Jeff Dike <firstname.lastname@example.org> said:
> > but spreading an instance across multiple nodes is nowhere as simple
> > as it seems;
> It is if you want to be sufficiently stupid about it :-)
> > where do you keep OS data,
> It gets faulted from host to host as needed.
Ugh Ugh Ugh. You need coherency algorithms for these.
> > IO access,
> This scheme (and any clustering scheme, I think) would need back channels
> for one node to access the devices of another
Right. You need communication services and coherency algorithms for these ;-)
> > scheduling decisions,
> This machine thinks it's a normal SMP box, so scheduling happens as normal
Ugh ugh ugh. Too many page faults; you need a scheduler capable of keeping
> > inter-node communication in the first place, how to deal
> > with node failure etc...
> Maybe I'm not familiar enough with the clustering world, but I was under the
> impression that with a normal SSI cluster, the nodes are like CPUs in an
> SMP box - if one fails, the whole thing dies. In other words, that SSI
> clustering and HA clustering are pretty disjoint.
No. In the optimal case, nothing dies. ;-) In the less than optimal case, only
the processes affected directly by the failure die (the processes which had
dirty pages there etc).
In the really useless case, the entire cluster goes down like on a SMP box or
an unpartitioned CC-NUMA.
"In search of clusters" is definetely highly recommended reading. It is very
Lars Marowsky-Brée <email@example.com>
-- Immortality is an adequate definition of high availability for me. --- Gregory F. Pfister
- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to firstname.lastname@example.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
This archive was generated by hypermail 2b29 : Tue May 07 2002 - 22:00:30 EST