VFS ROADMAP (and vfs01.patch stage 1 available for testing)

Martin P. Hellwig mhellwig at xs4all.nl
Mon Aug 16 03:50:27 PDT 2004


Matthew Dillon wrote:
<cut>
    Yes, I think you could.  The cluster abstraction would be
    a superset of the jail abstraction.  So much so, in fact, that the
    jail concept would be obsoleted.   e.g. the cluster will partition off
    cpu, network, and memory resources.  In fact, I think the cluster
    abstraction might even end up running a subset of the kernel itself as
    a loadable module... kind of a kernel within a kernel.   This would
    automatically isolate resource monitoring and related activities and
    sysctls to just what is visible within the pseudo-kernel, and the
    'real' kernel (or the 'parent' of the pseudo kernel) would control
    the actual hard resource allocation and scheduling.
    It will be a while before my thoughts solidify on how best to 
    implement a cluster (and input is always welcome).  

So, simply put (as far as I can understand this material), DragonFly 
will provide an API that has the hardware abstracted to an virtual 
hardware (which is clusterable over multiple nodes with different types 
of machine) kind of a distributed virtual machine (DVM). On top of that 
DVM there is a kernel which is specialized in the "DVM" hardware and has 
the understanding how to distribute, load-balance, replicate and 
fail-over system processen and other daemons. On top of that you can do 
anything like you would do with a normal OS, if you want only ,say 
PostgreSQL, thats fine but a whole full blown installation with multiple 
services is also possible.

This is very, very interesting and would be a elegant solution for tons 
of practical problems is have today.
One thing I don't understand is that if such a DVM is created wouldn't 
it be a problem when you have a mixture of 32 and 64 bits soliciting for 
a cluster?

Please forgive my ignorance on this matter.

--
mph




More information about the Kernel mailing list