Note: While I found this issue on Apache httpd, it may apply to any http server out there.
The Apache httpd server controls this feature through 3 configuration directives, all of them in the core module:
KeepAlive: on/off, default on;
KeepAliveTimeout: seconds, default 5;
MaxKeepAliveRequests: positive number, default 100.
There is no setting in Apache httpd to somehow allow KeepAlive and non-KeepAlive requests at the same time (e.g. to allow up to 100 KeepAlive requests, what comes above to be treated differently). One must choose the server behavior from the very beginning.
Now it’s math time. Starting from the MaxClients value (default 256), what is the request rate (new clients / second) that can be served without compromising the user experience? MaxClients you may say, but let’s not draw conclusions too fast. There are some issues to be considered:
The time between opening the connection and the KeepAliveTimeout expiration. On a default configuration, at the very least 5 seconds, but on a more typical side, maybe 7-10 seconds;
Traffic fluctuation (spikes);
Internal Apache httpd time (spawning new processes to handle new connections, etc).
It’s getting complicated. But assuming a typical 7 seconds time for every process that serves a client and an uniform behavior (all or almost all clients keep a server process busy or waiting for data for 7 seconds), on a default setting of 256 MaxClients, the uniform traffic rate that can be served is 256/7 = 36 (new) requests / second. Any spike larger than that will cause page loading delays and a poor user experience.
If disabling KeepAlive is not an option, all the planning should start from the worst expected spike that is to be handled (e.g. 2x or 3x the average during the busiest period). For a 50 req/s busy period average, the server should be able to allow 100 or even 150 req/s without compromising the user experience. If the configuration is already maxed out from the hardware point of view, then other solutions should be looked into (multiple servers, load balancers, I’m not dwelling into that).
Assuming a maximum rate of 150 req/s, with a KeepAliveTimeout at the 5 seconds default, one may need to adjust MaxClients (and ServerLimit if prefork mpm is used) to somewhere in the area of 1,000.
Don’t leave the defaults on; always start the parameter calculation from the desired outcome first. On suboptimal hardware (memory wise) disabling KeepAlive is the way to go to squeeze a bit more performance. Oh, and don’t assume the hardware is capable of keeping up with your calculations; turn any failure or mis-planning in a learning subject.
That’s it for today, have fun!