Hi Lars,
>> The network is congesting and some time (not regular frequency) some
packet
>> are loose, if a VRRP MASTER advert is dropped, the remote BACKUPs said :
>> "Wow our master is down" => then a new master election is done....
>
> Pardon my ignorance, but do you do this on a single dropped packet?
Wouldn't 3
> packets or so be a better check?
For example we have 2 routers (R1, R2). R1 in MASTER state & R2 in BACKUP.
So R2 is listening to the VRRP mcast group for remote MASTER R1 advert. The
VRRP state machine (and RFC protocol) consider remote MASTER down if don t
receive remote MASTER advert within 3s (the max takeover delay). If we
implement a "retry receiving MASTER advert" for exemple "wait 3 secs more"
we will increase the takeover delay by 2 setting the max default takeover
delay to (3+3)s. This can be acceptable, IMHO, if we are into a heavy
loaded env, but on a normal env this feature should be disabled. At this
time, we need a event that enable/disable this "retry receiving", that way
takeover time will not be affected in a normal loaded env and will be
delayed on heavy loaded env. At this time I would know the best way to
handle this event: If we are in a congestion env then we launch the
ratelimit estimator to avoid the protocol flapping issue. Otherwise we
function in a normal mode. So need to find an indicator to say : "We are
running on a heavy loaded network".
Best regards,
Alexandre
|