--- Christian Lyra <[EMAIL PROTECTED]> wrote:
> ok... broadcast was bad choice of word. I meant, when a edge node
> register to supernode what happens next? My guess was that supernode
> "tells" everyone about the new edge, is this correct?
No. Read on.
> Now about the broadcast... when a node send a ARP who-has, who replies
> to sender? the edge node who has the awnser or the supernode awnser,
> since he knows who has?
CURRENT BEHAVIOUR
Say A sends who-has B.
B replies to A via S. But B will also immediately try to register with A
because it received a packet from A. Unfortunately all edges receive the ARP
broadcast and will try to register with A.
On receiving the ARP reply A registers with B to allow direct packets from A to
B. Until the registration completes, A sends to B via supernode.
The issue for me is that the ARP who-has request reaches all nodes; and their
response to any incoming edge packet is to register with the sender. So this
triggers a flurry of activity.
The supernode cannot see the contents of the packets as it is not part of any
community so it cannot be smart. This is good.
Fortunately the supernode does not broadcast globally; only to edges in the
same community as the originating broadcast request. The problem remains that
as the size of a community increases the packets required to add each new edge
increases linearly.
PROPOSED CHANGE
The edge nodes can see inside the packets. One possible change would be to not
initiate a registration when receiving a packet with the broadcast MAC address
- only when directed to the edge's individual MAC address.
If this change was made the behaviour would be:
A sends who-has.
All nodes receive who-has; B responds via S. B does not register with A, nor do
any other edges register with A because dest_mac=FF:FF:FF:FF:FF:FF.
A registers with B. B then registers with A because it received a packet to its
individual MAC address.
Under this model edges do not form a fully-connect mesh; rather only connected
by the requirement to exchange packets between individual members. This would
have several benefits:
* Edges would have fewer registrations -> more efficient.
* System more immune to DoS by L2 broadcasts.
* Packet cost for adding a new node is O(1) instead of O(N) for community of
size O(N).
I cannot see a disadvantage yet.
--
Rich
Get the name you always wanted with the new y7mail email address.
www.yahoo7.com.au/y7mail
_______________________________________________
Ntop-dev mailing list
[email protected]
http://listgateway.unipi.it/mailman/listinfo/ntop-dev
