[WIP/RFC v4 0/5] Support abstract Unix sockets

[Alejandro Colomar]

Hi Maxim!

On 8/24/22 18:00, Maxim Dounin wrote:
 > I would rather say "it's a common misconception that it's faster
 > to communicate through Unix sockets".  While it Unix sockets can
 > be beneficial for some microbenchmarks, in most production setups
 > it makes no difference, yet used to introduce various issues.


On 8/25/22 01:14, Alejandro Colomar wrote:
> Anyway, if you're not convinced by the feature, let me come back to it 
> after I do some performance testing in Unit to see how significant are 
> the numbers.

I got some numbers (not with nginx, but with unit; if you want me to try 
with nginx, I'll do).  I think they are relevant enough to come revive 
this discussion.

I set up the following:

Unit, running two apps (written in C), one is the backend, and one is 
the frontend.  The user curl(1)s the frontend, which itself calls the 
backend (through either localhost or an abstract UDS), then responds to 
the user.

The exact code of the apps is  here:
<http://www.alejandro-colomar.es/src/alx/alx/nginx/unit-c-app.git/>

I used the 'main' branch for the backend, which is a simple Hello-world 
app.  The 'front' branch is a modified version of the backend which 
calls 1000 times in a loop to the backend and reads the first few bytes 
of those requests, and then responds to the user.  The 'ip' branch is 
the same as the 'front' one, but calling the backend through localhost.

The unit config file is here:

$ cat /opt/local/unit/state/conf.json
{
	"listeners": {
		"*:8080": {
			"pass": "applications/back"
		},
		"unix:@back": {
			"pass": "applications/back"
		},
		"unix:@front": {
			"pass": "applications/front"
		}
	},
	"applications": {
		"back": {
			"type": "external",
			"executable": "/opt/local/unit/libexec/app_back"
		},
		"front": {
			"type": "external",
			"executable": "/opt/local/unit/libexec/app_front"
		}
	}
}


The times, are as follows (I only copied a sample of each, but all of 
the tests that I run --around 10 each--, only varied +/- 2 seconds):

IPv4:

$ time curl --output - --abstract-unix-sock front localhost/ >/dev/null 2>&1

real	0m0.122s
user	0m0.006s
sys	0m0.000s


UDS:

$ time curl --output - --abstract-unix-sock front localhost/ >/dev/null 2>&1

real	0m0.078s
user	0m0.006s
sys	0m0.000s


I think these numbers are enough to make some users go for UDS instead 
of localhost if it's possible.  Also, I think my test is mostly 
measuring the latency of the sockets, but throughput should be even more 
different between UDS and localhost, from what I've read.  Of course, if 
the app is slow, the socket might not be the bottleneck, but from what 
I've tested, unit is fast enough so that socket speed is still 
measurable, so fast apps may legitimately want UDS to pair with them.


Cheers,

Alex

-- 
Alejandro Colomar
<http://www.alejandro-colomar.es/>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: OpenPGP_signature
Type: application/pgp-signature
Size: 833 bytes
Desc: OpenPGP digital signature
URL: <http://mailman.nginx.org/pipermail/nginx-devel/attachments/20220902/fec98b54/attachment.bin>