Cache coherency, clustering, and Kernel virtualization

[Matthew Dillon]

:with "multi-grid clustering" I meant joining multiple grids with one 
:logical machine, basically what you're trying to achieve for december; 
:MPP = Massively Parallel Processing, a shared memory machine, which is a 
:cluster but not in the sense that several individual hosts work on a 
:problem but are otherwise a system in their own right (i.e., MPP 
:machines are always SSI and share one virtual memory map). IRIX on a 
:cluster of Origins is MPP and SSI for example, but an application using 
:MPI to run on multiple nodes isn't.
:
:Cheers,
:-- 
:         Thomas E. Spanjaard
:         tgen at xxxxxxxxxxxxx

    DragonFly's SSI is from the point of view of user processes and 
    mechanisms within the kernel to guarentee cache coherency between
    distinct machines for user visible things like the process list,
    directory structure, file contents, devices, etc.  But the DragonFly
    kernel's own memory map, i.e. 'kernel memory', is *NOT* intended
    to be shared.

    Theoretically one could run a virtual kernel (which is a user process)
    in a clustered system and configure it with thousands of 'cpus', and
    the virtual kernel's memory would then be shared (and, in fact, this
    would be a very good way to test the clustering technology), but the
    DragonFly kernel is not designed to perform optimally that way.

					-Matt