[Beowulf] Teraflop chip hints at the future

[Eugen Leitl]

On Tue, Feb 13, 2007 at 11:14:55AM +0100, Vincent Diepeveen wrote:

> Can you explain the OS issue to us?

DRAM die piggy-backing is rather expensive and has
its own limit on the memory bandwidth issue, so long-term
memory will have to be embedded within the CPU. Because
of yield limits such embedded RAM will have only very small 
sizes, few MBytes at best. (This also opens the way
to wafer-scale integration, which has also been overdue
for a very long time). 

Current kernels would be hard-pressed to alone fit into
such tight memory spaces. Fortunately, there is no point to
include code for e.g. I/O, MMU (notice the Cell doesn't do
MMU for SPEs, instead using cache transistors as SRAM), 
video etc. in an effectively embedded node, so kernels can 
be slimmed down to few 10 kBytes, thus limiting redundancy. 
Another critical point is to put message passing (ideally, a 
subset of MPI) directly into machine instructions, to limit 
latency. The on-die/on-wafer mesh fabric has to send message 
with almost the same penalty as the on-die access. 
OOP does implicit memory protection if there are enough cores,
because the only way to impact another node's address space
is by sending a message. There also needs to be a machinery
which allows a large object to be recursively decomposed
into composite objects, which will eventually fit into such
a small node.

This is not exactly Beowulf anymore, even if such hardware
is commodity. Linux might have some life in it, though, if
it goes the L4/L3 way, and only runs the whole Linux hog on
the fat node (host system).

Caveat: my crystal ball might or might not be defective. It's
been showing me the same thing for the last decade.

-- 
Eugen* Leitl <a href="http://leitl.org">leitl</a> http://leitl.org
______________________________________________________________
ICBM: 48.07100, 11.36820            http://www.ativel.com
8B29F6BE: 099D 78BA 2FD3 B014 B08A  7779 75B0 2443 8B29 F6BE