Lynn W. Taylor wrote:
> If you really want to open that can of worms, how about the fact that a 
> floating point add (which is a floating point operation) is dramatically 
> easier than a floating point cos().
> 
> ... and the fact that an AMD processor might do adds dramatically faster 
> than an Intel, but do cos() slower.
> 
> A hypothetical "add...@home" could give dramatically higher credit just 
> because it does particularly fast floating point operations, especially 
> compared to "trigonome...@home."

Very true.

> I don't see a way out, short of doing exactly what Eric Korpela's script 
> tries to do -- normalize FLOPS credit to that predicted by the 
> (imperfect) benchmarks.

Except that assumes that all projects behave in the same way as s...@h. In 
reality, they appear not to. Note that s...@h is dominated by certain FFT 
calculations...


An improvement could be to improve the FLOPs counting so that due 
weighting is given to the different operations performed. For example, 
assume the optimum average ops needed to calculate "cos" or whatever and 
so add by that amount to the ops count as appropriate.

If new dedicated hardware comes along (FPGAs or GRAPE anyone?) that 
improves on that 'cost', then that new hardware should indeed gain a 
rewards advantage until the hardware becomes dominant. Then adjust the 
scoring for that one operation as needed. Those that then don't have the 
new hardware will get downgraded for their older hardware for that 
operation. But then, those users should already have migrated to other 
projects that can take better advantage (and score better) for their 
hardware...


So... reward proportionately for whatever resource is actually used. 
Don't try scoring for what is available (but unused). Let the users 
apply their hardware where it can be best utilised for whichever 
project, with the RAC showing a true score for utilisation.

(Note: Benchmarks assume maximum utilisation for whatever they measure...)

The problem then moves to how do you score for all the various resources 
and various operations? Including such as memory bandwidth, max memory 
used, and others?

Also, how do you easily add up the scores for a completed project WU?


At the moment, the present "assume everything is a FLOP" appears to be 
rather too coarse and noticeably inaccurate...

Regards,
Martin

-- 
--------------------
Martin Lomas
m_boincdev ml1 co uk.ddSPAM.dd
--------------------
_______________________________________________
boinc_dev mailing list
[email protected]
http://lists.ssl.berkeley.edu/mailman/listinfo/boinc_dev
To unsubscribe, visit the above URL and
(near bottom of page) enter your email address.

Reply via email to