Roger wrote:
> Tell that to the people who complained that the GB_FLIP
> period (_1+2^55) is not long enough.
Mainstream personal use computers these days have 4 effective CPUs, each
clocking in at 3 Ghz. Assuming you devoted one such
machine to entirely to generating random numbers using the GB_FLIP algorithm,
and each processor could generate one random number
per cycle, then:
;:^:_1 'MWDhms' ,&":&.>~ _ 4 7 24 60 60 #: <. (<:2^55) % 4*3e9
1M 0W 6D 17h 59m 59s
It would take over a month before algorithm cycled. Of course, this ignores
real-life concerns such as memory size, bus speeds,
cycles lost to an OS and other processes, not to mention you probably couldn't
generate a random number per cycle, and even if you
could, since each number depends on the previous, then you couldn't parallelize
the algorithm across 4 processors like this. So
add a little buffer and call it two months.
I'd be interested in learning which applications have such appetites for random
numbers that this period is too short for them (and
do they really run models over the course of a month?).
-Dan
----------------------------------------------------------------------
For information about J forums see http://www.jsoftware.com/forums.htm
