Geoff wrote:

Klaatu wrote:

[...]

http://www.climateprediction.net/video/openday/davea.wmv

on the overclockers return bad data, ya i agree, a lot of poeple overclock,
and don't test their machines properly, but then, normal clocked machines
will return bad data too, they have allowed for it
give 1 work unit to 3 machines

There is a huge difference between overclocking or overheating (high) error rates compared to the very occasional errors due to cosmic ray or radioactive hits flipping a transistor bit.

the tools for testing overclocks are
prime95 in torture mode
memtest+ (or memtest86, same thing)

The /essential/ tests!


In addition to the earlier summaries:

From that link, Dr Anderson also mentioned in the talk about a group programming s@h for nVidia GPUs to boost the crunching performance [by x20 or x30 or so over present x86 CPUs].

He also seemed nicely aware of the 'social' aspects of s@h and that their PR could be better (:-))

He was keen that BOINC should permit the available DC resource to be fully usefully utilised for science projects rather than sometimes idle on overly redundant work just for the sake of 'social' aspects.

Getting the server stuff going is no trivial task and they've had to write their own performance monitoring scripts to help identify bottlenecks.


Happy crunchin',
Martin

(And yes, you can play Windoze.wmf files in linux (:-P))