Folding Forum

7im wrote:I do want to change the forumula, even if just slightly. The curve was only form fitted to match scientific production from 1, 2, and 4 core systems. The FAQ does not say that 8 core or higher systems were even considered. That is an oversight at the least the 10x cap probably covered. And a mistake not to come back and address it when larger systems came online.

I am yet to be convinced the exponential curve isn't a match with science, even if the faq does not mention it the formula is not based on the faq but the other way around at least afaik?

Which is why I think it's premature to argue the 10x cap which was already artificial wasn't the first mistake, and removing it the right thing to do?

Seems we can't get an answer to this question other then the actions taken so far: a formula is given which is accompanied with the message 'this is a close to scientific value as we can get it'. That action alone for answers my question if the formula should be capped or not unless that information comes from the people who came up with the formula. As soon as that is done, that will convince me.

No one else then the people from PG and those close, which include you so I'm abit surprised with the chain of events till now, can value time when it comes to returning a completed work unit or am I wrong?

That is the only question I have, and the only reason I have yet to be convinced we need to change it.

I don't want to create artificial leagues either, but that's because I personally do not mind being someone who can play in the lower end of the ppd rankings. There are people with far more resources who I'm happy to see contribute as much as they do, I'm perfectly fine with someone buying 10 bigadv systems and pass my total point count in a flash.. it does not degrade my previous efforts, it shows the progress we have made since then!

I don't want to create artificial leagues either,

This is an interesting point. The point system for Big ADV. was not advocated by donors. It was not something that people lobbied, demanding to receive extra huge bonuses in some kind of a precondition to participation.
It was instituted originally to keep track of institutional donations. Panda Labs, University, Government and corporate donors. It was only after the fact that it was opened to individual donors that wanted to run 2p and 4p systems.

My point is that this History goes against 7im's contention that large numbers of CPU were not considered in the origination of the formula; That the formula is fundamental flawed due to a lack of foresight.
If the point system was designed and instituted to accommodate 1000 core systems, then it should certainly be able to handle 48 core systems.

7im wrote:Just for reference, there are people out there folding with just 11 -bigadv systems making more than 1.2 million points per day. Why is that 1,200,000 point mark interesting? That's because the 1100+ users folding under the default PS3 user account (the current #2 ranked user in the project) only makes 1.1 million points per day. 11 systems now equals 1100. Are the 11 SMP work units turned in daily by the first account worth more than the 1100 work units turned in daily by all those PS3 users?

The PS3 is one of the more limited systems in terms of the depth of the science it can do. It's where I started. It was an early backbone of FAH. It's good portion of why I purchased a second PS3 and ran both 24x7. Technology has marched on and now both my PS3 systems are not folding, that 400 Watts of power and waste heat is better serves science by even an SMP 4 folder, a high end GPU + computer is another factor 10 + better for the same 400 watts, and I have 1 bigadv running. Sure Grandpa is kicking my butt with his multiple bigadv systems. More power to Grandpa!!

The PS3, in it's day could do very, very simplified science really, really, really fast. One could even argue the PS3 was effectively the first Quick Return Bonus (QRB) folder, but there was no formula since the WUs were short (lessening the factor of not running 24x7 on a per WU basis) and all PS3s run the same speed. As I was running my PS3s, their 1,000 PPD totally swamped the measly 40 PPD, I got from my old Celeron, that I still have but it also sits powered off. Should we really change the points system and cap the PS3s at 400 PPD so there points will no longer swamp the points from the old Celeron? We'll limit the PS3 to only ten times the old Celeron. I'll plug my old Celeron back in and unplug my GPU folders so I don't trip the single breaker I have for all the folding computers. If you want different point categories for "Home" and "Folding Farms", go for it! If you want to make old Celerons feel valuable by capping the point system to not reflect the scientific value, you do everyone, including myself, a great social disservice. I may need the results of this research in the future if I get one of the cancers or Alzheimer's that this research relates to. If the formulas do not reflect the scientific value, please adjust them so that they do. Don't adjust them for anything other than scientific value, or you do us all, and society, huge disservice. If you need to encourage beginning folders on low end hardware, folding on computers not running 24x7, have the formulas STILL follow the science, but create the two categories, "Home" and "Folding Farms".

Given my current options, the PS3 is not science efficient or economically efficient.

For those that only have a PS3, it's a wonderful folding device, and more power to them!

But let's not artificially constrain the entirety of FAH science by comparing it to a gaming console from 5 years ago, even though that console is impressive and at the time of it's release, people were creating supercomputers (very loosely speaking) by clustering 8 of them. Technology has dramatically improved. Instead of clustering PS3's people are now using a single gpu, or several, or dozens (Atlas Folder), or thousands (recent asian supercomputer). I haven't heard anything lately about clustering PS3's. The PS3 heat and power costs are prohibitive compared to alternatives available today, that even cost less for the initial purchase if you have an existing computer.

If you really want to get upset about obsolescence, try farming. My limited understanding is that you have to buy the latest harvesting equipment to remain competitive
. but it's obsolete and uncompetitive before it's paid off, making farming economically very difficult. I heard that in the past, so maybe it's not as true today. More to this discussion, would you buy a 5 year old laptop today? Or more dramatically, a 10 year old laptop? It's not worth the money unless you are artificially constrained to no other option (and you don't pay for or realize how much you spend on electricity.)

Maybe we should change the thread title:

"point system is getting ridiculous..."

"Recent technology advances that produce more folding scientific value are getting ridiculous..."

And that's a REALLY GOOD THING!

If I could give you a hug I would, finally someone who feels the same way

Seeing it took so long for a post like yours, I almost doubted my own sanity

Edit:

More importantly you made me reconsider my stance on the diversification of contribution.

No need for a hug dude, just

FOLD ON!

An SR-2/x56xx class machine is now awarded about 425,000 points for a single p6703. That leads me to believe that project should focus entirely on fast multi processor machines and forget about the rest. The value of the quick return is so great, that perhaps these WUs should be run solely on supercompters. A single p6903 run in 100 minutes (merely 27x faster than the aforementioned SR-2) yields 33,000,000 ppd, more than my entire teams production in a week. One could conclude that DC isn't the way to go for FAH anymore. Running the uniprocessor, PS2, and regular SMP is certainly a comparative waste of electricity.

As I recall, one of the principal reasons the formula is exponential is to discourage donors from running multiple instances, in search of slightly higher ppd, and not to align the points to the scientific value of the WUs. The huge bonuses are the result.

mdk777 wrote:

I don't want to create artificial leagues either,

This is an interesting point. The point system for Big ADV. was not advocated by donors. It was not something that people lobbied, demanding to receive extra huge bonuses in some kind of a precondition to participation.
It was instituted originally to keep track of institutional donations. Panda Labs, University, Government and corporate donors. It was only after the fact that it was opened to individual donors that wanted to run 2p and 4p systems.

My point is that this History goes against 7im's contention that large numbers of CPU were not considered in the origination of the formula; That the formula is fundamental flawed due to a lack of foresight.
If the point system was designed and instituted to accommodate 1000 core systems, then it should certainly be able to handle 48 core systems.

Bigadv was in response to donors asking for a points system that better matched the points to the quick returns that Vijay was asking for. He even said so a few posts back. Institutional folders had absolutely nothing to do with this.

READ THE FAQ I linked to earlier. NO mention of institutional folders!!! No mention curve matching to large CPU count systems.

Edit: quote added to reflect who I was responding to:

ChasR wrote:An SR-2/x56xx class machine is now awarded about 425,000 points for a single p6703. That leads me to believe that project should focus entirely on fast multi processor machines and forget about the rest. The value of the quick return is so great, that perhaps these WUs should be run solely on supercompters. A single p6903 run in 1 minute (merely 27x faster than the aforementioned SR-2) yields 33,000,000 ppd, more than my entire teams production in a week. One could conclude that DC isn't the way to go for FAH anymore. Running the uniprocessor, PS2, and regular SMP is certainly a comparative waste of electricity.

As I recall, one of the principal reasons the formula is exponential is to discourage donors from running multiple instances, in search of slightly higher ppd, and not to align the points to the scientific value of the WUs. The huge bonuses are the result.

Why should I not run my regular smp? There are scientists enough who want to simulate protein folding on my cpu, trust me! Why approach this only from the argument of 'what would give the most points'. ->Those huge points for that supercomputer are nothing if not followed up with more people folding the next generations, that is still the reason why F@H is perfectly suited for distributed computing in my opinion.

The principal manner in which donors sabotaged the scientific progress of this project was by running dual smp's on their systems, and the QRB was put into place to finally put a value on quick returns not related to just the dual smp's saboteurs but solely to the match the coherency of time and science as close as possible.

Edit: to clarify, the last paragraph doesn't per se represent my personal conviction but just an alternate or opposite viewpoint which is at least equally valid, more so because until now that is the official standpoint.

ChasR wrote:An SR-2/x56xx class machine is now awarded about 425,000 points for a single p6703. That leads me to believe that project should focus entirely on fast multi processor machines and forget about the rest. The value of the quick return is so great, that perhaps these WUs should be run solely on supercompters. A single p6903 run in 1 minute (merely 27x faster than the aforementioned SR-2) yields 33,000,000 ppd, more than my entire teams production in a week. One could conclude that DC isn't the way to go for FAH anymore. Running the uniprocessor, PS2, and regular SMP is certainly a comparative waste of electricity.

Its only a waste of electricity if you have the option of doing something different. Many people want to contribute but can't afford an SR-2 or 4P rig. Just because their points are much lower than someone with an SR-2's doesn't mean that their contribution isn't valuable, but they may have to become used to the idea that it isn't AS valuable. I think one problem is that people don't like to be told (even implicitly through points) that their contribution isn't as valuable.. but this is science, and as a scientist I can say honestly that there are definitely things you can and can't wait around for.. and when you need results to confirm or get a paper published, you're willing to do a lot more to get them (like give out lots of points).

So do I think the point system is getting ridiculous? yea, maybe a bit... but I definitely don't think we need a drastic revamping of the system. if PG tells us that a bigadv unit is worth 10x the value of a given GPU unit to them, we have to believe that, since it really might be. who knows, in a couple years they might have a different paper in the works that needs GPU WU's, and GPU points might go through the roof... then people will be clamoring for change again..

Nothing in my post was a fabrication.

A collaborator has donated a large amount of compute time to this project; those clients were initially running under username Anonymous/team 1. To give proper credit for the donation, we have changed the username to PDC, team 1. During the period of this donation, there are at any time between 100 and 400 8-core clients running under this username (800-3200 cores total).

Please read this post for the complete background:

Does this have any relation to the large-points value work units and recent high-scoring users?
Yes. The initial projects are 2681 and 2682, valued at ~25K points base. Although these point values seem high, the work units are correspondingly larger, so the base PPD (points per day) value is roughly comparable to standard SMP.

viewtopic.php?t=10697#p104982

Just because you do not agree with an idea, there is no need for an ad hominem response.

That was the public announcement. I was around for the beta test. I watched the program develop. I even helped proof read some of the Qs and As from that post before it was posted.

PDC had NOTHING to do with the development of the bonus points program. They were simply the first large scale user to run it.

The Large WU program was up and being tested long before everyone noticed the points spike from PDC, hence the rename from Anon to PDC. That one quote from that post was simply an explanation of why we were seeing a huge points spike weeks before that post took place. PDC was the reason for the new really big work units. It was NOT the reason for the new bonus program. The new bonus points program was for all of us general users, so we could add to what PDC was already doing.

So you are trying to link two unrelated statements to draw some third unrelated conclusion. It doesn't hold water.

I don't disagree with the idea, I helped shape it. I very much agree with the idea, and now I'm trying to help shape it to be even better

That was the public announcement.

Correct, my reasoning was based entirely on the publicly available information.

A. Logical conclusions, not unrelated.(first post describing the program describing the current issue under debate

Although these point values seem high....)

B. Demonstrates that the Bonus system was designed with full cognition of large Processor counts.

(800-3200 cores total).

Which was the point of my post:

The original intention was to unify very different classes of computers under one point system, not to set up different classes as some have suggested.(as a possible solution)
Perhaps that is not possible.

I was just pointing out that extremely large core counts were certainly considered when the existing system was established.
These facts are at variance with the position that some type of short sighted oversight took place that now has to be addressed.

As I see the problem, it's not so much the exponential nature of the system, it's the un-constrained growth of the curve as the TPF gets down to very low times.

One option then would be some where along the line change the formula to a multiple of the number of points for that TPF. ie If we use a cutoff of 10 minute TPF, and a 10 min TPF gives 500K points, then we say that a 5 min TPF give 4 * 500K points, a 2 min TPF gives 7 * 500K etc:

So, for example:

If TPF (in seconds) > 600 then: points = min(1,sqrt(deadline_time * k / WU_time)) * Base
If TPF (in seconds <= 600) then points = sqrt(deadltime_time * k / ( 1000 / 1440)) * Base * (2*((600/TPF)-1)+1)

Using that formula - If you do it in 10 minutes, you get X points.
If you do it in 5 minutes, 2X as fast, you get 3X points
If you do it in 2.5 minutes, 4X as fast, you get 7X points
etc.

H.

A collaborator has donated a large amount of compute time to this project; those clients were initially running under username Anonymous/team 1. To give proper credit for the donation, we have changed the username to PDC, team 1. During the period of this donation, there are at any time between 100 and 400 8-core clients running under this username (800-3200 cores total).

So that's A LOT of 8 core systems, NOT a few systems with extremely large core counts... PDC is a computer farm, not a super cluster.

Please also check the date of that post you quoted. 8 core systems were rare at the time, so it really was a big deal. Just not as big as you claim in point C. above... (800-3200 cores)

And so far, I've seen nothing to suggest that larger count cores (more than 8) were benchmarked and used to curve fit the points curve mentioned in that Points FAQ I linked earlier.