SO_BUSY_POLL for latency critical cases

Maxim Dounin mdounin at mdounin.ru
Thu Feb 2 12:47:43 UTC 2017


Hello!

On Thu, Feb 02, 2017 at 09:40:56AM +0200, Donatas Abraitis wrote:

> I do not cover any tests, just putting this patchset and wanna ask if
> someone had tried this? I tried this for Redis and changes are really
> reasonably visible like here: https://github.com/antirez/redis/pull/3773

Just a side note:

The data provided in the pull request is not really convincing.  
Ministat shows no difference proven at 95% confidence for both min 
and max latency values:

$ ministat with.min without.min
x with.min
+ without.min
+------------------------------------------------------------------------------+
|x                                  x  x   x+ +          +              x +   +|
|            |________________________AM____|____________M__A__|___________|   |
+------------------------------------------------------------------------------+
    N           Min           Max        Median           Avg        Stddev
x   5        458752        702022        587727      585944.6     86500.993
+   5        605029        722469        649593      659764.4     53243.639
No difference proven at 95.0% confidence
$ ministat with.max without.max
x with.max
+ without.max
+------------------------------------------------------------------------------+
|x                                   + x   xx     +                +   +      *|
|             |__________________________A_M|________________A_____M|________| |
+------------------------------------------------------------------------------+
    N           Min           Max        Median           Avg        Stddev
x   5      26625442 2.0884278e+08 1.2690286e+08 1.2130263e+08      64656095
+   5 1.1113228e+08 2.0951387e+08 1.8428094e+08 1.6806744e+08      40111502
No difference proven at 95.0% confidence

-- 
Maxim Dounin
http://nginx.org/


More information about the nginx mailing list