6:54 a.m.
Not the way the RPCDispatcher in JGroups works. I always pass in a
List of buddies, and usually this list would have 1 element (if
n=1). Either way, even if n=3 in a 500 node cluster the multicast is
inefficient.
(moving this to the JBC dev list)
--
Manik Surtani
Lead, JBoss Cache
JBoss, a division of Red Hat
Email: manik(a)jboss.org
Telephone: +44 7786 702 706
MSN: manik(a)surtani.org
Yahoo/AIM/Skype: maniksurtani
On 5 Oct 2006, at 11:58, Ben Wang wrote:
I thought for BR, we do unicast now when there is only one buddy
(the default) and switch to multicast where is more than one?
From: Bela Ban
Sent: Thursday, October 05, 2006 2:53 PM
To: Ben Wang; Brian Stansberry; Manik Surtani; Vladimir Blagojevic;
Galder Zamarreno
Subject: Anycast
FYI,
currently, buddy replication generates the *same* amount of traffic
as multicasting to everyone. I created a JIRA issue, anycast, which
when implemented will allow for sending of messages to a subset of
the membership: http://jira.jboss.com/jira/browse/JGRP-338
--
Bela Ban
Lead JGroups / Manager JBoss Clustering Group
JBoss - a division of Red Hat
7:11 a.m.
New subject: Anycast
Manik Surtani wrote:
Not the way the RPCDispatcher in JGroups works. I always pass in a
List of buddies, and usually this list would have 1 element (if n=1).
Either way, even if n=3 in a 500 node cluster the multicast is
inefficient.
Not worse than TCP: in TCP, we send the *same message* 500 times, one
for each node. Everyone discards the message but the 3 who are targetted.
Not a big problem for now, because we usually have small clusters, but
something that need fixing in the future.
How the transport delivers a message M to X members of a cluster of size
N (where X <= N) is left to the transport.
TCP: here, the transport would probably simply send M to X members.
UDP: if X is close to N --> multicast M and everyone whose not in the
target set X simply discards M. If X is small --> unicast to X
--
Bela Ban
Lead JGroups / Manager JBoss Clustering Group
JBoss - a division of Red Hat
10:43 a.m.
New subject: Anycast
On 5 Oct 2006, at 13:11, Bela Ban wrote:
Manik Surtani wrote:
> Not the way the RPCDispatcher in JGroups works. I always pass in
> a List of buddies, and usually this list would have 1 element (if
> n=1). Either way, even if n=3 in a 500 node cluster the multicast
> is inefficient.
Not worse than TCP: in TCP, we send the *same message* 500 times,
one for each node. Everyone discards the message but the 3 who are
targetted.
Not a big problem for now, because we usually have small clusters,
but something that need fixing in the future.
How the transport delivers a message M to X members of a cluster of
size N (where X <= N) is left to the transport.
TCP: here, the transport would probably simply send M to X members.
UDP: if X is close to N --> multicast M and everyone whose not in
the target set X simply discards M. If X is small --> unicast to X
Precisely. Re: UDP, this could even be configurable (i.e., the X:N
ratio above which a MCAST is triggered)
6657
days inactive
6658
days old
jbosscache-dev@lists.jboss.org
2 comments
2 participants
participants (2)
-
Bela Ban -
Manik Surtani