Suggested Change to the PPD System

[k1wi]

I have just been trying to explain my original hypothesis.

[7im]

7im - I would never advocate a weekly schedule. Perhaps monthly is too frequent, but quaterly would be feasible, perhaps even half yearly. Put on a schedule just like the Fed Reserve and it becomes a known part of the folding calendar. Even annually would be fine, however I think there is a danger in only doing it too infrequently. It will still be possible to determine between project change and normalisation change because all projects would be affected equally and at the same time, so you can compare it to other current projects.

Okay, let's see where the goes ...

But IMO, we still must have a multi-socket benchmark PC first, or this goes nowhere fast.

[MtM]

I have just been trying to explain my original hypothesis.

Which wasn't that clear to me in the first posts, and as I read it you changed it in the progression of the thread. I was against normalisation based on anything other then computational throughput, when you started using that as where to base the normalisation difference on you had me on your side. The same goes for ChasR, the moment he showed the point of interest in the QRB which invalidated it for him as being 'in line with science' I had to agree with him and proposed a conceptual 'fix' for that. I'm just not capable of converting the concept into a formula, so I'm hoping others will pick up where I ( have to ) slack off

7im - I would never advocate a weekly schedule. Perhaps monthly is too frequent, but quaterly would be feasible, perhaps even half yearly. Put on a schedule just like the Fed Reserve and it becomes a known part of the folding calendar. Even annually would be fine, however I think there is a danger in only doing it too infrequently. It will still be possible to determine between project change and normalisation change because all projects would be affected equally and at the same time, so you can compare it to other current projects.

Okay, let's see where the goes ...

But IMO, we still must have a multi-socket benchmark PC first, or this goes nowhere fast.

It would be nice if this discussion ends in a usable formula this time

And I'm still interested in hearing about the need to benchmark gpu work unit's on both 48 core gpu's and 448 core gpu's because when you would apply the current smp qrb on gpu work unit's you would see the same problems as we're discussing now with smp except for the fact that we don't have different categories ( yet? ) there. We could reserve that one for later though if you want.

[Grandpa_01]

...

Just curious but where would the incentive be for people to buy the machinery and to run the bigadv WU's if it were normalised across the board. The purpose of the QRB was to encourage quick returns. Guess what it works just look at all the 4P out there right now and there are quite a few being planned and built. Why would I or anyone else run a bigadv WU if we could make the same PPD off of a smp with far less risk.

You can't seriously think a 1 core benchmark would be normalized against a 16 core benchmark? We don't do that now, and never would. Please excuse yourself from the discussion if you cannot keep it reality based.

ChasR said use the same 16 core machine for CPU, SMP, and BigAdv. But only use and bench for 1 core on CPU, 4 or 8 cores on SMP, and 16 cores on Bigadv. (or whatever PG finds appropriate to use) They'd all have a different base PPD, as one would realistically expect. Maybe Base points for CPU client, and Base x 16 for BA. And then add the QRB on top of that. Still PLENTY of incentive for BigAdv with QRB.

This also seems like it would make the benchmarks more accurate. Tough to argue against that...

Just curious where did I say 1 core anything. You comment was unjustified and rude 7im and quit trying to twist thing around.

[orion]

I'd like to see Grandpa's numbers on regular smp work on his 4P.

I'm not Grandpa but here's 24hr of vanilla SMP on a 4p. http://folding.extremeoverclocking.com/ ... =&u=102167

Look for the date of 03.13.12

I have one question...how many of these threads are going to be started?

[Grandpa_01]

He already has the smp PPD avg I posted it back when he asked for it. And just a little insight to what it takes to to avg 1,200,000 to 1,500,000 PPD
Hodware cost at bargain shopping prices for my folding rigs.

2 - 4P rigs cost $6,000.00 countless hours shopping My only use for them folding. If a person has to pay retail $10,000.00 to $15,000.00 ea.
5 - Hex rigs $7,500.00
1 - 2700k rig $1,500.00
Electricity $350.00 per month = $4,200.00 annually
up keep, upgrade and maintenance $1,000.00 + a year

I fold 1500+ smp and bigadv WU's a month I do fold GPU on occasion but not often. Now some of you say the privileged few are over compensated for their contribution. So what is the cost for those of you that believe the privileged few are over compensated. What is your return for $$$ spent for purely folding and nothing else. Are the people who have taken the time $$$ and effort to build these purely for folding machines over compensated, I do not think so. How many of you have built one how much does it cost you to fold. Most if not all of the people that have taken the time to build these are dedicated folders and then the rest of the folding public want to trash them why ? The bigadv folders have been the most picked on group of folders in the folding community (actually the only picked on group) why ? We are all trying to help

I am starting to take this personally and so are some others out there I am starting to wonder if it is worth it or not. Hell allot of you seem to have nothing better to do than sit around and gripe about how many points your fellow dedicated folder is getting and you are not getting. I think I am done with this for a while I am throughly disgusted with some of my fellow folders. Thanks for the support guys.

[7im]

...
And I'm still interested in hearing about the need to benchmark gpu work unit's on both 48 core gpu's and 448 core gpu's because when you would apply the current smp qrb on gpu work unit's you would see the same problems as we're discussing now with smp except for the fact that we don't have different categories ( yet? ) there. We could reserve that one for later though if you want.

I don't entertain questions that have no bearing on the topic no matter how many times you ask them. But since you won't seem to drop it, I have a simple answer so you can get back on topic... We only have 1 class of GPU work unit at this time. We don't have any BA-16 GPU work units earning 100x the points of normal GPU work units, so it doesn't really apply to this topic.

And if your really want a 48 core GPU benchmark, you can take the 448 core GPU they have now, and disable all of the shaders except 1 grouping. Unfortunately, Stanford can't do that kind of exact benchmarking for BA-16 with the PC hardware they have now.

[7im]

Just curious but where would the incentive be for people to buy the machinery and to run the bigadv WU's if it were normalised across the board. The purpose of the QRB was to encourage quick returns. Guess what it works just look at all the 4P out there right now and there are quite a few being planned and built. Why would I or anyone else run a bigadv WU if we could make the same PPD off of a smp with far less risk.

Just curious where did I say 1 core anything. You comment was unjustified and rude 7im and quit trying to twist thing around.

Sorry, but you said "across the board" which means everything, at least to me it does. Everything includes 1 core systems as much as 16 core systems. And you mentioned QRB, which applies to both 1 core systems and 16 core systems.

Clearly I did not understand you correctly.


And it's NOT personal, although I'm sure ChasR took it a little personally watching 10 years of his folding get wiped in 10 days of BA folding. But progress is disruptive. And it's time to be disruptive again.

Besides, it's not like k1wi is trying to take away your precious QRB. He, and many others want to tweak it slightly so the bonus curve doesn't approach infinity every 18 months as Intel and AMD keeps doubling the number of cores shoved in to their chips.

We've all seen the Intel press releases about 80 core chips coming. What happens to PPD next year when Punchy buys a quad socket for that chip? That kind of wheat and chess problem has to be solved NOW before it gets to the point where the grainery can't count the points fast enough to put them on the stats pages.

Orion asked how many times is this going to come up?! As many times as it takes to solve the wheat and chess problem with the current QRB system. It's not going away unless you can convince Intel to stop double the core counts every 18 months. And running away or ignoring it won't solve it.

I tried to step asside so k1wi could continue. Please let me step asside.

[MtM]

...
And I'm still interested in hearing about the need to benchmark gpu work unit's on both 48 core gpu's and 448 core gpu's because when you would apply the current smp qrb on gpu work unit's you would see the same problems as we're discussing now with smp except for the fact that we don't have different categories ( yet? ) there. We could reserve that one for later though if you want.

I don't entertain questions that have no bearing on the topic no matter how many times you ask them. But since you won't seem to drop it, I have a simple answer so you can get back on topic... We only have 1 class of GPU work unit at this time. We don't have any BA-16 GPU work units earning 100x the points of normal GPU work units, so it doesn't really apply to this topic.

And if your really want a 48 core GPU benchmark, you can take the 448 core GPU they have now, and disable all of the shaders except 1 grouping. Unfortunately, Stanford can't do that kind of exact benchmarking for BA-16 with the PC hardware they have now.

You failed so far to point out the reason why it's not possible for PG to set the base points of BA in line with science. Common sense would dictate it's possible to do so as they are capable of interpreting this scientific value, you are not when you only base it on your emotional reasoning that it's 'not fair'. If it's as simple as using a relative performance indication ( as you would get from disabling clusters groupings on gpu's ), it should be easy enough for them.

And it's NOT personal, although I'm sure ChasR took it a little personally watching 10 years of his folding get wiped in 10 days of BA folding. But progress is disruptive. And it's time to be disruptive again.

Wait a second, we already decided this will not change, we're only going to limit the scale of credit assignments so we don't run into the need to factorize them. We already decided that the relative numbers is not something WE can change, certainly not only because people feel it's unjustified that other people get so much points.

No one has proven QRB as it is now is actually invalid, with the exception of a bigadv project which is getting a very ( and probably to big ) bonus for just making the deadline. That makes it plausible to suggest a normalization between speed_ratio 0 and 0.X to smooth out the initial bonus slightly, nowhere does it make the entire schema wrong.

ChasR would do good to change his perception, which is other then a change to the QRB. And since we're not going to change the relative difference's between bigadv and ChasR's 10 years of folding, your comment seems to make out you're convinced this thread has a different consensus then there actually is.

Proof your point if you think you have one, proof the relative difference between BA and regular folding is so wrong ( without resorting to obvious facts like QRB not working since it's not rolled out completely making the scenario scewed towards BA.. apply the current BA multipliers to GPU and people will be making much more ppd then they are now and maybe not let envy or jalousy cause them to kick against the system which seems so unfair to them ).

If you can I will support you, but it will need more then asking for common sense.

[k1wi]

I would like to sum up that my position, as per my original post, that the point of this discussion is resolve the wheat and chess problem. The only way I see this being possible is by normalising the point system, which require a fundamental shift in how we consider points being earned. By shifting to a normalised system we remove the exponential component and have a tweaked system that looks at relative difficulty at a given point in time. It does not remove any of the proportionality, it simply keeps the points system on the first half of the chess board, perhaps on the first two rows of the chessboard.

It allows PG to continue to set the relative scientific value, but it says "if technological improvements have resulted in computer power improving by x amount, a point should require x amount more to earn." As I have stated from the output that is the overarching concept.

Theoretically, this most simply becomes PPD / Y, where Y can be measured - I cannot imagine that Stanford does not have the appropriate data to do so - and disclosed.

There are important psychological benefits of such a system, because differences of hundreds or thousands of points are much less mind blowing than differences of millions or, eventually, billions, even if the proportional difference remains the same.

By normalising for PPD on a regular basis, it becomes a 'feature' of the community. It will affect all people equally, because all donors will be affected the same by an x% drop in ppd. If done regularly any adjustment could result in relatively small absolute drops in PPD. My proposal affects all client types and projects equally and it does not remove the ability to weight different 'classes' of clients/wu's differently in order to manipulate shifts in 'scientific value'

My proposal notes that there are two issues - that the chessboard problem exists regardless of QRB, and that the QRB continues to be a source of contention. It then attempts to focus on the underlying chessboard problem and provide a solution to the first issue. The more I think about this problem, and the more discussion there is, the more I suspect that what we learn from solving issue one could lead to very real solutions for problem two.

If normalising underlying PPD removes the explosion in absolute growth associated with that aspect, perhaps we can also normalise the QRB system so that it too stops moving along the curve. Doing so will NOT stop the fact that a 4x increase in performance between an 'average/mainstream' computer and a 'top end' computer results in a 9x increase in PPD, but it will account for ever increasing increases in performance of 'average' and 'top end' computers and prevent it from increasing in absolute terms. Investing in a serious piece of hardware will still earn exactly the same proportion more points compared with someone investing in a less powerful piece of hardware, as is designed by the QRB. If people want to discuss the QRB, perhaps we can discuss it in these specific terms?

Onion - as 7im says, threads will continue to be created until the wheat and chess problem is resolved and as long as people care enough about folding to try and resolve the issues, or put forward solutions. Or until the last person gets cut down by those who aren't prepared to hypothesise, suggest and, most importantly, constructively discuss to create solutions to a problem that clearly exists. People once thought the world was flat, and that the heavens rotated around the earth, some of the first people to believe otherwise, that the earth rotated around the sun, were punished for their theories.

Grandpa - I have absolutely no beef with you getting a proportionately high PPD based on your extensive outlay. I HOPE my proposal retains the incentive, but prevents the psychological barrier that exponential increases in absolute values causes. Because I think that is the fundamental issue at heart.

MtM - Normalising PPD over time WILL prevent ChasR's 10 years of folding being wiped out in 10 days... It doesn't change the relative proportional distribution at a given time (or incentive to go large at said point in time), but it WILL change the relationship between different points in time.

[MtM]

let's see...

We make ba make instead off 1000 points, 100 points ( 10%), and smp instead of 100 points, 10 (10%).
We use 10% as a start as it's easy to use.

The next round, we use Y which I described should be based on the computation speed increase. This speed increase should be applied to both trajectories again, so let's assume we have a 10% speedup which have caused ba to be making 1000 points again, and smp 100 points = we normalize down by 10% for both and publish the 10% number so people can still see the how much computational/scientific effort was needed to make earn x credit. How does this influence the relative progression of a ba versus a regular smp instance? The BA instanc will still earn the same amount of points more then the SMP instance relative to the total points obtainable.

This is only to prevent the ppd numbers to get out of hand and needing to be shown in factors. This has nothing to do with making the gap between BA and smp smaller, and I say it should stay that way untill somoene can proof with facts that BA is getting disproportional amounts of ppd because of a fundemental flaw in the QRB concept.

Let's fix the weat and chess problem by the way I proposed.... and let the QRB alone untill someone is able to proof it needs fixing. This does not include the only solution I see possible for the QRB problem ChasR pointed out, a more gradual start of the qrb bonus after meeting the deadline, so anyone who can change the formula in the manner I proposed would make me happy.

And I already gave a method of controlling the upper part of the slope, you need to prevent people from getting there by better controlling which machines get which projects, if you don't you can't set kfactor and deadline so that the spread in speed is in a desireable range. But why do I have to repeat myself all the time?

Edit: to make it crystal clear ->

Doing so will NOT stop the fact that a 4x increase in performance between an 'average/mainstream' computer and a 'top end' computer results in a 9x increase in PPD, but it will account for ever increasing increases in performance of 'average' and 'top end' computers and prevent it from increasing in absolute terms.

You're trying to lower the gap, first start with showing the need to do so.

Proof BA is getting disproportional points, and that PG is not able to set points in accordance with science.

[k1wi]

MtM - Look at the graph on page 1 - the increasing gap between the blue line and the red line illustrates that the wheat and chess board problem also applies to that, and there is a need to apply a similar normalisation to that procedure.

[MtM]

I only said I agreed to fix the problem regarding unreadable numbers because of the increases in computation power along the board, I have no been convinced yet that correction needs to be applied to lower the differences between BA and SMP. I want them to keep reflecting scientific value, and I believe in the statement that time is an important factor now and I am hoping it will become even more relevant in the future.

Look at it this way: SMP when first introduced had an accompanying explanation which elaborated on the importance of smp allowing to simulate longer trajectories in shorter time. Time was important because you need lots of trajectories to find whatever it was a project is looking for, and having longer trajectories also increases the chanche of finding that event in any of the results which are already being submitted quicker then before. Again this relates to time as well, and emphasizes the importance of it.

I expect f@h to become more efficient with quicker returns of longer trajectories not only because of the increase in the possibilitie to encounter a certain event, but also because of a lack of event finding, or other factors like molecular positions after a certain gen not being able to lead to finding an event based on previous data. This will be a perpetual increase, and time is the important driving factor behind this increase.

Time should be rewarded as much as possible.

If people feel their epeen is left in the basement because of BA folders, move BA outside of the regular FAH stats ( create a seperate statistical category for them ).

Edit:

Look I want f@h to grow as much as possible in computational throughput, which may or may not be the same as in user count. If you can proof to me that the project is loosing computational throughput due to the QRB causing people to leave ( and this not being outweighed by the increase in computational throughput from the ba systems! ), then I'll stop defending the high ppd for BA, and only then because it would lower the overall project efficiency.

What do you support? The idea that the stats should be arbitrary based on 'fairness', or the idea that the stats should incite participants to make the project more efficient?

[orion]

Orion asked how many times is this going to come up?! As many times as it takes to solve the wheat and chess problem with the current QRB system. It's not going away unless you can convince Intel to stop double the core counts every 18 months. And running away or ignoring it won't solve it.

I didn't realize that this was intel's doing. All along I thought that it was PG. So here's a novel idea...why don't the complainers take it to an intel forum and lobby intel to stop doubling core counts every 18 months

Where were some of you when BA’s and their QRB were implemented back in '09? I don’t recall one iota about this back then but now it’s a problem. All I heard back then was how slow 2p and x6’s where slowing the science down and that the shortages of BA's were because of them. Nothing about points being too great for fast rigs that did BA's in a short amount of time.

So what's next? People folding on HPCS.

[Rattledagger]

But there is one thing that is for sure the points system works Stanford's F@H project is growing the number of people participating grows on a daily basis. Some people come here and say XX% of there team is quitting F@H and moving to other projects but if you check their stats their active membership is actually going up and in most cases so are their points. People have come and gone since the beginning of the project and the always will the important thing is that they continue to grow and guess what the stats say they are.

While number of users that once-upon-a-time ran FAH increases daily, the only place I can see anything about currently active users is KakaoStats, and appart for an extra dip around xmas-holidays #active has been basically a flat line for AFAIK atleast 6 months now, oscillating between roughly 63k - 65k active users. Now of course the number of active users from KakaoStats can be wrong, but atleast I've not seen anything that indicates FAH for the last 6 months has been adding more new users per day than they're losing old users per day.

Where were some of you when BA’s and their QRB were implemented back in '09? I don’t recall one iota about this back then but now it’s a problem. All I heard back then was how slow 2p and x6’s where slowing the science down and that the shortages of BA's were because of them. Nothing about points being too great for fast rigs that did BA's in a short amount of time.

In my opinion FAH's point-system was FUBAR before the introduction of QRB, so didn't see much point discussing it in 2009, and still doesn't see much point discussing it now.