[j-nsp] EX4550 (Un-)known unicast flooding at session start for up to 100ms

[Tim]

Hi everybody,

the last half year (since we extend some monitoring/logging things) we
observe multiple times a day strom control triggering across the
network. We could trace it back, that it happen on a regular basis if
heavy stream are initiated (database copy jobs, vmotion stuff, etc.)
it seems that the EX4550 Switches in the data path needs up to 100ms
for mac learning unit he stops the (un-)known unicast flooding.
Because of the heavy nature of the traffic streams we are talking
about 7-8 MB flooded traffic during the 100ms.

The topology is quite simple and straight.

SRV1 - EX4550 - QFX5100 - EX4550 - SRV2

With a sniffing server attached to a normal port on the EX4550 Switch
near SRV1. Normal port means, that it is of course not a mirror port
to only got the flooded stuff.

We've open a case at our service partner, but recieved a at least
disputable answer "Hey, works as designed. Normal behavior like Cisco,
etc.)"

In my personal 15 years experience with Cisco, Extreme/Enterasys
Networks and Nortel Networks i've never see this magnitude of MAC
learning latency / unicast flooding.
If i imagine that this behavior is normal, it would mean that just a
dozen concurrent heavy streams would fuck up all other ports and even
a 40G uplink has no chances and gets heavily overloaded in this time
frame.

What brings me to my question. Does anyone here have any experience
with "typical" MAC learning latencies or similar problems with Juniper
/ $vendor like this?

Best regards,
Tim