kutata · May 20, 2019 08:12
diff --git a/md b/md


 In short, you're running out of ports.

 The default ephemeral port range on osx is 49152-65535, which is only 16,383 ports. 
 Since each ab request is http/1.0 (without keepalive in your first examples), each
 new request takes another port.

 As each port is used, it get's put into a queue where it waits for the tcp 
 "Maximum Segment Lifetime", which is configured to be 15 seconds on osx. So if you 
 use more than 16,383 ports in 15 seconds, you're effectively going to get throttled 
 by the OS on further connections. Depending on which process runs out of ports first, 
 you will get connection errors from the server, or hangs from ab.

 You can mitigate this by using an http/1.1 capable load generator like wrk, or using 
 the keepalive (-k) option for ab, so that connections are reused based on the tool's 
 concurrency settings.

 Now, the server code you're benchmarking does so little, that the load generator is
 being taxed just as much as the sever itself, with the local os and network stack 
 likely making a good contribution. If you want to benchmark an http server, it's better 
 to do some meaningful work from multiple clients not running on the same machine.

 -- from stackoverflow JimB's answer 
 https://stackoverflow.com/questions/30352725/why-is-my-hello-world-go-server-getting-crushed-by-apachebench


	In short, you're running out of ports.

	The default ephemeral port range on osx is 49152-65535, which is only 16,383 ports.
	Since each ab request is http/1.0 (without keepalive in your first examples), each
	new request takes another port.

	As each port is used, it get's put into a queue where it waits for the tcp
	"Maximum Segment Lifetime", which is configured to be 15 seconds on osx. So if you
	use more than 16,383 ports in 15 seconds, you're effectively going to get throttled
	by the OS on further connections. Depending on which process runs out of ports first,
	you will get connection errors from the server, or hangs from ab.

	You can mitigate this by using an http/1.1 capable load generator like wrk, or using
	the keepalive (-k) option for ab, so that connections are reused based on the tool's
	concurrency settings.

	Now, the server code you're benchmarking does so little, that the load generator is
	being taxed just as much as the sever itself, with the local os and network stack
	likely making a good contribution. If you want to benchmark an http server, it's better
	to do some meaningful work from multiple clients not running on the same machine.

	-- from stackoverflow JimB's answer
	https://stackoverflow.com/questions/30352725/why-is-my-hello-world-go-server-getting-crushed-by-apachebench