Vernacular high concurrency (2)

background

This is the second part of the three high topics , I thought for a long time about how to connect this topic , Later, I thought I would talk in the direction of traffic , From user request to server to user .

Suddenly 200 Million requests

Our server can bear 200 Ten thousand requests ？ answer , Sure , But the condition can only be an empty link after three handshakes , We know linux All documents in , So one request for one linux It is also a file fd, Then just mine linux The number of files is large enough to receive 200 Million requests , A request defaults to 4k,200w * 4k however 8M There is no problem with the memory around , But the premise is that these are empty links , If the client sends a large amount of data to the server , The server needs to do cpu Do a lot of calculations , After processing, a large amount of data is returned to the client , Don't say that 2000 ten thousand ,1000 It's all hard .

For our secsha scene , A single machine cannot bear 200 Million requests , Therefore, it is necessary to introduce load balancing . Route requests to multiple places .

NGINX

First of all, I thought of nginx,nginx yes 7 Layer load balancing , The amount of concurrency is about 5w about , If you want to support 200 Million concurrent , Need to have 40 individual nginx Talent .

LVS perhaps F5 The server

a nginx 5 Million concurrent , So you need to nginx One more layer above 4 Layer load balancing , All say lvs Concurrent in hundreds of thousands , Assuming that 20 ten thousand , need 10 platform lvs.

DNS Load balancing

a lvs It still seems unbearable 200 Million concurrent , It seems that a layer is needed above him , This floor is DNS, adopt ip To do load balancing , One lvs One ip.

summary

Hundreds of millions of concurrency is also this routine , If one layer is not enough, add another layer ,ip If it's not enough, add a few more ip, If the machine room is not enough, add several more machine rooms .

Reference resources

Taobao double eleven