Page 12 of 17

Re: Suggested Change to the PPD System

Posted: Mon Apr 02, 2012 11:11 pm
by k1wi
That is a perception,

Removing the exponential growth assoicated with increasing PPD due to technological performance improvements removes the inflation, but it does not induce deflation - that would only occur if you decreased PPD at a higher rate than relative technological improvement.

The relative value of past contributions are maintained at constant. That is, above average contributions in the past remains equal to above average contributions in the present, Unlike the current system where above average contributions in the past are considered less worthy than even below average contributions in the present.

As to the carrot and the stick - because the point-in-time PPD curve does not change all it is doing is normalising the PPD at a given level. There is just as much carrot under the current system as there is in the old system, and there is just as much stick as there is in the old system, it just requires people to think of PPD in terms of relative computing power. This proposal does require people to make the mind-shift.

I understand you are fundamentally opposed to the QRB scheme and see it as a primary source of inflation. But I do not know of any theoretical change to the current system other than mine that will encourage people to invest in fewer faster machines without resulting in someone (read fast folders) facing decreasing PPD because of the removal of that source of point inflation.

Re: Suggested Change to the PPD System

Posted: Mon Apr 02, 2012 11:17 pm
by mdk777
If most countries in the world can determine a CPI then I don't see why FAH cannot determine the relative rate of computational improvement.
Because they have staffs of thousands gathering and analyzing the thousands of variables.
And then, the market..millions and millions of people debate if those numbers are really accurate and reflective of the actual economy.
Not for nothing, the market often diverges rather significantly from official projections.

Yes, your whole approach has the comical air of an old command and control economy.

Wheat production will increase 5% every year with our new collectives....Oh wait, didn't the Soviets and the Chinese try this already?

You discount not only know variables that you don't control...number of donors, hours per donor, watts per donor, FLOPS per donor, inherent lag time in communication and distribution of a distributed project, efficiency changes in high speed internet, data caps, etc. etc. etc.

But you also have the discount the unknowable unknowns.

My graphics card is rated at

3.79 TFLOPS Single Precision compute power
947 GFLOPS Double Precision compute power

However, it is generating something along the lines of 4000 ppd.

An i7 2600 is rated at what? 70 GFLOPS?

Now I accept this disparity in ppd production verses theoretically computational power because I know it reflects current reality.
My GPU is much better as a gpu shader calculation machine than a FOLDING computation machine.
I know that the current GPU client is using something like 1/8 of its potential even compared to other optimized gpu.

Hence the ppd reflects the actual donation to science and not the the potential of the hardware.

Now, can I predict if open CL, and Gromacs, and the FOLDING cores will eventually ever harness my 3800 GFLOPS as efficiently as an Intel CPU?

Perhaps, as I noted earlier in the thread, it might even gain a multiplying effect when used with a cpu, generating an effective multiple over a cpu.

I can't predict the efficiency of the entire software, hardware, science Eco-system. There are just too many variables.

By keeping ppd at some arbitrary constant, you are saying that you willing to not only predict, but also arbitrate the future value of this entire ecosystem.

Not even remotely as easy as you glibly hope.

Re: Suggested Change to the PPD System

Posted: Mon Apr 02, 2012 11:41 pm
by k1wi
mdk777 wrote:
If most countries in the world can determine a CPI then I don't see why FAH cannot determine the relative rate of computational improvement.
Because they have staffs of thousands gathering and analyzing the thousands of variables.
Funny... Statistics New Zealand has an entire work force of 1,000 employees, very few of whom work on the CPI calculation. The CPI calculation in New Zealand is determined by 690 goods and services, which are increasingly being recorded automatically through the internet and leading to greater efficiencies. Given that FAH is entirely computerised it would be very easy to create quite detailed reports automatically.
And then, the market..millions and millions of people debate if those numbers are really accurate and reflective of the actual economy.
Not for nothing, the market often diverges rather significantly from official projections.

Yes, your whole approach has the comical air of an old command and control economy.

Wheat production will increase 5% every year with our new collectives....Oh wait, didn't the Soviets and the Chinese try this already?
Thanks for calling me a Communist/Socalist..? Your comment is a blatent falsity. My proposal doesn't set the rate of future adjustments anywhere, adjustments are made based on measured results that have occured in the previous time period and apply for the next time period. If done at a short enough time period (again, 3 months) then it can be seen as setting current ppd adjusted to the improvement over the previous quarter. If a range of PPD measurements are used as a measure then that's the determined metric.
You discount not only know variables that you don't control...number of donors, hours per donor, watts per donor, FLOPS per donor, inherent lag time in communication and distribution of a distributed project, efficiency changes in high speed internet, data caps, etc. etc. etc.

But you also have the discount the unknowable unknowns.
PG knows quite a number of variables, they certainly know how many folders/clients there are. They would even probably be able to calculate the throughput of clients... Remember the adjustment is a aggregate adjustment and effects all users at a set rate.
My graphics card is rated at

3.79 TFLOPS Single Precision compute power
947 GFLOPS Double Precision compute power

However, it is generating something along the lines of 4000 ppd.

An i7 2600 is rated at what? 70 GFLOPS?

Now, can I predict if open CL, and Gromacs, and the FOLDING cores will ever harness my 3800 GFLOPS as efficiently as an Intel CPU?
PG already does this, it's their calculation of scientific value? It's their benchmarking system? They already work to ensure that folding is 'fair' across different architectures. As I suggested, it is possible for them to note that the relative value of one client is falling out of step with the other clients and make a revision. They already do that as it is.
Perhaps, as I noted earlier in the thread, it might even gain a multiplying effect when used with a cpu, generating an effective multiple over a cpu.

I can't predict the efficiency of the entire software, hardware, science Eco-system. There are just too many variables.

By keeping ppd at some arbitrary constant, you are saying that you willing to not only predict, but also arbitrate the future value of this entire ecosystem.

Not even remotely as easy as you glibly hope.
Once again, my system is not making predictions. It's measuring changes over time.

Re: Suggested Change to the PPD System

Posted: Mon Apr 02, 2012 11:52 pm
by mdk777
Thanks for calling me a Communist/Socalist..? Your comment is a blatent falsity. My proposal doesn't set the rate of future adjustments anywhere, adjustments are made based on measured results that have occured in the previous time period and apply for the next time period. If done at a short enough time period (again, 3 months) then it can be seen as setting current ppd adjusted to the improvement over the previous quarter. If a range of PPD measurements are used as a measure then that's the determined metric.
What you don't understand is that you making a value judgement by picking your starting point.
You are making a value judgement by picking the metric you choose to measure "improvement over the previous quarter"

You think these are simple statistics that everyone can easily agree on.

I don't agree with this premise. :mrgreen:

Perhaps others will. Good luck again. :wink:
Once again, my system is not making predictions. It's measuring changes over time.
Then it will be completely random. It will have no direct correlation to computational power, past or present, and will not "preserve" any past contribution.
instead of representing a fixed value, it will represent a Floating currency....a BITCOIN economy. ...

Re: Suggested Change to the PPD System

Posted: Tue Apr 03, 2012 12:17 am
by Jesse_V
mdk777 wrote:
My graphics card is rated at

3.79 TFLOPS Single Precision compute power
947 GFLOPS Double Precision compute power

However, it is generating something along the lines of 4000 ppd.

An i7 2600 is rated at what? 70 GFLOPS?

Now I accept this disparity in ppd production verses theoretically computational power because I know it reflects current reality.
My GPU is much better as a gpu shader calculation machine than a FOLDING computation machine.
I know that the current GPU client is using something like 1/8 of its potential even compared to other optimized gpu.

Hence the ppd reflects the actual donation to science and not the the potential of the hardware.

Now, can I predict if open CL, and Gromacs, and the FOLDING cores will eventually ever harness my 3800 GFLOPS as efficiently as an Intel CPU?

Perhaps, as I noted earlier in the thread, it might even gain a multiplying effect when used with a cpu, generating an effective multiple over a cpu.
FLOPS are not a good reflection of scientific value, that's been fairly well established by PG members and I've personally made the mistake of thinking otherwise and had to be corrected in that regard. There are other factors besides raw computational power like memory, bandwidth, lag, etc. As PPD attempts to align credit with the value of the scientific results and make a system of equal pay for equal work as it says on the F@h website, I see inflation over time slowly contradicting that goal. For example, if we look at F@h's 2000-2004 News blog at http://home.kpn.nl/tomhuls/folding/news.html, look how excited they are when someone breaks the million point record. For a dedicated folder even on a laptop that's getting more and more commonplace. Yes there's a lot more work done but you can kind of see what happened.

Can we summarize both sides of this discourse so that if the PG is ever sent a request for comment they don't have to read the entire thing? Or is it all right for them to read the entire thread? Whenever both sides are satisfied with explaining their positions, IMO we should summarize/organize it and then invite Bruce (our Donor Advisory Board representative) or someone else to present the plan and see the official reaction. After all, the proposal rests on the PG's inside statistical knowledge so eventually we should see what they say about this whole thing.

Re: Suggested Change to the PPD System

Posted: Tue Apr 03, 2012 12:21 am
by k1wi
Jesse - there are a number of summaries of the project, such as: http://foldingforum.org/viewtopic.php?p=210802#p210802, which could possibly be tidied up. I am all for continued debate on the project, and I am still very willing to hear from people who would like to improve it, even if it means I have to spend time discussing the change with people who have their set view in opposition and who attempt to continue to include comparisons between communism and bitcoin, all of which are in my opinion, fear-mongering.

Re: Suggested Change to the PPD System

Posted: Tue Apr 03, 2012 12:48 am
by mdk777
Not fear mongering...just different starting points.

I think points should represent some fixed computational metric.

If my computer increases my computational ability, 10x, 100x, or 10,000x over the years i expect my points production to follow.

You seem to ascribe to a worker deserving a "fair fixed daily wage" That as his ability to produce increases over time, 10x, 100x, 1000x, his daily ppd should remain constant.

This sounds like an economic theory I have heard of.

It is not one I ascribe to.

And yes, what you have described is very very similar to how the Bitcoin system was described as functioning.
If you are offended by this, I would be happy to hear how you think it is fundamentally different.

Re: Suggested Change to the PPD System

Posted: Tue Apr 03, 2012 1:04 am
by k1wi
mdk777 wrote:If my computer increases my computational ability, 10x, 100x, or 10,000x over the years i expect my points production to follow.
Right, because you don't want to measure PPD in terms of relative effort - that's fine. I happen to think that because computation power increases over time the 'scientific value' should be adjusted over time to ensure it remains relative.
mdk777 wrote:You seem to ascribe to a worker deserving a "fair fixed daily wage" That as his ability to produce increases over time, 10x, 100x, 1000x, his daily ppd should remain constant.

This sounds like an economic theory I have heard of.

It is not one I ascribe to.
The 'daily wage' isn't fixed, it's measured in terms of real production as opposed to nominal production, but not fixed. I see you like to use economic policy anaologies so I will dance at your party - I don't know of any economic system where a printer technician (worker) got paid twice as much simply because the printing press he worked on became more efficient and produced double the number of newspapers.

My policy is not unlike the Fed Reseve setting monetary policy in order to control the supply of money in the market.

Re: Suggested Change to the PPD System

Posted: Tue Apr 03, 2012 1:15 am
by 7im
Just wanted to make the distinction that fah's "equal pay for equal work" is not the same as "fair wage." One is a discussion about the goal of assuring that relative scientific contributions to fah are awarded relative points even from differing hardware, while the second is a discussion about how much those relative points should be.

Re: Suggested Change to the PPD System

Posted: Tue Apr 03, 2012 1:23 am
by mdk777
The 'daily wage' isn't fixed, it's measured in terms of real production as opposed to nominal production, but not fixed. I see you like to use economic policy anaologies so I will dance at your party - I don't know of any economic system where a printer technician (worker) got paid twice as much simply because the printing press he worked on became more efficient and produced double the number of newspapers.
perhaps not instantly, but worker productivity drives GDP which drives the total standard of living of an economy.

Yes, I make 10,000 x what my great-grand father ever dreamed of making.
I drive 1000 miles a day in comfort. (just drove from Chicago to New Orleans the other day)
I fly 5.000 miles in mere hours. I do things never dreamed of in the History of the world...things no King ever dreamed of...and yes, it is all because of that insane increase of productivity due to capital investment, technology, doubling of productivity.

My wages might not increase over-night...but over time, as the economy expands, my standard of living increases.

If my sales go up 4x or 10 x and my income doesn't change, you can bet I will be looking for a different employer. :mrgreen:
My policy is not unlike the Fed Reseve setting monetary policy in order to control the supply of money in the market.
close. Except the money supply increases over time a fairy good clip to reflect the growth of the economy. Look up the figures.

What you are proposing is closer to the Chinese who keep their currency at a fixed exchange rate even though their economy has increased 10x over the last few years....A major concern and intentional disruption in the world economy.

Re: Suggested Change to the PPD System

Posted: Tue Apr 03, 2012 1:29 am
by ChasR
k1wi, your quest is noble. I appreciate seeing a new idea and understand where you're coming from. I just don't think it's doable without causing a huge disruption at each "normalization" and in the end, you just measure the relative performance on a different scale.
As for deflation, as I understand your thinking, if I choose not to upgrade my machine(s), over time I will make less and less ppd on an exponential curve, while doing the same amount of science. I would call that deflation.

Re: Suggested Change to the PPD System

Posted: Tue Apr 03, 2012 1:51 am
by k1wi
ChasR - I appreciate that my proposal is unwelcome to some.

To clarify my proposal, the points of an individual machine would decrease over time. But at the moment the exact same thing happens it's just relative to other users - if you don't upgrade your computer then your PPD relative to other users follows the exact same curve - your PPD remains constant and everyone else's skyrockets as people upgrade their computers to new technology. The relative point proportion at a given point in time will remain exactly the same under the new system, it's just under the current system we end up with exponentially large PPDs that become relatively meaningless to the regular user. At the same time, the new proposal continues the incentive to return work units faster, just without the compounding inflation currently present under the QRB (because we account for that inflation too).

I would hope that 'huge' disruption can be avoided by making the adjustment process something that is ingrained into the FAH system - make it expected and make it regular and people will expect it, just as we get use to the Fed Reserve adjusting the official cash rate. Make it known that at a regular interval PPD will be adjusted to keep it relative to computer production. That in itself would be a huge improvement over the current system where PG faces a massive malstrom everytime the points system is tweaked in order to better achieve their goals and scientific value. I feel that '3 monthly point updates' would be the best balance, because adjustments would be small, but not so regular that you wouldn't know the PPD from one day to the next, and who knows, perhaps once folders have gotten use to the points system being adjusted on a regular basis PG can bring additional adjustments into the system - so that Scientific Value continues to be massaged to suit PGs goals, but in a systematic method rather than the current process. I.e. at the 'March Update' PG could notify that overall PPD will decrease by 4%, and that at a given point in time (say a month later, or at the next update) the premuim on Bigadv (over SMP) will be increased by 5% in order to better align with their scientific value, and the GPUs differential will decrease by 5% so that they remain aligned to their scientific value.

After all, why not create a systemic expectation around the current expecation that you should "not base future PPD on current PPD"!

Furthermore, the second benefit to my proposal is that at the moment all your accrued points deflate exponentially, which means your previous work is also devalued relative to the work being done today. At least bank accounts have interest rates that (in the long run) compensate for inflation!

Re: Suggested Change to the PPD System

Posted: Tue Apr 03, 2012 2:35 am
by ChelseaOilman
I sold my 6 x i7 920 systems and my 3 x 2600K systems. Then spent about 8k building 2 4P systems. All I hope is PG gives me a heads up if they decide to neuter the points system for 4P systems, so I can sell them while they're worth something. Then I'm done with F@H. I'm never going back to 1P systems. If PG would rather have a bunch of old 17 systems folding their WUs, more power to them. It's been a nice ride the last 10+ years.

Re: Suggested Change to the PPD System

Posted: Tue Apr 03, 2012 2:56 am
by Grandpa_01
Fundamentally I really see no problem with the current point system. What k1wi is proposing as I see it is a theoretical way to curb inflation. I do not know that curbing inflation is a good thing or a bad thing. I do know that people look at what there neighbour has and what his neighbour is getting paid for his efforts. The reality of it is if you want a person to upgrade and then you are going to have to give him an incentive to do so. I myself look at the value of what is available and if the value is there and I have or can comfortably achieve the means to upgrade then I will upgrade. if any of those factors are missing I will not upgrade. The current point system does a very good job of giving incentives to upgrade and promotes it. It gives clear messages as to what is needed and what direction I should take, (At least it did until they started listing to public input)

I could easily look at the PPD given for any class of WU and tell what Stanford wanted and needed, that is becoming a little more difficult now because people are muddying the water with there complaints about there neighbour is getting too many points and it is not fair. I on the other hand have a very hard time believing that Stanford did not know the value of the work / science and that they made any mistake in assigning that value. But some believe they are smarter than Stanford and that Stanford does not know the value of the science as well as they do.

What can I tell you by looking at today's values of the bigadv WU's . I can tell you that if they stay in line with the 6903, 6904 it is worth it to upgrade to a 4P system if you have the means to do so and are willing to purchase a machine that has no other use to you than folding. When I am making a decision I have to remember these 4P rigs have no other practical purpose for me other than folding so that has to be taken into consideration. Now when I look at the value of the 8101 well that is a different storey. They have a very low comparative value, the last 1 that I ran I actually only made 170,000 PPD I can make 200,000+ PPD folding smp on average. Now I have a Gulftown rig running at 4.5Ghz with a GTX 580 and I can get up to 96,000 PPD on it. If the new bigadv is going to be the standard then there is no way I would invest in a 4P rig. It is just not worth it. There is way to much cost and maintenance involved for to little reward compared to other choices. So as you can see it is a little muddy and confusing. Did I make the right choice, I do not know for sure at this point in time.

Perhaps that is what Stanford wants is more smp and GPU rigs if you listen to some that is what they are going to get, which is fine with me I will follow there lead which ever way it takes me but I do not think that we as donors should be setting the points value to our own personal choices. It is there project not ours and they should be able to use there own tools and personal knowledge to determine the value of the work. I am positive that Stanford knows far more than I or 7im does about what the value of any given WU is. :wink:

Re: Suggested Change to the PPD System

Posted: Tue Apr 03, 2012 3:31 am
by k1wi
Grandpa_01 wrote:I could easily look at the PPD given for any class of WU and tell what Stanford wanted and needed, that is becoming a little more difficult now because people are muddying the water with there complaints about there neighbour is getting too many points and it is not fair.
Thanks for your input Grandpa01,

I really do think that the above issue is fundamentally because the absolute numbers are getting pretty large - keep the proportion of points the same (i.e. a computer twice as fast earns four times the points or however much it is under the QRB) but keep the absolute range tight and people will be less likely to get gobsmacked by PPD. I'm much less likely to find issue with a 10,000 PPD difference than I would a 10,000,000 PPD difference, even if the proportion was the exact same. I'd be even less likely to find an issue with the system if a faster computer earned the same proportion more points than me, but didn't sweep past my years of folding in two weeks... If they fold twice as fast and earn four times the points then I'm happy for them to accumulate the same number of points as I have in a quarter the time, but not in a thousandth just because they bought their computer now and not last year!

Control the proportion of users folding BigAdv by controlling the deadline and also adjusting the differential between SMP and BIG so the scientific value is aligned with Stanfords intentions (that is, a shift of the curve)