|
Hey there,
we have a problem with dropped tcp connection requests.
The setup looks like this, using a classical LAMP but with an ipvs direct
routing based loadbalancer for the mysql replication cluster:
[ipvs database loadbalancer]
|
|
|
[=================switched network==================]
| | | | | |
| | | | | |
| | | | | |
[web 1] [web 2] ... [web n] [db 1] [db 2] ... [db n]
This setup works perfectly most of the time, but sometimes the database
servers drops tcp connection requests
originating from the web servers. The tcp syn packets went from the
webserver through the ipvs machine, the mac adresses are replaced there
and the packet arrived at one of the database servers (it's visible in
tcpdump at the database server and looks correct). But the database server
is ignoring it, not sending a syn/ack, tcp rst packet or something else.
Also neither the kernel nor the database is logging anything (eg. tcp hash
table overflow or something you'd expect in such cases).
This does not happen when the webservers connect directly to
the database servers without using the ipvs loadbalancer but using
application based load balancing on the webservers. Because of the we think
that it could be a problem related to ipvs and maybe somebody at this list is
able to give us a hint what to do.
Thanks in advance
Jan Abraham
|
