> You might look into distributing swap as a first
> step toward doing the
> distributed shared memory implementation; that is,
> allowing any node to
> use any swap device. This was supported in NonStop
> Clusters for Unixware
> but I don't think that has made it into OpenSSI.
> The HP folks will probably
> have hints for you as to the implementation, or
> possibly even code, if any
> of the dswap support code was freed with the rest of
> Frank Mayhar frank@... http://www.exit.com/
> Exit Consulting
by "distributing swap" would this mean
-the virtual memory system would have to advertise all
the changes to page tables to all nodes and to
advertise its memory map of the swap space.
-a subsystem would have to translate the virtual
address to a geographical (node no.) address, then
pass the page fault or swap out request to the correct
node (what would this involve?) at which the local vm
system would handle it and if necessary return the
page or a confirmation.
-copy-on-write and other memory mangement can be done
locally after the above translation and request to the
identified node is complete (?)
-since the virtual address will be resolved to first
of all a node, and then locally to a linear address
will this be enough to facilitate per-process
allocation and deallocation of main memory across the
cluster? or is this a much too simplified view?
These are just some ideas of what is needed to to
share memory if someone can point me to more
implementation specifics it'd be great
p.s. frank you mentioned the "folks" at hp may be able
to help. do you know of any folk in particular?
Yahoo! Messenger - Communicate instantly..."Ping"
your friends today! Download Messenger Now