Re: Clustering for Linux 2.3.x?!

[Erik Elmgren <erik.elmgren@swipnet.se>]

Quoting Robert G. Werner (rwerner@lx1.microbsys.com):
> I'm not sure if you could ever see a networking scheme that mere mortals could
> get ahold of that is going to be able to keep up with processors wanting to
> talk to each other.

Depends on the app, distributed.net is a fine example of a message-passing
application with lots of CPUs and poot latency.

> The problem isn't bandwidth as much as latency between
> when the message leaves one processor and arrives at the other processor.
> Between we have several layers of protocol processing that is going to have to
> be done on both ends.  

You claim less latency for distributed-shared-memory or that shared-memory
programs are more latency tolerant or both? Or are you just telling me the
stuff I forgot to include? :-)

My argument:

1. distributed-shared-memory and message-passing have the same latency and
   the same bandwidth (approx and for clustes).
   
2. Applications designed with message passing:

2.1 Uses resources /slightly/ better in a cluster as the system don't need
    to guess 'how much' and knows 'to' when 'send' rather than 'receive'.

2.2 Tend to tolerate latency better.

2.3 Tend to be more fault tolerant.

It has been claimed (not here) that shared-memory is superior because many
programmers know it better than message passing.

I also forgot to say that message-passing could indeed benefit from kernel
modification, but not from total redisign.

Erik

--

Has it ever occurred to you that God might be a /committee/?
--- Jubal Harshaw
-
Linux-future: thinking about the future of the Linux kernel
Archive:      http://humbolt.nl.linux.org/lists/
Wish list:    http://users.ox.ac.uk/~mert0236/linux-future.html