Recently, we are always asked the same question by our customers: our overseas users are always stuck watching videos as PPT, is there any reliable solution? To be honest, what lies behind this question is actually the hard core of global network latency. I have handled a lot of multinational video projects, the actual test found that many teams think that just buy a CDN can be fixed, the result is the reality of the face - nodes did not choose the right, scheduling strategy sucks, the so-called “global coverage” is simply a paper exercise.
The latency problem of overseas video transmission is essentially a game of physical distance and network topology. Packets traveling across the ocean, every 1000 kilometers jump latency increased by 5-10ms, not to mention the possibility of encountering congestion in the middle of the line or unreliable operators. Some CDN service providers boast of hundreds of nodes, but take a closer look at all crowded in Europe and the United States, Southeast Asia and South America users are still watching the film in circles. These days even CDN have to “prevent teammates” - to prevent those who do not take into account the real network conditions of piggyback program.
The real high defense CDN global coverage, you have to play like a game of Go points. I used to divide the world into eight key regions: North America East and West Coast, Europe Frankfurt and London, Southeast Asia Singapore, Japan Tokyo, Brazil Sao Paulo, Australia Sydney. Each region must be deployed at least 2 high defense POP points, and need to support BGP anycast routing. Don't believe those vendors with a single line node, South African users access the U.S. node bypassing Europe this kind of operation I've seen a lot.
Last year, we did a comparison test for a live broadcast platform migration. The same is from Los Angeles to push the stream to the Japanese users, the use of a traditional CDN average latency reached 380ms, and changed to use CDN5 global intelligent scheduling directly after the pressure to 92ms. the key difference is: CDN5 in the Pacific link went to their own private fiber, while the traditional program is still squeezing the public gateway. Video transmission of this thing, sometimes it is really “money ability” game - not willing to spend money to build a dedicated line, the effect is very different.
The high defense capability must be front-loaded to the edge nodes. A common misconception is to put the cleaning center in the core server room, and so the traffic around the center and then cleaning, latency has long been burst. My strategy is to make each edge node have more than 200G DDoS mitigation capability. Especially for UDP Flood and CC attacks in the video industry, fingerprinting needs to be done at the first access point.
This is the protection configuration template we use on the CDN07 platform:
Special attention should be blocked HTTP1.0/1.1 protocol - many attack tools are too lazy to implement HTTP2, this trick can block out 70% of CC attacks. The test found that after opening the adaptive protection, CPU consumption increased less than 8%, but can resist 800Gbps mixed attacks, the deal is absolutely cost-effective.
Intelligent route scheduling is the real technical barrier. Don't believe in programs that only assign nodes based on GeoIP. The most outrageous case I've seen is: a Chilean user was assigned to a Miami node, and the latency, which should theoretically be 120ms, actually spiked to 400ms - because the traffic took the odd path of Chile-Argentina-Brazil-Florida. Now our strategy is a four-dimensional decision:
This is the core fragment of our routing decision algorithm:
During last year's Double Eleven, a cross-border e-commerce company's live streaming business relied on this set of scheduling algorithms to maintain an average latency of <120ms in Asia-Pacific and <80ms in Europe, despite the global peak traffic reaching 3.2Tbps. Especially for German users, the latency of getting video streams from Frankfurt nodes was pressed down to an amazing 43ms - faster than many local service providers.
The caching strategy must match the business scenario. Simply caching the entire video file is long overdue. For popular live broadcasts, we pre-cache the next 2-3 seconds of video slices in the edge node; for on-demand content, we use dynamic segmented caching - slicing the video into 2-second ts files, and deciding the cache depth based on the heat. Tests show that this scheme saves 37% of storage cost compared to full file caching, while ensuring that the first frame time is controlled within 800ms.
Don't overlook the importance of protocol optimization. Many teams are still sticking to TCP, but they don't know how much advantage QUIC protocol has in cross-border transmission. We have tested in Southeast Asia: for the same video stream, the first screen time of TCP program is 1.8 seconds, while QUIC only takes 0.9 seconds. Especially for wireless network environment, QUIC's 0-RTT handshake and multiplexing can greatly improve the experience. However, it should be noted that QUIC consumes more CPU, so you need to choose a node that supports hardware acceleration.
To be honest, there is no silver bullet to do a good job of global video acceleration, and the key is still piling up the details. From node selection to protocol tuning, from security protection to cost control, each link must be keyed to the extreme. Some customers choose to shrink the program in order to save 20% costs, and as a result, the loss brought about by the loss of users far exceeds the savings. How to calculate this account, smart people should understand.
In the last three years, I have handled the project to verify a rule: no CDN vendor can do the best in all regions. CDN5 is preferred in North America, CDN07 is used in Europe, and 08Host is chosen in Asia-Pacific, especially around China - hybrid multi-cloud is the ultimate solution. Intelligent switching between different service providers through self-developed scheduler, and ultimately realize the goal of latency <150ms anywhere in the world.
Finally, I would like to say something offensive: Don't look at the official website of some vendors blowing up the sky, when you really want to test, remember to ask them to provide monitoring data of real customer cases. I've seen too many cases where customers were fooled with ideal data from the lab. In a real network environment, a triple-optimized solution that can simultaneously achieve low latency, high protection and low cost is a partner worth trusting.
