Re: [sl4] The Jaguar Supercomputer

From: J. Andrew Rogers (
Date: Tue Nov 17 2009 - 18:03:20 MST

On Nov 17, 2009, at 2:23 PM, Daniel Yokomizo wrote:
> The Monte Carlo AIXI is embarrassingly parallel:

That code will not scale on a conventional cluster computing architecture the way I think some people are assuming unless you significantly restrict the scope. I do not see how it is fully parallelizable in a general way unless I missed something -- the Monte Carlo piece is just a small part. Regardless, it is still a fairly pathological graph-like code with all the scaling and efficiency issues implied on real-world hardware.

In practice, these codes frequently run faster on good desktop machines that a conventional supercomputing cluster. Unfortunate but true. Unless, of course, you use a supercomputer designed for graph-like codes; those would run MC-AIXI very efficiently. You just won't see them in the Top 500 because they aren't built for that.

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:01:05 MDT