[Spread-users] membership algorithm and token loss

Steven Dake scd at broked.org
Sat Jan 24 14:29:32 EST 2004


After reading Yair's PHD thesis and some of the spread code, I have a
question about token loss.

It appears on my test networks, a UDP packet is dropped every 1 million
- 20 million packets.

Does this result in the loss of the token in Spread, which causes the
membership algorithm to partition the network, and start the process of
remerging the split rings?

I would like to avoid this costly membership transition as my
distributed application has costly processing when membership
changes...  It is best if this could be avoided...

In some test software, I have found a p2p ACK can recover from
unreliable packet loss of the token.

Was an ACK with retries considered when moving the token from one ring
member to the next?  The cost of the ACK is minimal, loss of the ACK is
detectable because multiple resends of the token can be detected and
managed, and only requires a timer.  When the ACK timer expires and no
ACK is received, and the token is resent a configurable number of
retries with no ACK, the membership algorithm could then be started.

Are there any negatives to this approach?


More information about the Spread-users mailing list