Opened 8 years ago

Closed 8 years ago

Last modified 8 years ago

#1210 closed defect (wontfix)

connections not distributed evenly after restart under high load with accept_mutex enabled

Reported by: gallatin@… Owned by:
Priority: minor Milestone:
Component: other Version: 1.10.x
Keywords: Cc:
uname -a: FreeBSD c099.ord001.dev.ix.nflxvideo.net 12.0-CURRENT FreeBSD 12.0-CURRENT #0: Fri Feb 24 13:40:24 EST 2017 root@viserion:/data/NCD-3545/.obj/nano.nflx/data/NCD-3545/FreeBSD/sys/RED1 amd64
nginx -V: nginx version: nginx/1.10.0
built with OpenSSL 1.0.2k-freebsd 26 Jan 2017
TLS SNI support enabled
configure arguments: --prefix=/usr/local/etc/nginx --with-cc-opt='-I /usr/local/include' --with-ld-opt='-L /usr/local/lib' --conf-path=/usr/local/etc/nginx/nginx.conf --sbin-path=/usr/local/sbin/nginx --pid-path=/var/run/nginx.pid --error-log-path=/var/log/nginx-error.log --user=www --group=www --add-module=/var/tmp/portswork/_.p/netflix/nginx-1.10/work/nginx-1.10.0/nflx_smooth --add-module=/var/tmp/portswork/_.p/netflix/nginx-1.10/work/nginx-1.10.0/nflx_geo_check --add-module=/var/tmp/portswork/_.p/netflix/nginx-1.10/work/nginx-1.10.0/nflx_signed_uri --add-module=/var/tmp/portswork/_.p/netflix/nginx-1.10/work/nginx-1.10.0/nflx_range_path --add-module=/var/tmp/portswork/_.p/netflix/nginx-1.10/work/nginx-1.10.0/nflx_variables --add-module=/var/tmp/portswork/_.p/netflix/nginx-1.10/work/nginx-1.10.0/nflx_nccp_conf --add-module=/var/tmp/portswork/_.p/netflix/nginx-1.10/work/nginx-1.10.0/nflx_logstat --add-module=/var/tmp/portswork/_.p/netflix/nginx-1.10/work/nginx-1.10.0/nflx_loghot --add-module=/var/tmp/portswork/_.p/netflix/nginx-1.10/work/nginx-1.10.0/nflx_status --add-module=/var/tmp/portswork/_.p/netflix/nginx-1.10/work/nginx-1.10.0/nflx_upstream_jdomain --add-module=/var/tmp/portswork/_.p/netflix/nginx-1.10/work/nginx-1.10.0/nflx_rorder --add-module=/var/tmp/portswork/_.p/netflix/nginx-1.10/work/nginx-1.10.0/nflx_delay --add-module=/var/tmp/portswork/_.p/netflix/nginx-1.10/work/nginx-1.10.0/nflx_sendfile_flags --with-debug --with-file-aio --with-ipv6 --http-client-body-temp-path=/var/tmp/nginx/client_body_temp --http-fastcgi-temp-path=/var/tmp/nginx/fastcgi_temp --http-proxy-temp-path=/var/tmp/nginx/proxy_temp --http-scgi-temp-path=/var/tmp/nginx/scgi_temp --http-uwsgi-temp-path=/var/tmp/nginx/uwsgi_temp --http-log-path=/var/log/nginx-access.log --with-http_v2_module --with-pcre --with-http_ssl_module --with-threads

Description

On a heavily loaded server (90,000 long lived connections) restarting nginx results in a few nginx workers with the connection limit, and others with very few connections.

conf file snippets:

worker_processes     64;
worker_rlimit_nofile 32768;
events {
    worker_connections  4096;

    accept_mutex on;

}

Before service nginx restart, server is serving 90+ Gb/s across ~90,000 conenctions:

            input        (Total)           output
   packets  errs idrops      bytes    packets  errs      bytes colls
   2237651     0     0  167011176    7789342     0 11633619064     0
   2214132     0     0  165215742    7719175     0 11531815268     0
   2255186     0     0  168221493    7816878     0 11679556820     0
<...>

nginx restarted. Now top shows uneven load in new nginx workers:

last pid:  4557;  load averages: 16.89, 19.71, 17.60                    up 0+00:52:43  20:00:44
227 processes: 5 running, 222 sleeping
CPU:  4.8% user,  0.0% nice, 10.8% system, 21.1% interrupt, 63.2% idle
Mem: 2547M Active, 91G Inact, 3260K Laundry, 27G Wired, 1585M Buf, 2710M Free
Swap: 32G Total, 79M Used, 32G Free

  PID USERNAME    THR PRI NICE   SIZE    RES STATE   C   TIME    WCPU COMMAND
 4364 www           1  48   -4   270M   177M CPU18  18   0:50  40.25% nginx
 4371 www           1  39   -4   246M   183M CPU4    4   0:56  33.52% nginx
 4347 www           1  35   -4   236M   170M kqread  4   0:54  26.62% nginx
 4398 www           1  32   -4   230M   167M kqread 18   0:55  26.00% nginx
 4391 www           1  30   -4   216M   151M kqread 15   0:53  24.04% nginx
 4381 www           1  31   -4   214M   150M kqread  7   0:50  23.07% nginx
 4372 www           1  27   -4   220M   157M kqread 20   0:38  22.09% nginx
 4346 www           1  26   -4   228M   166M kqread  5   0:37  19.00% nginx
 4380 www           1  26   -4   232M   158M CPU26  26   0:48  18.59% nginx
 4363 www           1  25   -4   206M   144M kqread 20   0:31  17.16% nginx
 4396 www           1  26   -4   210M   147M kqread 14   0:34  16.60% nginx
 4402 www           1  24   -4   186M   123M kqread  3   0:27  14.66% nginx
 4345 www           1  24   -4   198M   136M kqread  6   0:30  14.58% nginx
 4349 www           1  24   -4   200M   137M kqread  6   0:31  14.30% nginx
 4365 www           1  20   -4   152M 91584K kqread  5   0:14  10.41% nginx
 4355 www           1  20   -4   138M 76796K kqread  7   0:12   7.28% nginx
 4351 www           1  20   -4   146M 84196K kqread 15   0:14   6.81% nginx
 4384 www           1  20   -4   120M 56260K kqread 21   0:06   4.76% nginx
 4400 www           1  20   -4   126M 63348K CPU17  17   0:08   4.49% nginx
 4383 www           1  20   -4   116M 53096K kqread  9   0:04   3.80% nginx
 4376 www           1  20   -4   116M 52824K kqread 20   0:04   3.76% nginx
 4397 www           1  20   -4   118M 53928K kqread 22   0:05   3.45% nginx
 4367 www           1  20   -4   112M 46252K kqread  9   0:03   3.32% nginx
 4368 www           1  20   -4   110M 43836K kqread  8   0:02   3.30% nginx
 4352 www           1  20   -4   112M 48280K kqread 22   0:03   2.84% nginx
 4373 www           1  20   -4   110M 42596K kqread 28   0:02   2.33% nginx
 4392 www           1  20   -4   112M 45620K kqread  5   0:03   2.14% nginx
 4377 www           1  20   -4   104M 31236K kqread 22   0:01   1.85% nginx
 4379 www           1  20   -4   108M 39072K kqread 30   0:02   1.68% nginx
 4361 www           1  20   -4   106M 36536K kqread 30   0:01   1.65% nginx
 4369 www           1  20   -4   108M 41188K kqread 20   0:02   1.63% nginx
 4388 www           1  20   -4   106M 36644K kqread  7   0:01   1.52% nginx
 4405 www           1  20   -4   106M 35744K kqread 13   0:01   1.32% nginx
 4395 www           1  20   -4   106M 37380K kqread 17   0:02   1.31% nginx
 4378 www           1  20   -4   104M 34352K kqread 13   0:01   1.28% nginx
 4407 www           1  20   -4   106M 39596K kqread  1   0:02   1.18% nginx
 4354 www           1  20   -4   106M 37312K kqread 29   0:02   1.17% nginx
 4382 www           1  20   -4   104M 31644K kqread  4   0:01   1.17% nginx
 4350 www           1  20   -4   104M 33508K kqread  8   0:01   0.94% nginx
 4401 www           1  20   -4   104M 30320K kqread 14   0:01   0.93% nginx
 4406 www           1  20   -4   104M 32572K kqread 12   0:01   0.92% nginx
 4408 www           1  20   -4   104M 31300K kqread 10   0:01   0.85% nginx
 4360 www           1  20   -4   106M 35236K kqread  1   0:01   0.72% nginx
 4362 www           1  20   -4   104M 32712K kqread 26   0:01   0.68% nginx
 4366 www           1  20   -4   102M 28800K kqread  3   0:00   0.65% nginx
 4353 www           1  20   -4   104M 30888K kqread 28   0:01   0.58% nginx
 4375 www           1  20   -4   104M 32060K kqread 18   0:01   0.57% nginx
 4403 www           1  20   -4   104M 29636K kqread  6   0:01   0.52% nginx
  568 root          1  20    0 10496K  2588K select 17   0:09   0.45% syslogd
 4385 www           1  20   -4   104M 29536K kqread  5   0:01   0.45% nginx
 4404 www           1  20   -4   102M 26208K kqread 22   0:00   0.44% nginx
 4387 www           1  20   -4   102M 22384K kqread  6   0:00   0.41% nginx
 4370 www           1  20   -4   102M 22328K kqread 10   0:00   0.41% nginx
 4399 www           1  20   -4   102M 23440K kqread 10   0:00   0.40% nginx
 4358 www           1  20   -4   102M 26216K kqread 12   0:00   0.38% nginx
 4390 www           1  20   -4   102M 26612K kqread 17   0:00   0.34% nginx
 4356 www           1  20   -4   102M 24788K kqread  7   0:00   0.33% nginx

Most popular workers have near connection limit, least popular has almost no connections:

c099.ord001.dev# lsof -n -p 4364 | grep TCP | wc -l
    4058
c099.ord001.dev# lsof -n -p 4380 | grep TCP | wc -l
    4058
c099.ord001.dev# lsof -n -p 4389 | grep TCP | wc -l
      77
c099.ord001.dev# ps augxw | grep 4389
www     4389    0.0  0.0 104392  27436  -  S<   19:57      0:00.60 nginx: worker process (nginx)
root    4618    0.0  0.0  14812   2484  0  S+   20:01      0:00.00 grep 4389

The uneven loading seems to persist indefinitely. Eg, the connection count never seems to even out.

As a workaround, disabling the accept mutex helps, but causes other problems (kernel lock contention in kqueue).

Change History (2)

comment:1 by Maxim Dounin, 8 years ago

Resolution: wontfix
Status: newclosed

With accept mutex enabled, distribution of connections is expected to be uneven as long as processes are not fully loaded. This is the result of how accept mutex work.

If you want to improve distribution of connections between worker processes, consider the following:

  • use smaller accept_mutex_delay;
  • use smaller number of worker processes;
  • switch accept_mutex off.

comment:2 by gallatin@…, 8 years ago

Thanks; I had not realized that.

Reducing the accept_mutex_delay does indeed seem to mitigate the problem.

Thank you again!

Note: See TracTickets for help on using tickets.