nginx-1.9.12 performance on DragonFly-g680c4d.

Sepherosa Ziehau sepherosa at gmail.com
Thu May 26 07:08:45 PDT 2016


nginx-1.9.12 performance on DragonFly-g680c4d.

Since the client machine is the bottleneck (all cpu time is used
up on client side) to measure further improvement, one more client
machine is added; they have same hardware configuration.

New network configure:
+--------+                           +--------+
|        |192.168.3.254   192.168.3.1|        |
| Server +---------------------------+ Client |
|        |10Ge        DAC        10Ge|        |
+---+----+                           +--------+
10Ge|192.168.4.254
    |
    |
 DAC|
    |
    |
10Ge|192.168.4.1
+---+----+
|        |
| Client |
|        |
+--------+

Each client generates 15K concurrent connections, so total 30K
concurrent connections.

Other configurations are same as:
http://lists.dragonflybsd.org/pipermail/users/2016-May/249581.html

                 1KB object
             setsockopt optimize
               before | after
             ---------+---------
 1  req/conn   156337 | 186442  (19% improvement)
 4 reqs/conn   221003 | 232046
14 reqs/conn   243643 | 247083

                 8KB object
             setsockopt optimize
               before | after
             ---------+---------
 1  req/conn   137622 | 162765  (18% improvement)
 4 reqs/conn   189352 | 200795
14 reqs/conn   208292 | 211986

                 16KB object
             setsockopt optimize
               before | after
             ---------+---------
 1  req/conn   121801 | 136605  (12% improvement)
 4 reqs/conn   139164 | 139093
14 reqs/conn   138843 | 138623

And for 16KB object, we now max out 2x10Ge (19.2Gbps, was 16.8Gbps
before 680c4dd)

This will serve as the baseline for further network optimization.

Best Regards,
sephe

-- 
Tomorrow Will Never Die


More information about the Users mailing list