A kernel level message passing service
hahn at coffee.psychology.mcmaster.ca
Fri Sep 1 13:09:26 PDT 2000
> > I have recently become aware of work being done to incorporate certain
> > types of services (ex: web servers) into the kernel of unix systems, so
the justification offered is that http is just another filesystem,
and since we have NFS in the kernel, why not khttpd? but putting
hhttpd in the kernel is nothing short of an admission
that Linux user-space is not viable in performance.
> > Does is make sense that this sort of thing could be applied to a message
> > passing interfaces for any sizable gain in performance? If so, is there
I don't understand how it could make any sense, outside of loading
your whole MPI/PVM system (app and library) into the kernel.
there already exist distributed-shared-memory kernel hacks;
that's about as close as you could come to a cogent kernel assist
for cluster programming. note that this topic is different from VIA-type
user-level hardware access. and if you want a low-overhead message
passing scheme that's still protected in the conventional sense,
why not use good old UDP or even IP datagrams. (I wonder sometimes
if the heavy machinery of TCP really makes sense in the context
of private, switched networks. does anyone have a count of dropped
frames inside a normal cluster?)
regards, mark hahn.
More information about the Beowulf