A question of fastcgi_pass method

[Maxim Dounin]

Hello!

On Sun, Oct 17, 2010 at 06:42:51PM +0800, SanCao Jie wrote:

[...]

> I use webbench to do this test  :  webbench -c 100 http://IP/
> 
> This is the result:
> 
> USE TCP Method:
> 
> Speed=5730 pages/min, 14038 bytes/sec.
> Requests: 2865 susceed, 0 failed.
> 
> USE UNIX SOCKET Method:
> 
> Speed=5548 pages/min, 13592 bytes/sec.
> Requests: 2774 susceed, 0 failed.
> 
> Why the test shows that TCP method is faster than Unix Socket ?

There are no real reasons why tcp over loopback should be slower 
than unix sockets.  It's up to OS implementation details.

In your particular case it looks like the speed is limited by your 
fastcgi app, not by nginx or OS.  The difference (if any, as your 
test results don't include statistical errors) is most likely due 
to different connection queueing with different sockets used by OS.

> And an  other quertion is :during the TCP Method test ,I typed : netstat -an
> | grep 9000  ,this were many ports shows
> TIME WAIT .almost over 3,000 . when the webbench is over ,zhe TIME WAIT
> ports number is decrease.
> Why this ,And is it safe ?

This is per TCP, see RFC 973 for details.

See your OS guides to find out what will happen if all possible 
(src, srcport, dst, dstport) between frontend and backend happen 
to be in TIME-WAIT.  Usually it's good idea to tune portrange and 
check for tw reuse/recycle options if you are going to work with 
high backend connection rates.

Maxim Dounin