Description of problem: We found while trying to scale the summit presentation that the head haproxy gear is hitting the nofile ulimit (which is set to 4096 by default). We added these lines to this file /etc/security/limits.d/84-51ae792fe0b8cdddf2000177.conf: 51ae792fe0b8cdddf2000177 soft nofile 16384 51ae792fe0b8cdddf2000177 hard nofile 16384 And that helped out tremendously. Version-Release number of selected component (if applicable): rhc-node-1.9.15-1.el6oso.x86_64 How reproducible: very Steps to Reproduce: 1. create a scaled app with at least 4 backed gears (5 gears total) 2. pound on the app 3. watch netstat -nap on the head gear's node and notice that the connections are being limited to the head gear. 4. bump the limit 5. notice that the head gear can scale better Actual results: Slow head gear because of file descriptors are being limited. This makes the backend gears not be able to contribute enough. Expected results: fast connections
Only the hard limit needs to be increased. Haproxy will request an increase up to the hard limit based on the number of connections its configured for. Release ticket updated. Pull request for devenv: https://github.com/openshift/li/pull/1618 To Q/A: 1. create a scaled app. 2. log into your gear and edit haproxy/conf/haproxy.cfg 3. change the value of "maxconn" to 8000 4. restart your app 5. log back into the gear and inspect haproxy/logs/haproxy.log Expected: No errors like "FD limit (1024) too low for maxconn=8000..."
Commit pushed to master at https://github.com/openshift/li https://github.com/openshift/li/commit/d4b70a6a4a31fd33ac0d25a76d9d42630bd16bf3 Bug 971610 - Increase the hard limit on file descriptors to 16k
Checked on devenv_3360, Step as comment#1, After restart app, no such errors in haproxy.log Move bug to verified.