Overview of Reliable Multicast Protocols

The table below is not meant to be exhaustive but rather to set the context and framework for presentations and ensuing discussion:

protocol (authors) reliability semantics session? Participant structure knowledge of participants? ACK/NAK

SRM (LBL et al.) reliable no local recovery groups via wb session msgs NAK, receiver reliable

rmtp (Bell Labs) reliable yes, with late join hierarchy of regions, DR's optional, may be known window of pkts ACK/NAK'ed

MTP-2 (TU Bremen) also RCF1301 reliable, totally ordered atomic deliver yes, w/ explicit join master, producer, consumer known NAK(?)

RAMP (TASC) RFC1458 reliable yes w/ late join none known NAK w/ selective ACK (sender and rcvr reliable

TMTP (Kentucky) reliable no tree-based, via ring search no restricted NAK's, NAK suppression, periodic ACKs

log-based (stanford) reliable no logging hierarchy estimated NAK w/ statistical ACK

LORAX (UCSC) stable, reliable no(?) tree-based parent/child hierarchical ACKs

sce (Ga Tech) reliable, weak ordering yes, no late join no known ACK, sender-reliable

HORUS (cornell) reliable up to total ordering membership microprotocol tree possible with PARCLD via membership NAK at lowest level

RMP (NASA
/UC Berkeley)
reliable up to totally ordered explicit membership peers, full members, clients via membership ACK w/ NAK, fully stable

NTE (M. Handley) reliable no local recovery via session pkts triggered NAKs with randomization

Some comments made during the discussion at workshop included:

need to include something in above taxonomy about flow control
need to include something in above taxonomy about stability (who gets to release packet) and the issue of notification (who knows when something is delivered correctly)
are re-tranmissions multicast or unicast (either, both) in the above and who controls this decision?

Observations:

many protocols proposed and implemented
essentially two classes of reliability semantics
- 100% data reliable with or without …..
- ordering and causality
protocols differ widely in design
- logical structure of communication pathways (ring versus tree versus none)
- group membership mechanisms and assumptions
- receiver-reliable versus sender reliable
- ACK/NAK
and the application for which they were designed

Up to main page for reliable multicast session.

protocol (authors)	reliability semantics	session?	Participant structure	knowledge of participants?	ACK/NAK
SRM (LBL et al.)	reliable	no	local recovery groups	via wb session msgs	NAK, receiver reliable
rmtp (Bell Labs)	reliable	yes, with late join	hierarchy of regions, DR's	optional, may be known	window of pkts ACK/NAK'ed
MTP-2 (TU Bremen) also RCF1301	reliable, totally ordered atomic deliver	yes, w/ explicit join	master, producer, consumer	known	NAK(?)
RAMP (TASC) RFC1458	reliable	yes w/ late join	none	known	NAK w/ selective ACK (sender and rcvr reliable
TMTP (Kentucky)	reliable	no	tree-based, via ring search	no	restricted NAK's, NAK suppression, periodic ACKs
log-based (stanford)	reliable	no	logging hierarchy	estimated	NAK w/ statistical ACK
LORAX (UCSC)	stable, reliable	no(?)	tree-based	parent/child	hierarchical ACKs
sce (Ga Tech)	reliable, weak ordering	yes, no late join	no	known	ACK, sender-reliable
HORUS (cornell)	reliable up to total ordering	membership microprotocol	tree possible with PARCLD	via membership	NAK at lowest level
RMP (NASA /UC Berkeley)	reliable up to totally ordered	explicit membership	peers, full members, clients	via membership	ACK w/ NAK, fully stable
NTE (M. Handley)	reliable	no	local recovery	via session pkts	triggered NAKs with randomization