Bigadv points change

[JPinTO]

First, it's simple. Single modification to formula, single data point modified for each WU.
Second, it's universal. Applies to all clients and fits into current unified benchmark scheme.
Lastly, and perhaps most important, it provides the necessary data needed to form a basis for statistical analysis.

For example, here is a graph showing the differences between the two formulas:


Given the above, the question is fairly straightforward. Which of these curves best describes the relationship between scientific value and time, the red line or the green one?

Thank you for putting a concrete proposal down. You propose a small formula tweak to the existing system which was what Vijay Pande earlier requested.

Your last question is what it boils down to: Which curve does PG team best describes the science value?

If the existing SQRT curve better fits the science done, then GPU/Single socket folding is now old tech, just like single threading CPU's&PS3 before it. Moore's law marches us along and the protein folding show must go on. New folding farms require higher investment in fewer multi-socket enterprise class machines that favour less WU's being done much quicker.

If CUBERT or some linear formula better fits the science, then the folding farm comprised of lots of cheap consumer grade hardware is still a viable methodology.

Its really only an answer that Pande group can provide. Until that answer is forthcoming, I'm on the sidelines with failing older gear watching my previous top 50 project position erode as multi-socket enterprise class machines pass me like I'm driving a Prius on the Autobahn.

- JP

[7im]

@ Amaruk, Thank you for posting your new comparitive graph. It answers the question I was hoping you would voice. Far left or far right side of the graph doesn't really make a difference. One could just as easily swap the axis on the same graph. It still shows the extreme scaling of points.

Would you please post another version of the dual graph, where to top Y points value is only 4 million (about where the new curve stops)? It will add more "resolution" to the graphed values at more than 12 hours. That's the general area of the graph we're looking at today and in the near feature. Thanks again.


There are other issues that cubert does not solve, but it does a very good job of illustrating how small tweaks can make small improvements without scrapping the whole system.

For example, there is no standard definition for setting the deadline on -bigadv workunits as there was with SMP and CPU work units, and hence the points tend to float up and down a bit from one project to the next. The K value can compensate for that, but it should still be standardized to help keep the points system consistent.

It also doesn't solve the issue of how the new shorter deadlines on fahcore_a4 (CPU/SMP) work units are causing part time or slower folders to miss deadlines they once made very easily on fahcore_78 (CPU) work units. If nothing changes, it significantly raises the minimum specs needed to contribute once all of the fahcore_78s are retired, from s P3-500 to about a P4-1.8 system. Double that if you only fold part time.

I've voiced my positions on several aspects of the current points system. And PG is meeting to discuss various issues. They made change a lot, a little, or nothing at all, but at least they are having that conversation.

[Napoleon]

Please-oh-please, some mod lock this topic. From someone who doesn't have a stake in it, but follows this forum closely anyway, and folds too... here's a clue: ULTRA-BORING.

[MtM]

Then stop posting and just read

[bruce]

By definition, ANY time bonus whatsoever will result in an exponential increase in PPD

That's simply not true.

I'm not suggesting that this is a better system than the present one -- it's just an example that I don't really recommend -- but it's really easy for the Pande Group to decide (for instance) that if anyone can complete a bigadv WU in less that 1 day, there's no scientific value in being faster than that, so the add a cap that ignores return times faster than one day. For fast/big systems that can complete projects in less than a day, they come up with projects that can only be assigned to those systems and they adjust the baseline points by some amount that does NOT become exponential.

The QRB doesn't have to be represented by a single formula that has a sqrt and an inverse relationship in it. We've lived with SMP WUs that follow the current QRB formula for a while and the Uniprocessor WUs which have followed the no-bonus rules. Who's to say the PG can't institute a third system? ... or find a way to adjust the present system?

[Napoleon]

Please consider my previous post as motivated by sheer frustration.
I'll try to be a little bit constructive this time.

Let's consider the formula k*deadline_length / elapsed_time. Cuberoot would only postpone the inevitable. Can't divide by zero, so how about throwing in one more parameter, call it P (Priority, Parameter, Pfatever)? (k*deadline_length) / (P + elapsed_time)

P = 0 is how things are now.

Code: Select all

P = 0 x deadline_length / 100
P = 1 x deadline_length / 100
P = 2 x deadline_length / 100
P = 3 x deadline_length / 100
...

P = 0 would still leave the door open to those people who are willing to invest in the latest and greatest tech and dedicate it to FAH, pay for the wattage, and all that. I have great respect to those people, and I think they deserve to be on a divide-by-zero curve.

However, for us mere mortals, P > 0
<idealism>
In a more perfect world, P is something you could choose in the v7 client, to match your hardware and level of dedication. I believe the extra Parameter would give Pande group a better way to communicate to more or less casual donors just what are the PG Priorities. IMHO, prolonged arguments about the value of quick returns are useless until QRBs get extended to all WUs, not just the select few. Basically, I'd like to see much smarter way of WU assignment.
</idealism>

I sure hope I explained my take on this topic a bit better this time.

[Amaruk]

IMHO, prolonged arguments about the value of quick returns are useless until QRBs get extended to all WUs, not just the select few.

They're working on it. I get QRB qualified Uni clients on a regular basis.

...more "resolution" to the graphed values at more than 12 hours. That's the general area of the graph we're looking at today and in the near feature.

Is this what you have in mind?



A quick look at the above graph reveals a serious flaw. The lines never connect. In fact, they do not connect at any point all the way back in time to the bonus cutoff of four days.

First, this goes against PG's normal practice of anchoring the new formula to the old.

Second, this means everyone loses. ALL SMP folders lose points, and PG loses folders. Not the desired outcome.



The solution is simply to adjust base points. But where should the two curves intersect?

The point I chose was the 36 hour mark. No, I didn't chose it because that's where the graph ends. I chose this point because it matches up with a TPF of about 22 minutes, which roughly corresponds to the performance of a 970/980X. They are both fairly common enthusiast level chips, and they qualify for the newer bigadv WUs. This would make them a popular choice for bigadv, so it's reasonable to think there are a number of them out there folding.

This sets the new base points for 6901 at 19,190 and results in the following graph:



Those folders who's completion time falls within the range of the above graph would average about 15,000 PPD less. (-11%)

The PPD for those on the right side (36 hours) would be unchanged, while a folder on the left side (12 hours) would get 261,637.11 PPD instead of 318,777.88 PPD, 57,140.77 less. (-18%)


The faster folders who fall beyond the left side (completing 6901 in less than 12 hours) would average 877,500 PPD less. (-39%)

If a folder did manage to complete 6901 in one hour, they would get 7,187,993.53 PPD instead of 13,251,347.45 PPD, 6,063,353.92 less. (-46%)

Just for reference, that's a TPF of 00:00:36, not accounting for download/upload time.


Because of the different slope between the formulas, folders off the right side (36 hours+) would get a small bonus, averaging about 1,800 PPD. (+6%)


This is the full graph covering completion times less than 36 hours.



I will not be readily available over the next several days, so I have decided attach the data I used to create these graphs.
This will allow anyone to recreate these or customize your own if desired.

Code: Select all

Column A: number of hours to complete Project 6901

~~~

Column B: PPD based on existing formula and current base points

bigadv 6901, A5, 26.4k, 8,955 points, 4/6 days

final_points = base_points * max(1,squareroot(k*deadline_length/elapsed_time))

~~~

Column C: PPD based on cuberoot formula and 2.14X base points (19,190)

bigadv 6901, A5, 26.4k, 19,190 points, 4/6 days

final_points = base_points * max(1,cuberoot(k*deadline_length/elapsed_time))

Intersect point is at approximately 36 hours.

~~~

Column D: PPD based on cuberoot formula and 1.28X base points (11,491)

bigadv 6901, A5, 26.4k, 11,491 points, 4/6 days

final_points = base_points * max(1,cuberoot(k*deadline_length/elapsed_time))

No intersect point. Bad idea.




A	     B		     C		     D

1	13251347.45 7187993.53	4304187.74
2	4685058.82	2852557.13	1708118.04
3	2550223.01	1661292.12	994785.70
4	1656418.43	1132038.04	677867.09
5	1185236.55	840712.82	503420.85
6	901639.99	659284.22	394780.97
7	715505.51	536797.27	321435.49
8	585632.35	449249.59	269011.73
9	490790.65	383958.53	229915.29
10	419044.40	333637.11	199782.70
11	363221.05	293821.85	175941.23
12	318777.88	261637.11	156668.93
13	282712.50	235152.65	140809.97
14	252969.40	213028.14	127561.76
15	228098.88	194305.90	116350.84
16	207052.30	178284.82	106757.38
17	189054.34	164440.62	98467.44
18	173520.70	152374.04	91241.94
19	160003.56	141776.05	84895.84
20	148154.57	132403.98	79283.82
21	137699.10	124064.82	74290.31
22	128418.04	116603.27	69822.32
23	120134.65	109893.13	65804.27
24	112705.00	103830.76	62174.11
25	106010.78	98330.37	58880.46
26	99953.96	93320.38	55880.47
27	94452.70	88740.66	53138.12
28	89438.19	84540.28	50622.92
29	84852.19	80675.88	48308.91
30	80645.13	77110.35	46173.86
31	76774.59	73811.73	44198.64
32	73204.00	70752.37	42366.69
33	69901.92	67908.23	40663.61
34	66840.80	65256.92	39076.83
35	63996.76	62784.18	37595.32
36	61348.83	60469.68	36209.39
37	58878.60	58300.47	34910.46
38	56569.80	56263.87	33690.94
39	54508.05	54348.58	32544.06
40	52380.55	52544.55	31463.80
41	50475.92	50842.77	30444.77
42	48683.98	49235.16	29482.13
43	46995.62	47714.42	28571.51
44	45402.63	46274.05	27709.01
45	43897.65	44908.07	26891.06
46	42474.01	43611.11	26114.44
47	41125.70	42378.34	25376.25
48	39847.23	41205.26	24673.81
49	38633.67	40087.87	24004.71
50	37480.47	39022.44	23366.73
51	36383.53	38005.59	22757.84
52	35339.06	37034.22	22176.18
53	34343.63	36105.48	21620.05
54	33394.07	35216.76	21087.88
55	32487.48	34365.61	20578.21
56	31621.17	33549.83	20089.72
57	30792.70	32767.34	19621.16
58	29999.78	32016.24	19171.40
59	29240.31	31294.76	18739.38
60	28512.36	30601.26	18324.11
61	27814.12	29934.22	17924.68
62	27143.92	29292.20	17540.24
63	26500.20	28673.92	17170.01
64	25881.54	28078.09	16813.23
65	25286.57	27503.61	16469.23
66	24714.06	26949.41	16137.37
67	24162.83	26414.42	15817.02
68	23631.79	25897.78	15507.65
69	23119.92	25398.55	15208.71
70	22626.27	24915.92	14919.71
71	22149.94	24449.12	14640.19
72	21690.09	23997.42	14369.71
73	21245.93	23560.11	14107.85
74	20816.73	23136.56	13854.23
75	20401.78	22726.16	13608.48
76	20000.44	22328.33	13370.26
77	19612.09	21942.53	13139.24
78	19236.15	21568.25	12915.12
79	18872.06	21205.00	12697.60
80	18519.32	20822.26	12486.42
81	18177.43	20509.77	12281.30
82	17845.93	20176.97	12082.02
83	17524.39	19853.49	11888.32
84	17212.39	19538.98	11699.99
85	16909.53	19233.09	11516.82
86	16615.46	18935.48	11338.61
87	16329.81	18645.85	11165.18
88	16052.25	18363.87	10996.33
89	15782.47	18089.27	10831.90
90	15520.16	17821.77	10671.72
91	15265.04	17561.14	10515.65
92	15016.83	17307.08	10363.52
93	14775.28	17059.40	10215.21
94	14540.13	16817.85	10070.57
95	14311.15	16582.23	9929.48
96	14088.12	16352.32	9791.81

[Amaruk]

Green and red still have to cover the sempron running classic client on the one side and bigadv on the other side?

These charts are graphing the PPD of a singe bigadv WU (6901) over time. The horizontal axis is the number of hours taken to complete this WU.

Red line is current PPD, green line is predicted PPD.

How do you think that affects the changes not on the extreme left ( in above example ) but the range where you would expect the largest part of the donors would be?

I did have to make an adjustment regarding this. Please see previous post.

Would it still give enough bonus to quicker return times?

Excellent question, one that PG will ultimately have to answer.

No one seems to really give me an answer to my questions, I must be either just so wrong no one feels called to point out the obvious or I'm not making sense to anyone but myself at this point

I hope I have not caused you to feel this way. FWIW, my answer to your third question is a perfect example that Grandpa_01 is right.

You forgot #3 nobody knows the answer to the question.

[MtM]

Long post in here yesterday, didn't post it then, changed this post atleast 10 times already, so beware I'm going to maybe jump from one thing to the other without making much sense.

I need to say I'm sorry, I don't want to post negative things in this thread I just felt like either no one understood me or like I was so wrong people were ridiculing me by not answering

The issue is there are atleast 3 people in this thread who are here to get others to talk, and while I should have understood that much much sooner it does make it hard right now to decide what to post and what not. I don't think I can post anything which isn't already known to those 3, posted publicly or in pm's at some time before or during this thread. In that light, I'm going to try and limit my input to this thread to the tool I'm working on. My views are known and if not clear I'll try and explain but I don't think I'll get allot of questions.

[Jester]

Well I'm glad I read through all of this and will now postpone any further hardware changes or upgrades until there is a clear direction of where the "Bigadv" project is heading,
When it was first released there were fairly steep hardware requirements and the "carrot" of big bonus points to encourage the uptake, hopefully this was totally for the science
of having these Wu's returned as fast as possible, time doesn't have a "proportional sliding scale" in value as some seem to think if it is concerned with a potential "life saving discovery"
in the future,
On a shorter time scale, say a $10k ambulance can reach the edge of it's service area in 30 minutes, but a $100k ambulance can do it in 29 minutes 10x the expenditure is way out of
proportion to the performance, unless you are the poor soul who needs that minute..... (an overly simplistic example, but you get the idea ?)
If projects like Bigadv are to be released it may be prudent to allow them to run unchanged for a set period stated in advance, only after that time can it then be changed due to "public opinion",
at least that way the onus is on the team responsible to consider the possible impact of high bonus points before release, and those deciding to take up the challenge can have some guarantee
that expensive hardware will not not be suddenly "devalued",
I started Folding when there were some Wu's that were worth 0.6 points and "big ones" were 4.5 points, but I don't consider their value as less now I'm running a machine capable of 100k+ ppd, that was
the point value of that Wu at that time, nothing more,
To those who remember that far back I also went through both the QMD and GPU1 projects, and if history repeats with the Bigadv project it shows nothing has been learned.
Ps: While it's true I love the "carrots" don't assume that I'm automatically a Donkey....

[MtM]

Not finished, some bug's and something just not working right ( one of the formula's ). the nice thing is one can pan and zoom, though it does only use the number of points you started with so it's not really accurate. Ideally it should recalculate for each zoom level. Can change project info details in the browser ( which get's the info from the summary of an url you enter manually ). Will add a download link when I'm really done. Sorry, said I would be quicker but sometimes things just don't go as easy as I wanted them to Doesn't store the point pairs to db either, and doing lots of projects at the same time is rather slow.

Without the comparison, even just this view should support the issue with 'one formula for all'.

[MtM]

i had edited this in the old post but I wouldnt mind the thread bump giving me a quicker shot at getting a math issue fixed

Edit:

I said my math is quite bad, so here goes:

Code: Select all

            Dim iKfactor As Double = CDbl(Projects.Project(Project).kFactor.Replace(".", ","))
            Dim iPworth As Double = CDbl(Projects.Project(Project).Credit.Replace(".", ","))
            Dim pCompletiontime As TimeSpan = EndTime.Subtract(BeginTime)
            If iKfactor > 0 Then
                'check if eta is before preferred
                If EndTime < BeginTime.AddDays(Projects.Project(Project).PreferredDays.Replace(".", ",")) Then
                    Dim bMulti As Double = Math.Sqrt((Projects.Project(Project).PreferredDays.Replace(".", ",") * iKfactor) / pCompletiontime.TotalDays)
                    iPworth = Math.Round(iPworth * bMulti)
                End If
            End If
            'iPworth = iPworth / 100.0F
            'How many frames per 24/h
            Dim iPPD As Double = 0
            Dim tsDay As TimeSpan = TimeSpan.FromDays(1)
            Do
                If tsDay.Subtract(pCompletiontime).TotalSeconds >= 0 Then
                    iPPD += iPworth
                    tsDay = tsDay.Subtract(pCompletiontime)
                Else
                    Exit Do
                End If
            Loop
            'get fraction of _tsFrame to be done in remaining seconds
            Dim iRfraction As Double
            If tsDay.TotalSeconds > 0 Then
                iRfraction = tsDay.TotalSeconds / pCompletiontime.TotalSeconds
                iPPD += iRfraction * iPworth
            End If
            Return (Math.Round(iPPD, 2))

This works ( has done for a long time ), but the other formula isn't

Code: Select all

 Dim iKfactor As Double = CDbl(Projects.Project(Project).kFactor.Replace(".", ","))
            Dim iPworth As Double = CDbl(Projects.Project(Project).Credit.Replace(".", ","))
            Dim pCompletiontime As TimeSpan = EndTime.Subtract(BeginTime)
            If iKfactor > 0 Then
                'check if eta is before preferred
                If EndTime < BeginTime.AddDays(Projects.Project(Project).PreferredDays.Replace(".", ",")) Then
                    'Dim bMulti As Double = Math.Sqrt((Projects.Project(Project).PreferredDays.Replace(".", ",") * iKfactor) / pCompletiontime.TotalDays)
                    Dim bMulti As Double = Math.Pow(Projects.Project(Project).PreferredDays.Replace(".", ",") * iKfactor / pCompletiontime.TotalDays, 1 / 2)
                    iPworth = Math.Round(iPworth * bMulti)
                End If
            End If
            'iPworth = iPworth / 100.0F
            'How many frames per 24/h
            Dim iPPD As Double = 0
            Dim tsDay As TimeSpan = TimeSpan.FromDays(1)
            Do
                If tsDay.Subtract(pCompletiontime).TotalSeconds >= 0 Then
                    iPPD += iPworth
                    tsDay = tsDay.Subtract(pCompletiontime)
                Else
                    Exit Do
                End If
            Loop
            'get fraction of _tsFrame to be done in remaining seconds
            Dim iRfraction As Double
            If tsDay.TotalSeconds > 0 Then
                iRfraction = tsDay.TotalSeconds / pCompletiontime.TotalSeconds
                iPPD += iRfraction * iPworth
            End If
            Return (Math.Round(iPPD, 2))

Math.pow(....,2) is the same as math.sqrt(...) but my head hurts and I can't get the formula cuberoot translated in pow....

Anyone willing to post how to solve this will get a big notice in the 'about'

[Napoleon]

Math.pow(....,2) is the same as math.sqrt(...) but my head hurts and I can't get the formula cuberoot translated in pow....

Not entirely sure if this is what you're looking for.


And so on.

EDIT: If I understood the code right, the cuberoot you're looking for is "Math.Pow(..., 1/3)".

[MtM]

Thank you! That works as it should

Don't know why but for some strange reason I only tried even numbers

I didn't do anything about writing a dump to db for the datapoints yet ( or export to cvs/xml ), my axis are reversed and my scale is based on fractional hours, still some things to do. I'll upload current work to a public repository later this day, want to have a go at making it more functional.

[MtM]

source and binary in bin for those interested -> http://code.google.com/p/cftunity/downl ... p&can=2&q=

It's just the graph, no legend/details. Could be added to save all step info to csv/xml, might have to edit the total ticks being used to get the interval to something which snaps to the grid more ( like the closest minute if the total minutes is large enough ). There are bugs but you can work around them to get a graph:

select the project(s) you want on the left list, press the button with the >>>> on it to move them to the active project list, the projects will be marked with an * on the end. You need to attach a min and max timespan for each project, but you can use multiselect if you need/want to. The numeric updown controls are wrongly placed and show only one number instead of two. The first control is days, then hours, then minutes. A validation is done to check if there is a valid timespan between those two, and pressing save will prepare projects for the points calculation ( the checked list box doesn't do anything, leave them unchecked as both formula's are always used ). After pressing save, the go button will be endabeld. After pressing go, right click on the graph and 'zoom to default'.

Edit:

Not so usefull in current form with allot of projects, need to edit the placement of the index on another form/graphpane ( think that is possible but not sure how yet ).