Mikus Grinbergs wrote:
>>> Goal: The credit granted for all tasks for a single WU needs to be
>>> identical, no mater what computation devices it runs on.
>> Sorry, disagree... What happens for the case where the CPU application
>> is highly optimised whereas the GPU application is brazenly wasteful of
>> the GPU resource? The GPU does more physical work and gains less credit
>> per unit of physical processing work done.
> I'm one of the people to whom "expanding the knowledge of the world"
> is important, rather than anyone's BOINC credit score. As such, I
> believe that credit ought to be given for the __contribution__ to
> the project made by the participant. Let's say one participant uses
> inefficient hardware/software, and contributes two "sacks" of
> answers per 24 hours. If another participant uses better performing
> hardware/software, and contributes six "sacks" of answers per 24
> hours, the 2nd participant should get three times the credit of the
> 1st participant - because he has __contributed__ three times as much
How do you *measure* "expanding the knowledge of the world" so that
'something' can be 'awarded' fairly? That is all very subjective. That
'something' hopefully attracts a participants interest in a project in
the first place. But how can that 'interest' be *measured* or
proportionately 'rewarded'?
Hence, instead we have the "cobblestones" that are based on one very
(hopelessly inaccurately for Boinc) small aspect of CPU performance.
That has _nothing_ to do with the (subjective) 'value' of the science done.
Note that to have one WU awarded exactly the same cobblestones credit
regardless of the hardware used to process that WU, then we must have
some scheme to calibrate the cobblestone award for each WU based on, or
referenced to for _ALL_ resource parameters, some golden standard computer.
Also, will all hardware platforms process that WU to the same level of
thoroughness, precision, finesse?
What of WUs that run a Monte Carlo simulation that can randomly take any
of many different paths/lengths?
I think that with the diversity of Boinc projects and the heterogeneity
of supported platforms, we have moved a very long way away from:
"1 WU = 1 Credit".
We have moved a long long way away from just 'rewarding' for s...@h FFT
calculations.
Hence, instead we should reward *fairly and accurately* that which can
be fairly and accurately *measured*...
[...]
> he should *not* get more than 1/4 of the credit given to the 1st
> participant - even though that 3rd participant may have put in
> herculean efforts in getting answers out of that obsolete system.
Exactly so, I agree. You get the expected 'rewards' if you reward for
the compute resources *used* (note that for your example, the obsolete
system takes 4 times longer to provide that resource).
>> I consider a fairer scheme is that credit is granted for the resource
>> actually used.
>
> I disagree. Let me suggest that credit be granted for results
> (i.e., __contributions__ to the project), rather than for
> "resources" being applied.
[...]
> It should not matter if results are produced from CPUs or from GPUs;
> nor if particular chips ("resources") get underutilized or not; nor
> if the software running those chips is optimized or inefficient --
> the credit ought to be based on the __contribution__ (meaning
> "usable" answers) delivered by the participant, irrespective of the
> tools the participant used to produce that contribution. If a
And there is the problem in your assertion. You can't measure the
'contribution' in terms of 'value' of 'results'. You can't /compare/ a
subjective 'value' of results across multiple projects.
The present units being used for "credits" are in terms of used
resource. We should abandon trying to senselessly corrupt the
cobblestone units for trying to meaninglessly put a 'value' on
'results'. Instead, just use them for what they actually represent:
*resource used*
We might then finally get away from trying to perpetually statistically
establish a new field of Bistromathematics[*] in trying to prove how
many oranges equal one tomato.
Regards,
Martin
* See:
http://www.geocities.com/CapeCanaveral/Lab/3550/bistro.htm
--
--------------------
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.
