Max of PHP FPM_ Some misunderstandings of children

Now? nginx + fpm Basically become the mainstream configuration , What we pay more attention to is pm.max_chindren Configuration of

First , We focus on a premise setting ： pm = static/dynamic,

This option is identification fpm Generation mode of subprocess ：

static ： It means that fpm The runtime directly fork Out pm.max_chindren individual worker process ,

dynamic： Express , Runtime fork Out start_servers A process , As the load increases , Dynamic adjustment , Not more than max_children A process .

Generally recommended static, The advantage is that there is no need to dynamically judge the load , Lifting performance , The disadvantage is to occupy more system memory resources .

The above tells us max_chindren Representative worker The number of processes . It is generally believed that , The more this configuration is, the more concurrency it can handle at the same time , This is a big mistake ：

1） Look at the fpm Related source code , Manage processes and worker The process is through pipe For data communication . So there are many processes , Increase the overhead of process management , Overhead of system process switching , More importantly , Can be executed concurrently fpm The process will not exceed cpu Number . By opening more worker To increase the number of qps, It's a misunderstanding , I won't say that you have opened several more processes , Just a few more cpu To deal with it .

2） but worker Less progress , If server If you are busy , It leads to nginx Type the data to fpm When , Find all woker All at work , There is no free worker To accept the request , Which leads to 502.

that worker How to configure numbers ？

Theoretically woker Number of processes =cpu The number of is the most reasonable , But because of the 2 Reason for point , Maybe each worker Didn't finish processing the request , such , It will appear frequently 502 了 . But ask more , Just say avoid 502, Temporarily put the request hang live , But this is only the way to alleviate , In fact, this will not increase the concurrency of the system , And it will increase the load of the system , therefore , Set a reasonable worker Number is more important .

Martial arts in the world , Fast break not only , Only try to improve the efficiency of the program , Minimize the time of a single request , such , Single worker The processing time of is shortened , There are naturally more requests that can be processed in unit time .

Then you can pass each worker Estimate the number of requests processed per unit time max_children The number of . Suppose the processing time of the largest request is 100ms Inside , And in the 100ms At the same time 100 Please come here , In theory, it needs to be configured 100 individual worker process , First give the request to hang live .

But the biggest request time may be affected by many external conditions , Not a good estimate , In fact, there is a shortcut , To configure your max_children Count , It's that you put max_childnren Set to a larger value , After a period of stable operation , Observe fpm Of status Inside max active processes How much is the , And then put max_children If the configuration is larger than him ok 了 .

I hope these articles can help you .