|
Thanks to everyone with suggestions. Karl's email pretty much
summarized the suggestions, so I'll respond to his. In the end,
I ended up not using "persistent" for this VIP, and connections
(even from the same machine using "siege"!) are balanced, now.
Karl Kopper wrote:
If all your connections are from a single client computer you may always get
sent back to the same real server due to the various timeouts (UDP, TCP,
connection tracking records, persistent connection template entries).
Better yet, test from two client computers.
Testing from two clients indeed showed requests going to the other
machine. Success!
You can use ldirectord to remove real servers from the cluster when they no
longer offer services and set:
quiescent=no
in your ldirectord configuration file. If you do this the connection
tracking records and the persistent connection template entries will be
removed (expired) from the Director.
I'm going to fool with "quiescent" a bit more.
You might also want to set "/proc/sys/net/ipv4/vs/expire_nodest_conn" to 1.
Didn't try this.
Make sure the VIP is hidden on the real servers, or you never know who is
going to respond to your ARP broadcast (from the client computer) for the
VIP--sometimes the packet might go through the Director and sometimes not.
Important, but I had already checked this and the interfaces were
definitely hidden.
Thanks again for alll the suggestions,
JZ
| 