Folding Forum

Posted: **Sat Jul 19, 2014 10:04 pm**

Things seem pretty quiet now, so I'll stir the waters a bit by proposing a modification to the QRB system. I tried to keep the formula simple, yet reasonably closely matched with the existing QRB scheme. Without further ado, here's what I came up with, using P9201 Core17 project as an example: http://1drv.ms/1n41wNJ

You're better off downloading the spreadsheet and trying it out yourself, but I'll try to elaborate the key points in greater detail here.

Base points B, timeout T and expiration E remain the same, but K-factor is replaced with a bonus points multiplier cap C which can be used to adjust the points curve - not too different from adjusting K
Maximum Points Per WU = PPWU = B * C
Let's yank back the concept of performance fraction p from v6 client. Let p = % of Timeout to complete a WU. Converting between p and TPF (minutes) is pretty straightforward
The new proposed points formula per WU:
- PPWU = B*(1 +(C-1)*(1-p/100%) for p values [0%, ..., 100%],
- If( p*T/100% <= E ) then PPWU = B/2 else PPWU = 0 for p > 100%
  - I figured we might just set PPWU = B/2 after Timeout, so that there's some sort of actual penalty for forcing some other(s) to complete the very same WU you worked on

There are some obvious benefits:

It is no longer possible to gain an insane jump in points by completing a single WU, no matter how fast your brand new quantum computing folding wonder gets
However, PPD still approaches infinity if the combination of your HW & internet connection & Stanford server speed is so fast that the overall processing time of WUs approaches zero

Summarizing the practical example from the spreadsheet (P9201, B=8000, T=2.6, E=3.4, K=0.75 / C=4.25):

Code: Select all

p	PPD prop.	PPD current	WUsPerDay	TPF	PPWU prop.
0%	#DIV/0!	#DIV/0!	#DIV/0!	0.00	34000
1%	1297692	3047194	38.4615	0.37	33740
2%	643846	1077346	19.2308	0.75	33480
3%	425897	586433	12.8205	1.12	33220
4%	316923	380899	9.6154	1.50	32960
5%	251538	272549	7.6923	1.87	32700
6%	207949	207335	6.4103	2.25	32440
7%	176813	164533	5.4945	2.62	32180
8%	153462	134668	4.8077	3.00	31920
9%	135299	112859	4.2735	3.37	31660
10%	120769	96361	3.8462	3.74	31400
11%	108881	83524	3.4965	4.12	31140
12%	98974	73304	3.2051	4.49	30880
13%	90592	65011	2.9586	4.87	30620
14%	83407	58171	2.7473	5.24	30360
15%	77179	52452	2.5641	5.62	30100
16%	71731	47612	2.4038	5.99	29840
17%	66923	43474	2.2624	6.36	29580
18%	62650	39902	2.1368	6.74	29320
19%	58826	36793	2.0243	7.11	29060
20%	55385	34069	1.9231	7.49	28800
21%	52271	31664	1.8315	7.86	28540
22%	49441	29530	1.7483	8.24	28280
23%	46856	27625	1.6722	8.61	28020
24%	44487	25917	1.6026	8.99	27760
25%	42308	24378	1.5385	9.36	27500
26%	40296	22985	1.4793	9.73	27240
27%	38433	21720	1.4245	10.11	26980
28%	36703	20567	1.3736	10.48	26720
29%	35093	19512	1.3263	10.86	26460
30%	33590	18545	1.2821	11.23	26200
31%	32184	17655	1.2407	11.61	25940
32%	30865	16834	1.2019	11.98	25680
33%	29627	16074	1.1655	12.36	25420
34%	28462	15370	1.1312	12.73	25160
35%	27363	14716	1.0989	13.10	24900
36%	26325	14107	1.0684	13.48	24640
37%	25343	13539	1.0395	13.85	24380
38%	24413	13008	1.0121	14.23	24120
39%	23531	12511	0.9862	14.60	23860
40%	22692	12045	0.9615	14.98	23600
41%	21895	11607	0.9381	15.35	23340
42%	21136	11195	0.9158	15.72	23080
43%	20411	10807	0.8945	16.10	22820
44%	19720	10440	0.8741	16.47	22560
45%	19060	10094	0.8547	16.85	22300
46%	18428	9767	0.8361	17.22	22040
47%	17823	9457	0.8183	17.60	21780
48%	17244	9163	0.8013	17.97	21520
49%	16688	8884	0.7849	18.35	21260
50%	16154	8619	0.7692	18.72	21000
51%	15641	8367	0.7541	19.09	20740
52%	15148	8126	0.7396	19.47	20480
53%	14673	7897	0.7257	19.84	20220
54%	14217	7679	0.7123	20.22	19960
55%	13776	7471	0.6993	20.59	19700
56%	13352	7271	0.6868	20.97	19440
57%	12942	7081	0.6748	21.34	19180
58%	12546	6899	0.6631	21.72	18920
59%	12164	6724	0.6519	22.09	18660
60%	11795	6557	0.6410	22.46	18400
61%	11438	6396	0.6305	22.84	18140
62%	11092	6242	0.6203	23.21	17880
63%	10757	6094	0.6105	23.59	17620
64%	10433	5952	0.6010	23.96	17360
65%	10118	5815	0.5917	24.34	17100
66%	9814	5683	0.5828	24.71	16840
67%	9518	5556	0.5741	25.08	16580
68%	9231	5434	0.5656	25.46	16320
69%	8952	5317	0.5574	25.83	16060
70%	8681	5203	0.5495	26.21	15800
71%	8418	5093	0.5417	26.58	15540
72%	8162	4988	0.5342	26.96	15280
73%	7914	4886	0.5269	27.33	15020
74%	7672	4787	0.5198	27.71	14760
75%	7436	4691	0.5128	28.08	14500
76%	7206	4599	0.5061	28.45	14240
77%	6983	4510	0.4995	28.83	13980
78%	6765	4423	0.4931	29.20	13720
79%	6553	4340	0.4869	29.58	13460
80%	6346	4259	0.4808	29.95	13200
81%	6144	4180	0.4748	30.33	12940
82%	5947	4104	0.4690	30.70	12680
83%	5755	4030	0.4634	31.08	12420
84%	5568	3958	0.4579	31.45	12160
85%	5385	3888	0.4525	31.82	11900
86%	5206	3821	0.4472	32.20	11640
87%	5031	3755	0.4421	32.57	11380
88%	4860	3691	0.4371	32.95	11120
89%	4693	3629	0.4322	33.32	10860
90%	4530	3569	0.4274	33.70	10600
91%	4370	3510	0.4227	34.07	10340
92%	4214	3453	0.4181	34.44	10080
93%	4061	3398	0.4136	34.82	9820
94%	3912	3344	0.4092	35.19	9560
95%	3765	3291	0.4049	35.57	9300
96%	3622	3240	0.4006	35.94	9040
97%	3481	3190	0.3965	36.32	8780
98%	3344	3141	0.3925	36.69	8520
99%	3209	3108	0.3885	37.07	8260
100%	3077	3077	0.3846	37.44	8000
101%	1523	3046	0.3808	37.81	4000
102%	1508	3017	0.3771	38.19	4000
103%	1494	2987	0.3734	38.56	4000
104%	1479	2959	0.3698	38.94	4000
105%	1465	2930	0.3663	39.31	4000
106%	1451	2903	0.3628	39.69	4000
107%	1438	2876	0.3595	40.06	4000
108%	1425	2849	0.3561	40.44	4000
109%	1411	2823	0.3529	40.81	4000
110%	1399	2797	0.3497	41.18	4000
111%	1386	2772	0.3465	41.56	4000
112%	1374	2747	0.3434	41.93	4000
113%	1361	2723	0.3404	42.31	4000
114%	1350	2699	0.3374	42.68	4000
115%	1338	2676	0.3344	43.06	4000
116%	1326	2653	0.3316	43.43	4000
117%	1315	2630	0.3287	43.80	4000
118%	1304	2608	0.3259	44.18	4000
119%	1293	2586	0.3232	44.55	4000
120%	1282	2564	0.3205	44.93	4000
121%	1271	2543	0.3179	45.30	4000
122%	1261	2522	0.3153	45.68	4000
123%	1251	2502	0.3127	46.05	4000
124%	1241	2481	0.3102	46.43	4000
125%	1231	2462	0.3077	46.80	4000
126%	1221	2442	0.3053	47.17	4000
127%	1211	2423	0.3028	47.55	4000

Low-end GPUs gain a little, but nowhere near enough to discourage the use of v6 client to finish off Core_15 work only
Mid-to-High range of GPUs get a more substantial PPD boost, which might encourage some additional people to fold on their GPUs, or put in an additional (relatively cheap) dedicated folding GPU
Today's top-notch cards shouldn't suffer much, if at all. The "hurt" starts at p < 6% in this example, equivalent to TPF < 2.25min (2min 15s) and about 207k PPD on a P9201 WU
- Even if the current top-notch (GPU) folders might get stung a little bit now, you could consider this kind of scheme an investment in the future. The proposed PPD curve in the 0% - 10% pic may be a bit misleading, since PPD gain is practically linear for p <= 8%. For each halving of p (and thus TPF), the PPD gain is [2.07, 2.03, 2.02]. With this proposed scheme, the upcoming top-notch Maxwell GPUs are less likely to blow your current top-notch Kepler GPUs completely out of water just because an uncapped PPD curve is getting crazier and crazier...

Just for fun, I set up a poll for you... comments and votes are most welcome.

Posted: **Sun Jul 20, 2014 12:13 pm**

Thanks for your efforts !
One question: Why would we still need C (or today k) ? Wouldn't it be easier in incorporate that direct in a higher base point B ? Faster hardware still get a better PPWU based on smaller p.

Posted: **Sun Jul 20, 2014 12:31 pm**

Why is a change required at all? How much productivity would be lost from non-"power Folders" getting less recognition for their volunteer efforts?

Posted: **Sun Jul 20, 2014 9:27 pm**

ChristianVirtual wrote:One question: Why would we still need C (or today k) ? Wouldn't it be easier in incorporate that direct in a higher base point B ? Faster hardware still get a better PPWU based on smaller p.

Without expressing an opinion about the proposed change, the theory behind the original QRB plan was related to some benchmarking results related to how well the SMP cores scaled. Suppose your machine has 16 cores and that with 4 CPUs, a particular WU earns X PPD. Running four CPU clients (or four CPU slots) will earn 4*X PPD. Without any QRB, running a single client (or slot) using all 16 cores will earn maybe 3.5*X PPD and will produce MORE (better) science. Somehow QRB needs to adjust the 3.5*X to EXCEED 4.0*X sufficiently to encourage people not to run multiple CPU clients/slots on a single system.

It's reasonable to question how much actual adjustment makes sense, and without launching a long discussion about the actual hardware, let's just say that the observed benchmark data varies, depending on the actual hardware.

Even reasonable people will argue about the details, but in the final analysis (several years ago) the Pande Group decided that none of the options proposed then would have avoid endless arguments nor would they be universally seen as improvements.

Posted: **Sun Jul 20, 2014 10:28 pm**

bruce wrote: Even reasonable people will argue about the details, but in the final analysis (several years ago) the Pande Group decided that none of the options proposed then would have avoid endless arguments nor would they be universally seen as improvements.

As the saying goes- if it ain't broke, don't fix it.

The current system may be a bit bent and battered in places, but it isn't broken!

Posted: **Mon Jul 21, 2014 3:27 am**

But I think the system will break eventually, just like any other system which goes without maintenance for extended periods will. Presuming you have a car, you do change the engine oil in it regularly, don't you? Because changing it after the engine has ground to a halt won't fix anything. I'd like to think what I'm proposing here is commonsense maintenance, as opposed to a fix. I'm not saying my particular brand of oil is the best choice, or even a good choice. What I'm suggesting is that choosing inaction over and over again is worse.

Let's face it, the current points system has a potential point of failure. Because there's no PPWU limit, theoretically someone could instantly leap from zero to superhero, points wise, by folding just one WU really, really, really fast. Here's a thought experiment for you: some folding newbie finishes his 1st WU ever for 50000000000 points later today and then decides folding is not for him. What would that do to the credibility of the points system and FAH as a whole?

I know such a thing is highly unlikely, but something like that might happen one day. The mere possibility is enough to keep me awake at night...

Posted: **Mon Jul 21, 2014 3:55 am**

So, even if it did happen, what is the perceived problem? It might even spur several others to try to replicate the feat or "break the record", adding to the Folding pool!

F@H points are not "worth" anything, except for the friendly competition they foster. As for those few teams who reward their Folders with outside prizes of whatever description, I am happy that the Pande Group does NOT feel the need to accommodate those specifics.

Posted: **Mon Jul 21, 2014 6:30 am**

ChristianVirtual wrote:One question: Why would we still need C (or today k) ? Wouldn't it be easier in incorporate that direct in a higher base point B ? Faster hardware still get a better PPWU based on smaller p.

I'm not sure if I understood you right, but simply removing k (or C) and boosting base points would eliminate QRB altogether. Without either, PPWU would be just the base points, and the whole concept of QRB is about adjusting base points somehow. Perhaps the adjustment wouldn't even have to be a smooth curve. Might just as well define something like this:

for 0% <= p < 20% PPWU = 32 * B
for 20% <= p < 40% PPWU = 16 * B
for 40% <= p < 60% PPWU = 8 * B
for 60% <= p < 80% PPWU = 4 * B
for 80% <= p < 100% PPWU = 1 * B
for 100% <= p < Expiration/Timeout * 100% PPWU = B
for Expiration/Timeout * 100% <= p PPWU = 0

That'd define 5 classes within the Best Practices guide, and there would be a clear distinction between the classes (whether or not the distinction needs to be exponential is debatable). Then the competition within a class would be linear, proportional to p only.

Alternatively, interpolate between the classes:

Maybe I'm just a worrywart, but still... the thought of PPWU approaching infinity when p approaches zero gives me the willies.

Posted: **Mon Jul 21, 2014 8:29 am**

Napoleon wrote:Here's a thought experiment for you: some folding newbie finishes his 1st WU ever for 50000000000 points later today and then decides folding is not for him. What would that do to the credibility of the points system and FAH as a whole?

Probably no more than bigadv did… the scenario you outline isn't greatly different to what I saw when I first started folding with a PPD around 15,000- others with PPDs in the hundreds of thousands or more who I simply couldn't compete with (until I discovered how to join them

). It caused me to have reservations about the way user ranking is presented, which I still have, but no more than that.

Napoleon wrote: I know such a thing is highly unlikely, but something like that might happen one day.

Let me present you with a thought experiment- a new algorithm evolves which scales much faster with increasing numbers of cores (cpu or gpu) than the current one. That would break the current QRB calculation and a new one would be needed but until then it's just playing with definitions of "fair".

Napoleon wrote: The mere possibility is enough to keep me awake at night...

Doesn't bother me, good luck to 'em

Posted: **Mon Jul 21, 2014 3:16 pm**

jrweiss wrote:So, even if it did happen, what is the perceived problem? It might even spur several others to try to replicate the feat or "break the record", adding to the Folding pool!

Snip.

How many does that inspire, and how many does that discourage? Net gain or net loss? Don't forget to look at the bigger picture while you examine a small portion of the picture.

And with ocores coming, QRB change is inevitable (more of that big picture stuff).

Posted: **Mon Jul 21, 2014 5:35 pm**

billford wrote:Let me present you with a thought experiment- a new algorithm evolves which scales much faster with increasing numbers of cores (cpu or gpu) than the current one. That would break the current QRB calculation and a new one would be needed but until then it's just playing with definitions of "fair".

Umm... more or less my point, minus a desperate attempt to do something about it?

Being able to span a single WU across multiple GPUs is old news already, it just doesn't scale well at the moment. But that could change "soon". There have also been some hints that CUDA JIT could speed up things a lot, without need for a brute increase in HW power. Considering P9201 as an example, top-notch GPUs are already at the p < 7% part of the PPWU curve and the abovementioned things serve as a reminder that we might hit the utterly absurd part of the existing PPWU curve sooner rather than later.

Be that as it may, you're making a valid point. Perhaps it would be best to let the system degrade up to a point where stating something like "your points are a reasonable estimate of the scientific value of your contribution to FAH" is widely accepted as grounds for involuntary committal.

IMHO, we aren't that far off from a situation where someone could fold 14 P9201 WUs in a month for about 124k Points Per Month (nonstop 24/7, about 30min TPF), while someone with a 36s TPF could pause after 2 WUs, collect his 125k PPM reward for two straight hours of folding and claim to have contributed slightly more than the slowpoke at the end of the month. Which is obviously false: 2 WUs vs 14 WUs... come on. Applying the PPWU cap would make such an approach less attractive, because on the average you'd have to return at least x/Cap WUs for every x WUs the 24/7 p=100% slowpoke is producing just to keep up, no matter how fast your system.

Alas, it could eventually force the true trailblazers to have their "friendly competition" on a practically linear basis... pretty much the way it was before QRB was introduced in the first place. Not necessarily a change for the worse, but hardly a popular change. Anyway, now that I've looked into my idea a bit more thoroughly, I just realized that this poll might as well have been worded thus: "Start a gradual retirement of the QRB system? Yes/No".

Scout's honor, I thought I had come up with something smarter than that when I started writing my original post. Apologies for the inconvenience...

Folding Forum

Yet another QRB proposal

Yet another QRB proposal

Re: Yet another QRB proposal

Re: Yet another QRB proposal

Re: Yet another QRB proposal

Re: Yet another QRB proposal

Re: Yet another QRB proposal

Re: Yet another QRB proposal

Re: Yet another QRB proposal

Re: Yet another QRB proposal

Re: Yet another QRB proposal

Re: Yet another QRB proposal