granularity/balancing issue with a small number of client IPs

[Stefan Schmidt]

Hi there,
i have a problem loadbalancing DNS udp/53 via LVS DR.
Whenever i have a small number of clients, in the example below it is DNS
realtime blackhole lists used by 10 mailservers.
Redundancy is working fine but granularity seems to be an issue and there is
times when i end up with something like this:
#ipvsadm --list --rate --numeric
UDP  1.2.3.4:53                          0     1889        0   173253        0
  -> 10.1.53.15:53                       0        0        0       18        0
  -> 10.1.53.9:53                        0     1889        0   173235        0

configuration looks like this:
virtual_server 1.2.3.4 53 {
        delay_loop 10
        lb_algo wrr
        lb_kind DR
        protocol UDP
        ! persistence_timeout 1
        ! persistence_granularity 255.255.255.255
	real_server 10.1.53.9 53 {
	...

Same picture if i uncomment the persistence statements.

I noticed there apparently are some slewing algorythms doing things like a slow start whenever i add a realserver but i never stumbled over any documentation,
can you give me a hint on where to look for this?
I would want it to be a little less 'soft'.

best regards,
		Stefan
-- 
panic("Oh boy, that early out of memory?");
2.2.16 /usr/src/linux/arch/mips/mm/init.c