On Wed, Jul 28, 2004 at 10:03:05AM -0700, Ernest Prabhakar scratched on the wall:
> The confusion is probably due to the fact that Firewire -- as an
> arbitrated bus -- offers some *theoretical* possibility of being used
> as a very low-latency transport for, e.g., MPI. Even, presumably,
> compared to non-IP utilizations of Ethernet (e.g., GAMMA
> <http://www.disi.unige.it/project/gamma/index.html#aBSTR>).
One thing that I've always wondered about is using an OS-bypass
system for Ethernet, just like InfiniBand does. Ethernet port-to-port
latency, especially with modern high-end stitches, is within an order
of magnitude to InfiniBand and other high speed interconnects. The
killer for Ethernet is the IP stack, and to some extent, the context
switches and memory copies to get in and out of kernel space.
Something like GAMMA still takes the second cost (as would a
traditional kernel based Firewire driver to some extent, I would
assume), but has the added advantage that you can still use the
Ethernet port for more traditional things, like IP. A MPI/OS-bypass
system would likely take over the Ethernet port and wouldn't allow it
to be used for anything else. On the other hand, the Xserve has two
ports (although let me guess... integrated controllers?), and Ethernet
is dirt cheap, as networks go-- I'm sure there are people that would
be happy to give up an Ethernet port on each node if they could save
a bundle on their interconnect switch.
> However, since there are currently no MPI over Firewire implementations
> available, this is just speculation.
Kind of like this quote; I love the way this huge and massive project
is slipped in as a sub-clause of a sentience, making it sound so easy
and casual (a good weekend project):
http://www.apple.com/macosx/tiger/unix.html
Under "64-bit Compatibility", second paragraph:
------------------------------------------------------------
For example, [...] a computer scientist can create a shared
memory architecture for an entire cluster.
------------------------------------------------------------
Although I suppose it is true; it could be done, and it would be
pretty cool. An S-COMA memory system, or something similar, for
Mac clusters could be very interesting.
-j
--
Jay A. Kreibich | Integration & Software Eng.
email@hidden | Campus IT & Edu. Svcs.
<http://www.uiuc.edu/~jak> | University of Illinois at U/C
_______________________________________________
xgrid-users mailing list | email@hidden
Help/Unsubscribe/Archives: http://www.lists.apple.com/mailman/listinfo/xgrid-users
Do not post admin requests to the list. They will be ignored.
