[extropy-chat] COMP: Distributed Computing

J. Andrew Rogers andrew at ceruleansystems.com
Thu Nov 18 18:41:24 UTC 2004


Henrique Moraes Machado <hemm at openlink.com.br>:
> This would be interesting. and what if some guys (and girls) that
> waste their time developing viruses and trojans started to develop
> the ultimate trojan, based on AI to turn every infected computer
> into a neuron of the world brain. Good plot for a SciFi novel.



This is implausible.  Latency bounds scalability for algorithm spaces
like these.  At some point, it costs more to distribute the work than to
do it yourself.  And for real people doing real work on real systems,
shaving a microsecond off the interconnect latency is a big win in terms
of overall performance, scalability, and how much work you can
distribute.  You'd be better off with a single machine with a gobs of
disk swap space than a global cluster with hundreds of milliseconds or
seconds of latency.

You can no more geographically distribute a problem like this to make it
scale than you can geographically distribute an ACID-compliant database
to make it scale -- same theoretical limitation in both cases.

j. andrew rogers




More information about the extropy-chat mailing list