Category CentOS

If you test Galera synchronous replication with Percona XtraDB Cluster or MariaDB Galera Cluster you must have tried to use a load balancer like HA Proxy or Galera Load Balancer. On very heavy load, you may have issue with a large amount of TCP in TIME_WAIT like this one:

tcp        0      0 127.0.0.1:59035         127.0.0.1:3306          TIME_WAIT

This can lead to a TCP port exhaustion as explained on this post. On HA proxy since version 1.4.19, you can use the nolinger option also on TCP backends. This terminate the connection (TCP RST) as soon as the loadbalancer finished the communication. The counter part is that Aborted_clients status counter in MySQL increases with each connections' end. This counter becomes then useless. This option is not available on glb (with -l parameter) if you apply the patch attached to this post. I provide also rpm package with the patch applied :

Name        : glb
Version     : 0.9.2
Release     : 2
Architecture: x86_64
Install Date: (not installed)
Group       : Productivity/Networking/Routing
Size        : 208489
License     : GNU General Public License version 2 or later (GPL v2 or later)
Signature   : (none)
Source RPM  : glb-0.9.2-2.src.rpm
Build Date  : mer 27 fév 2013 17:15:54 CET
Build Host  : percona1
Relocations : (not relocatable)
URL         : http://www.codership.com/products/galera-load-balancer
Summary     : TCP Connection Balancer
Description :
glb is a simple user-space TCP connection balancer made with scalability and
performance in mind. It was inspired by pen, but unlike pen its functionality
is limited only to balancing generic TCP connections.

Features:
* list of backend servers is configurable in runtime.
* supports server "draining", i.e. does not allocate new connections to server,
  but does not kill existing ones, waiting for them to end gracefully.
* on Linux 2.6 and higher glb uses epoll API for ultimate performance.
* glb is multithreaded, so it can utilize multiple CPU cores. In fact even on a
  single core CPU using several threads can significantly improve performance
  when using poll()-based IO.
* connections are distributed proportionally to weights assigned to backend
  servers.
* this is a patched version providing SO_LINGER

Example:

[root@macbookair ~]# glbd -K -l --threads 6 --control 127.0.0.1:4444 127.0.0.1:3308 127.0.0.1:3306
glb v0.9.2 (epoll)
Incoming address:       127.0.0.1:3308 , control FIFO: /tmp/glbd.fifo
Control  address:        127.0.0.1:4444 
Number of threads: 6, max conn: 493, policy: 'least connected', top: NO, nodelay:
 ON, keepalive: OFF, defer accept: OFF, verbose: OFF, linger: ON, daemon: NO
Destinations: 1
   0:       127.0.0.1:3306 , w: 1.000
Router:
------------------------------------------------------
        Address       :   weight   usage    map  conns
      127.0.0.1:3306  :    1.000   0.000    N/A      0
------------------------------------------------------
Destinations: 1, total connections: 0 of 493 max

Pool: connections per thread:     0     0     0     0     0     0

If you test it please post a comment.

Codership released a new version of the load balancer for Galera. I made new rpms but I forgot to share them ;) Here they are ! This new version provides ] a "single" balancing policy where all connections are directed to a single destination chosen by highest weight, a --top option that forces balancing only between the destinations with the highest weight, and a SO_KEEPALIVE option on destination connections (default: on) for timely detection of the destination failure.