From: Steve Arnold (stevearn_at_microsoft.com)
Date: Wed Feb 25 2004 - 17:21:30 PST
In Porcupine, the authors are trying to achieve a goal very similar to
that in the Fox, et.al. project - that is a clustered system for
managing an application with many users that can scale simply by adding
more worker machines. The implementation of this system, however, is
quite different. The system here is completely functionally homogeneous,
and it is not centrally controlled like in the other project.
The application being implemented is a mail system. Users connect to the
system, but their requests may be directed to any of the machines in the
pool. Each machine looks just about the same. They contain fragments of
the mailbox (which can be replicated). However each node contains a list
of where to find the mailbox. Each node has a replication manager that
communicates with all of the other nodes.
The system - just as with BASE - does not maintain strict coherency. It
does take some time for all data to propagate throughout the network. In
their system, this did result in some interesting duplicated messages
(or missing ones) but this was only transient. They authors found that
their system is very close to being linearly scalable.
One limitation that they mention is that each node in the system
communicates with every other node in the system. While this allows for
complete replication, I have my doubts as to whether this is very
efficient and a good use of network resources. Also a large part of
memory on each node is used for the same thing. I would think that
having a central manager to manage resources would be a better approach.
This archive was generated by hypermail 2.1.6 : Wed Feb 25 2004 - 17:21:03 PST