The Clustering and Userland VFS transport protocol - summary

[Matthew Dillon]

:Unresearched... Matt, didn't you have a distributed multi-master
:database for backplane? If we just consider each memory page a row in
:a table, then the same algorithms used for the database could be used
:for the memory coherency; that's not to mean the actual implementation
:may be more complex due to pagefaults and related stuff
:
:-- 
:Greetz, Antonio Vargas aka winden of network

    Yup, I did/do.  But it wouldn't work with memory per-say because the
    database can always return a commit failure, whereas we can't return
    a failure if a program modifies memory which is later determined to be
    stale.

    I'm working on a cache management infrastructure.  Actually two
    cache-management infrastructures.  One for serious cache management,
    and another to allow large blocks of data (such as the data in a
    read() or write()) to be represented out of band. 

    But before I get into that I am first working on expanding the 
    journaling protocol to become a true RPC abstraction, similar to
    VOPs, but capable of transparent operation over arbitrary 'reliable'
    connections (TCP streams, pipes, memory mapped interfaces, direct calls,
    etc).  The key is to make it degenerate into a direct call when the
    source and target are in the same context.  Then both userland VFS and
    clustering support can be built on top of it.  Stay tuned!

					-Matt