What is the principle of realizing 10 million connections to a single server?

[ahfuzhang]

请问实现单台服务器 1000 万连接的原理是什么呢？

• 假设每个连接需要协议栈的内存为4kb，那么 1000 万连接光协议栈就需要38GB的内存
• 处理 1000万连接的计算量，仅靠 epoll 的能力无认为无法胜任

我看了其他关于 C10M 问题的实现：几乎都是 DPDK+用户态协议栈+多队列网卡等一系列技术的组合才能做到这一点。

谢谢

What is the principle of realizing 10 million connections to a single server?

• Assuming that each connection requires 4 kilobytes of stack memory, then 10 million connections would require 38 gigabytes of stack memory alone.
• The amount of computation required to handle 10 million connections cannot be handled by epoll alone.

I've looked at other implementations of the C10M problem: it's almost always a combination of DPDK + userland stack + multi-queue NICs and a host of other technologies that make this possible.

Thank you!

[lesismal]

欢迎来交流！

目前只试过1M连接，没太大压力。 至于 C10M，epoll的内核协议栈流程太长了应该是确实不适合，DPDK这些，或者路由设备之类的直接在内核实现的方案更适合。

对单节点承载量，我觉得要根据服务节点的协议层次定位来做对应的在线和吞吐追求：

1. 比如路由器交换机这些主要处理网络协议栈的，可以去追求单节点更大连接数和更大吞吐极限
2. 应用业务，单节点承载C10M的意义可能不大，应用层的业务种类太多，不同种类对应不同的、不可预知的开销和优化方案，对应不同的响应性能、对应不同的对其他基础设施的依赖比如数据服务的交互，而很多应用业务节点的扩容也可以通过简单的堆机器就能解决、更容易些，只要追求适度的承载量和性能就可以了

[github-actions[bot]]

This issue is stale because it has been open for 30 days with no activity.

[github-actions[bot]]

This issue was closed because it has been inactive for 14 days since being marked as stale.