Recently, I always see people complaining about game servers being hung up by DDoS, and the latency is ridiculously high after adding a high defense CDN. To be honest, these days, the game server, light to carry the fight can not, but also have to run fast. I have tested several high-defense CDN, and found that many teams simply do not take the node selection and routing optimization seriously, the result is that players are cursing the card while dropping the line.
In fact, high defense CDN latency problems, eighty percent in the node deployment and routing strategy. Many vendors are not moving to blow their own defense bandwidth of how many T, but the nodes are all rented second-hand resources, lines around the earth half circle. Players connected up, the packet must first travel abroad before coming back, this can not card? Especially those who play the “cheap high defense” banner, in fact, is the traffic to the same cleaning center plug, not fried strange.
I previously measured a vendor called CDN07, publicized that the world's 500 nodes, the results of a check of the IP segment, 80% concentrated in North America and Europe. Domestic players connected to the past, the delay directly soared to more than 200ms. There is a more outrageous - a manufacturer of Southeast Asian nodes, the actual physical location in Brazil, this wave of operation is simply called magic transmission.
The program that really works has to start from the two levels of node distribution and routing strategy. First of all, the nodes: must choose the physical distance from the players close, and prioritize the use of BGP multi-line nodes. For example, if you have a lot of domestic players, you have to use the nodes in China, Japan, Korea and Southeast Asia, and open another cluster if you have a lot of players in Europe and the United States. Don't believe those “global unified scheduling” nonsense, transcontinental transmission is not fast.
For example, I optimized a shooting game by replacing a vendor's generic node with 08Host's Asia-Pacific BGP node, and simply adjusted the routing strategy to reduce latency from 180ms to 110ms. Simply adjusted a routing strategy, latency directly from 180ms to 110ms. the key operation is actually two points: one is to disable the default Anycast routing, change to according to the player ISP to select the corresponding single line node; two is to manually configure back to the source of the route, to avoid going to the public network across the operator.
Route optimization is more interesting. Most CDN vendors give the default configuration is “fully automatic intelligent routing”, sounds lofty, in fact, is a brainless selection of the shortest AS path. But the network transmission is not a straight line distance is not necessarily fast, we have to look at the actual link quality. For example, some paths may seem to be a detour, but they go to the operator's intranet, but more stable than a direct connection.
I usually start by measuring a wave of real-time link quality with a tool. For example, do a route trace with MTR and focus on packet loss and jitter:
mtr --report --report-cycles 10 cdn-node.example.comThen manually specify the priority according to the result. For example, telecom users are prioritized on CN2 lines, and mobile users on CMI backbone. Here is a pit: do not rely on the CDN vendor's automatic routing function, they are in order to save costs often stuff the traffic into the cheap links.
I tested a vendor called CDN5, and their “intelligent routing” actually routed the requests of Guangdong users to Indian nodes for the reason that “the load on Indian nodes was lower at that time”. Later, we directly locked the regional routing policy in the background:
After the adjustment, the delay of Asian players directly dropped by 40%, and Europe dropped by 35%. this effect is much more real than adding bandwidth.
Another tip is protocol optimization. Game traffic actually do not need to go to the full TCP proxy, part of the UDP packets can take the direct connection channel. For example, voice chat and real-time location synchronization of such latency-sensitive data, you can go to the point-to-point transmission, only in the first packet through the CDN verification. This can prevent DDoS, but will not introduce additional delay.
One final note on the balance of defense and latency. Many people feel that high defense must sacrifice speed, in fact, not. Like 08Host's Anycast + unicast hybrid architecture is quite smart - usually go unicast low latency line, was hit automatically switch Anycast cleaning. In addition, their TCP acceleration algorithm is really useful, especially in the network environment with a packet loss rate of 3% or more, much faster than traditional TCP.
In fact, there is no silver bullet for latency optimization, the key has to be continuously monitored and adjusted. I used to use Prometheus+Grafana to do real-time monitoring to see a few key indicators: regional delay distribution, TCP retransmission rate, DNS resolution time. Once a node is found to be fluctuating, immediately cut the alternate line.
At the end of the day, a high defense CDN is not something you can buy and rest easy. You have to be like tuning a race car constantly tossing lines and strategies. Those vendors that claim “fully automated optimization”, to put it bluntly, do not want you to see how simple the background. The real reliable program is always the technical team to control the routing right.
Recently, I helped a hand game team to do optimization, migrated their North American node from shared cluster to CDN5“s exclusive node, with custom routing rules, the delay from 140ms to less than 90ms. The feedback from the players is simply two worlds - those who used to curse lagging are now asking ”what kind of black technology have you changed".
So stop blindly stacking bandwidth. Save that little bit of money and instead study the node map and routing table. Sometimes the effect of changing a node and changing a route is more obvious than expanding bandwidth by 10G. Remember, players will only vote with their feet, and a second higher latency may result in one more point of churn.
