Very nice, and very reminiscent of the "processor farms" we used to
build out of gumstix-sized transputer TRAM modules in the late '80s.

Of course you'd need a whole boxful of 20Mhz transputers to compete
with a single 600Mhz omap, but the transputer did make it surprisingly
easy to get linear speedups by adding more processors, thanks to the
low latency of communication on its serial links.  Because link
read/write was performed with a single instruction, with process
scheduling and interrupt handling microcoded into the cpu (no kernel
required), the overhead for a send or receive (from user code to bits
on the wire or vice versa) was about 20 cycles, or 1μs.  I would be
interested to know the equivalent figure for overo's ethernet (with
linux or plan 9 helping out).

Reply via email to