[jboss-jira] [JBoss JIRA] Commented: (JGRP-1243) FD_SOCK: reduce number of messages sent on a suspicion

FD_SOCK: reduce number of messages sent on a suspicion ------------------------------------------------------ Key: JGRP-1243 URL: https://jira.jboss.org/browse/JGRP-1243 Project: JGroups Issue Type: Feature Request Reporter: Bela Ban Assignee: Bela Ban Fix For: 2.11 - When B suspects C, B multicasts a SUSPECT(C) message - Everyone receives the SUSPECT(C) message and passes it up and down the stack as a SUSPECT(C) event - VERIFY_SUSPECT *on every member* sends one (or more) ARE_YOU_DEAD messages to C - C replies to the sender with a I_AM_NOT_DEAD messages, or not if crashed - However, only the coordinator (or next in line) actually processes the SUSPECT(C) event in GMS ! --> All of the VERIFY_SUSPECT processing is superfluous unless it is the coord or next-in-line ! The number of messages used for a false suspicion are (1 SUSPECT mcast) + ((N-1) ARE_YOU_DEAD unicasts) + ((N-1) I_AM_NOT_DEAD unicasts)) ! SOLUTION: - The SUSPECT(C) message could be sent as a unicast only to the coordinator and the next-in-line member. Maybe we could use a max_rank=2 for this, similar to the suggested solution for FD_ALL ? This would be good for non multicast based transports, e.g. TCP - The SUSPECT(C) message is multicast to everyone, but only the coord and next-in-line start the VERIFY_SUSPECT processing Issue: if we have {A,B,C,D,E}, what happens if A,B and C crash at the same time ? - E's connection to A closes: E sends a SUSPECT(A) to B and C (excluding suspected A) --> B and C are dead and won't process the message ! - Then E suspects B and sends a SUSPECT(A,B) to C and D (excluding suspected A and B) - C adds A and B to its suspect list and finds out it is the next-in-line - C then runs the VERIFY_SUSPECT protocol - C passes the SUSPECT(A,B) event up the stack - C becomes the new coord