Re[2]: [xmlblaster-devel] clustering

[David Kantowitz <davidk at aeonics.com>]

Marcel,

MR> For lookup/discovery i have planned to use multicast as one possibility
MR> as well, i wanted to borrow it from JINI :-)

A Group Communication Protocol can be used for more then just group
membership. Most of the tool kits allow you to construct protocol
stacks to create different message ordering properties. You use these
communication channels to pass data among the severs and keep servers
in a cluster synchronized.  For example, you can impose a total
ordering across all messages or a causal ordering across the messages.
So you might want a causal order between subscription/unsubscription
message and data messages and a total order across all data messages
for a specific subject.

GCPs are not normally used to attach a message client to the cluster
because most GCPs do not normally perform well enough for 100's of
nodes and it is usually expensive to join or leave the group.


On the client side a pragmatic approach of putting something in the client
library to move from node to node in the cluster when the node appears
to fail, from the client perspective, or put some flavor of load balancing
switch in front of your cluster.


MR> Yes, there are more solutions around than i've expected,
MR> the performance of some of those tools is impressive.

A I recall, the Totem and Transis systems have even better perform
then Ensemble, but I don't think either group makes their software
generally available.


-- 
Best regards,
 David
 
R>  >
MR>  > Have you considered using a group communication protocol instead of a
MR>  > master/slave routing tree?


MR> They have a well defined discovery schema, which i thought to
MR> use.
MR> The problem with this approach is that i don't like it very much,
MR> it must be an optional possibility.
MR> I had a lot of trouble with 'osagent' the discovery of the Visibroker
MR> CORBA implementation - you never know what happens.

MR>  >
MR>  > See:
MR>  >   http://javagroups.sourceforge.net/

MR> This is a cool tool, probably a simpler approach than JINI discovery
MR> and ready implemented.
MR> As soon as basic xmlBlaster clustering is running (i hope the next days)
MR> i need to play with it.

MR>  >   http://www.cs.cornell.edu/Info/Projects/Ensemble/index.htm
MR>  >
MR>  > JavaGroups provides an implementation in java of several different
MR>  > protocol stacks. Ensemble is written in ML, but there are C++ stacks
MR>  > available for the Ensemble protocol.
MR>  >
MR>  > There is plenty of material to fan out from these two links.