fail due to low processing capacity due to dynamic requests and the need to query many related systems. User-requested files will not be able to reach the user end smoothly, and users will enter a long wait. Therefore, Yunfan Accelerator has made a new attempt in scheduling, marginalizing scheduling and adopting decentralized scheduling. All caches are scheduling servers, and there are as many schedulers as there are cache servers. When accessing the scheduler, the scheduler will distinguish whether this is a scheduling request or a user's final access request, and then process it according to the result. Not only 302 can be adjusted in this way, but even DNS can be adjusted in this way. "In order to prevent DNS from being killed, every machine is also a DNS server."

There is also a classic scenario in the process of dynamic back-to-source. In this structure, edge users request the nearest node, and the nearest node returns to the nearest upper source. But sometimes the edge node cannot access the upper layer. In this structure, any point or edge point can immediately switch to other edges as the proxy upper layer. In Yunfan's architecture, all nodes can be used as the upper layer or the lower layer at the same time. There is no need to worry about who is the upper layer and who is the lower layer, so that the file transfer path is unobstructed and can be reached at one touch.

Future Trends of CDN

Fu Kai, as a senior CDN expert with 15 years of experience in the industry, has witnessed the growth of CDN from scratch. In his eyes, the future development of CDN is bound to develop towards a differentiated route, such as the CDN+P2P approach. Yunfan Accelerator has P2P cooperation with many well-known companies at home and abroad. On the other hand, Yunfan is vigorously promoting its own security services such as anti-DDOS services, traffic cleaning services, and WAF services. It can also do a lot of AI things, such as sending the number of connections and IO usage of its own data center machines to big data for learning, automatic verification, and automatic detection of faults and anomalies.

Fu Kai pointed out that most people used to think of CDN as a static service, but it is actually the most universal gateway interface at the front end? It is closest to users and can do a lot of things. I believe that in the future, enterprises will need more small innovations like Yunfan Accelerator has made for network transmission. "Exquisite and beautiful" is the "original intention" and pursuit of CDN companies and customers for a long time.