[FRIAM] Google & IBM giving students a distributed systems lab using Hadoop

[Marcus G. Daniels]

Owen wrote:
>> My guess is any realistic solution will be hybrid, combining the
>> features of all these large scale architectures.
>>     
By designing circuits to do special purpose compute tasks, the lengths 
of wires can be reduced (and thus their diameter) and this is ultimately 
the limiting factor on serial performance and circuit density.   In 
practice, designing these circuits is hard to automate and optimize 
(even with FPGAs) and even relatively general hybird computing 
approaches like the Cell broadband engine still require quite a bit of 
programming finesse (e.g. prefetching and keen awareness of memory 
access patterns, etc.)

Of course, at some point the wire diameters can get no smaller, and we 
have to look to programming approaches to find parallelism.   One 
software technology that looks promising to me is software transactional 
memory:

http://en.wikipedia.org/wiki/Software_transactional_memory

Apparently Sun is working on hardware support for it..

http://www.theregister.co.uk/2007/08/21/sun_transactional_memory_rock/