On Thu, Mar 23, 2000 at 08:36:17PM -0800, Nick Hill wrote:
> http://www.cs.umd.edu/~keleher/bib/dsmbiblio/dsmbiblio.html
> thanx for that. But how much is the overhead on
> communicating the page tables on the network, what
> kind of network will u need? that was one of my
> questions too.
Page tables are not communicated. Basically if page A is shared between
machines M1 and M2, that page table entry for A is made read only on
both machines. If one of them writes to it, a page fault is generated.
The fault handler transfers the modified page to the other machine. So
you end up getting a lot more page faults than you intended.
You can run it on any kind of a network (the OS doesn't care - it just
sees it as a device) - but your performance will vary. I don't have any
numbers - check out the bibliography I mailed earlier.
There are variations of the above theme to make it more efficient. Wisconsin
guys deviced a method where coherency is maintained at a cache line level
using hardware mechanisms, but replication happens at a page level using
software mechanisms.
Oh, the wonderful world of CS research :)
-Arun
-----------------------------------------------------------------------
The LIH mailing list archives are available at:
http://lists.linux-india.org/cgi-bin/wilma/linux-india-help
