All things related to multicast

[Phil Pishioneri <>]

We've noticed an interesting (and to us, undesirable) behavior of a 
couple of different vendor's switches that implement IGMP snooping, and 
were wondering what folks thought about it.

Background: multicast source within our domain, PIM sparse mode.  Source 
is on a different router (differnet subnet would probably be sufficient) 
from that used by the receiver.

Scenario:

  1. Host H is receiving multicast group G via router R (having sent an
     IGMP report earlier).  It is the only host on the subnet joined to
     that group. (group G is -not- one of the groups which
     unfortunately map into the "flood all data" addresses.)
  2. Host H wishes to leave group G and sends a Leave Group message.
  3. Router R sends the first Group Specific query for  group G, waits
     for timeout (1 second).
  4. Router R sends the second Group Specific query for  group G, waits
     for timeout (1 second).
  5. Router R stops forwarding traffic for group G to the subnet of host H.

This all seems reasonable -- it's what the vLAN's switch (single switch, 
vLAN only leaves the switch on a trunk back to the router) does which 
isn't good.

When the switch snoops the Leave Group message in step 2, it removes 
that port from its forwarding table for that multicast group.  Since 
that was the last (only) port receiving traffic for that group, it 
removes all IGMP knowledge of that group.

While the router is waiting to hear membership reports,  traffic for 
group G is still being forwarded to the subnet.  Since the switch does 
not have any IGMP state for the group, it defaults to flooding the 
traffic to _all_ ports of the vLAN for the two seconds it takes for the 
router's queries to time out  (we haven't checked to see if the traffic 
leaks out of the vLAN yet).  Not good for some of the hosts on that 
subnet.  (e.g., 5+ megabit streams, 10 megabit connections: host A 
subscribed to group Ga, host B subscribed to group Gb.  Host B leaves 
group Gb, host A's reception of group Ga is impaired for a couple of 
seconds.)

Would it be reasonable for a switch to be a little smarter about how it 
handles this situation?  Either a global setting that says "always use 
IGMP snooping for (appropriate) groups on this vLAN (perhaps only if a 
querier or any IGMP messages are discovered)" or setting a timer to keep 
the IGMP state around (timer based on last subscriber, or of last IGMP 
message received for that group), or something else?

-Phil
Cornell University