Lynn W. Taylor wrote:
> 
> 
> Martin wrote:
[...]
> 
>> There is also a third "correct" in whether for the runtime 
>> statistics/credits, the credits score can be considered to be correct 
>> or just arbitrary random numbers that vaguely look to be in proportion 
>> to CPU runtime as compared to a s...@h WU...
> 
> If we're talking about BOINC-as-black-box -- about BOINC from a project 
> scientist's view, this is pretty unimportant.  That's one of the reasons 
> we have inaccurate credit.
> 
>  From a participant view, it's a huge deal.
> 
> That's why I'm in favor of anything that can dial-in credit 
> automatically for a project without having to tie up the project 
> scientists.  They're busy enough doing their research.

Nicely put and exactly so.

Hence the steady progression of how to better compute the credits as 
Boinc itself has progressed. We're now up to the idea of "auto-FLOPs"...


Perhaps now is a good time to jump in with the calibrated credits 
whereby you let your reference machine(s) do the "auto-FLOPs" bit for 
the entire project?...

The calibration certainly avoids all the headaches of trying to use 
client-side benchmarks despite highly disparate hardware, or of trying 
to instrument the client software, or of trying to do FLOPs counts and 
guesstimates on the client...

Also, to avoid all the projects reinventing the reference calibration, 
if wished, multiple projects can be referenced from the same one (or 
similar) machine(s) pretty much as is already done for NTP (and NIST).


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