Change in BA requirements

[Joe_H]

This post has been preemptively recorded so evidence of tampering, if any, can be provided.

That is an unnecessary precaution. Ideas expressed by individuals regarding FAH are NOT censored.

...

I will post just one correction here, the actual quote is from tear, not gimpy. gimpy quoted another post and the quoting got messed up somewhere along the line. I have looked at the post a couple times and still not figured out where to patch up the tags to get it right.

[kerryd]

This is where I get fubar if I am folding only on my 3930 k with out 2 cores to run gpu's . I get same points as my 16 core 4p's running smp. But if 16 core 4p's are doing BA they get from 2 x to 2 1/2 times the ppd, <<<< sames work same points you would think.And it is same work sorry Grandpa_01 but a cpu can only do so much work at 100%
Well what I am thinking is smp needs a point boost I am not asking for them to match BA but with core 17 needing a cpu core or more for each gpu something is not right in the hen house. I really do not know how to fix it with out making ppl mad one way are other.
But i get off the topic up there kind of, the topic should be what to do with 16 core and 24 core ankle biters. There really no good for smp folding fyi 3930k and cost 4 or more times the $ to run then a 3930k with 2 gpu's .
So what it all comes down to is what are we going to do with are ankle biters , we are not going to do smp on them or at-least I will not at $60 a month each and I got low power rates.
I really do not want to stop folding but I need to know that my money is not going down the drain every 3 months.
.

[Bill1024]

Adak meet in the middle 80%+

There are quite a few helping with the Kill The SMP Backlog drive. I am pleased with the effort.
We should make it a contest between teams. Who can fold the most SMP WUs in 5 days.

ME, I have done 375 in 5 days

I am impressed, I thought I was kicking butt at around 100 done.
Thank you.

[gwildperson]

Long sermon, but I actually came up with an idea. How about demoting some of the existing BA projects to SMP on April 17th, giving them high priority on the AS and requiring 16 or more cores to fold them? They would no longer receive the BA-specific 20% bonus, but other than that, business as usual. The projects remaining in BA, and new BA projects launched after April 17th of course, would require 32 or more cores/threads and receive the BA-specific bonus.

Nobody likes pay cuts, so to speak, but I think it might be prudent to set up a poll about it...

From that quote, it appears that the original policy granted a bonus increase of 20% over SMP. Does anybody know why or when it became a 200% increase?

If it were1 still 20%, I don't think we'd be having all this controversy.

[7im]

It's 20% applied after the exponential QRB is applied, so it scales up very quickly. It is not 200%.

[Adak]

When you say 20% bonus, are you referring to the k factor being set initially to 2?

Review of BA history. Summary of Dr. Kasson's post, July 16, 2009

Why are BA wu's earning more points, and a different work unit category?

1) Simulations are larger and more bandwidth required for up/down loads.

2) Faster turn around needed. A new bonus scheme to reward faster return will be in effect.
The calculations for it are:

Base points are roughly equal to SMP units.
Total points=base pts * bonus factor.

bonus = sqrt(deadline * k/WU_time).
k started at 2, but is adjustable.

I agree with you that if the drop in points was 20% for the "ankle biters" about to be dropped from BA, then there would be no problem. Some dissatisfaction, to be sure, by the owners of the "ankle biters", but this change in BA threshold was not unexpected.

[Adak]

This is where I get fubar if I am folding only on my 3930 k with out 2 cores to run gpu's . I get same points as my 16 core 4p's running smp. But if 16 core 4p's are doing BA they get from 2 x to 2 1/2 times the ppd, <<<< sames work same points you would think.And it is same work sorry Grandpa_01 but a cpu can only do so much work at 100%
Well what I am thinking is smp needs a point boost I am not asking for them to match BA but with core 17 needing a cpu core or more for each gpu something is not right in the hen house. I really do not know how to fix it with out making ppl mad one way are other.
But i get off the topic up there kind of, the topic should be what to do with 16 core and 24 core ankle biters. There really no good for smp folding fyi 3930k and cost 4 or more times the $ to run then a 3930k with 2 gpu's .
So what it all comes down to is what are we going to do with are ankle biters , we are not going to do smp on them or at-least I will not at $60 a month each and I got low power rates.
I really do not want to stop folding but I need to know that my money is not going down the drain every 3 months.
.

Is your 3930k cpu overclocked? I'm guessing that it is, since it's an unlocked cpu. But your 16 core 4P system is running either at stock speed, or with very little overclocking. Is that correct? And one cpu is from Intel, and the others are from AMD, and they don't have any resemblance to equal performance, core to core.

These are the kinds of issues that PG has avoided, simply because there is no way to even all the work unit point, into a perfectly fair system, for everyone. The only advice that makes sense is for you to try and look ahead 3-5 years, on all of your hardware purchases. For example, if I were to purchase a system today, with only 32 cores, and barely able to make the deadlines, I would expect to be bumped out of BA folding, far sooner than three years.

PG may not know what hardware will be best for BA folding, 5 years from now, but everyone knows that the hardware will be improving substantially, and that BA folding will try very hard to follow those leading edge hardware trends.

[kerryd]

This is where I get fubar if I am folding only on my 3930 k with out 2 cores to run gpu's . I get same points as my 16 core 4p's running smp. But if 16 core 4p's are doing BA they get from 2 x to 2 1/2 times the ppd, <<<< sames work same points you would think.And it is same work sorry Grandpa_01 but a cpu can only do so much work at 100%
Well what I am thinking is smp needs a point boost I am not asking for them to match BA but with core 17 needing a cpu core or more for each gpu something is not right in the hen house. I really do not know how to fix it with out making ppl mad one way are other.
But i get off the topic up there kind of, the topic should be what to do with 16 core and 24 core ankle biters. There really no good for smp folding fyi 3930k and cost 4 or more times the $ to run then a 3930k with 2 gpu's .
So what it all comes down to is what are we going to do with are ankle biters , we are not going to do smp on them or at-least I will not at $60 a month each and I got low power rates.
I really do not want to stop folding but I need to know that my money is not going down the drain every 3 months.
.

Is your 3930k cpu overclocked? I'm guessing that it is, since it's an unlocked cpu. But your 16 core 4P system is running either at stock speed, or with very little overclocking. Is that correct? And one cpu is from Intel, and the others are from AMD, and they don't have any resemblance to equal performance, core to core.

These are the kinds of issues that PG has avoided, simply because there is no way to even all the work unit point, into a perfectly fair system, for everyone. The only advice that makes sense is for you to try and look ahead 3-5 years, on all of your hardware purchases. For example, if I were to purchase a system today, with only 32 cores, and barely able to make the deadlines, I would expect to be bumped out of BA folding, far sooner than three years.

PG may not know what hardware will be best for BA folding, 5 years from now, but everyone knows that the hardware will be improving substantially, and that BA folding will try very hard to follow those leading edge hardware trends.

I kind of like this one (The only advice that makes sense is for you to try and look ahead 3-5 years, ) and just how do you do that.
I got into computers late but at that time a 478 socket was really fast about 3 years later come along socket 775. Not all that much faster but with a few more core so was a bit faster. i ran each for 3 years then I missed a socket now on socket 1155 but then I get a chance and put together a socket lga 2011.<<< that was going to be future prof ., Now I am sure the mother board will not run the new socket 2011 when they come out.
So this is where we are at . Say a layman orders a computer at this time you most likely will get a i5 computer BUT you could get a socket 778, i3 i5 i7 or a .
My point is folding at home was for the OR A not for the best out there.
Now it seems they want the best out there but heck my gtx770's are no better then my gtx 580's if I do not change drivers 1 to fold other to game.
But what panda has forgotten is every day folders I know it all comes down to money in the end but most of the folders are running socket 778,i3,i5.Most do not have a i7 and fewer yet are running 2011 cpu's. <<<<< now that is most of there folders. Then you get BA folders I can not give a % but its got to be under 10% so you got 10% getting 90% of the points just a guess.Hay do not get that wrong I liked being in the 10%. Hay that sounds like the USA where the top 10% get richer well the bottom do what ever, Hay we find a way.

[Adak]

From Kasson's post on BA, I took it that BA was a completely different species, as far as the wu deadlines. These were wu's that normally were sent to a supercomputer cluster. Turn around time is not just important, like it is for SMP work units, it's absolutely critical. If the speed isn't there, then FAH can't do the BA projects - they'll be sent to the supercomputer cluster instead.

Intel showed off a prototype card for the PC that was basically a glorified GPU on a small card, which drew very little watts, and had stunning speed. I couldn't help thinking to myself "dang, that one card is going to make my folders obsolete, some day". That card hasn't come to the market yet, but I'm sure some day it will. That's just the kind of thing that neither PG nor us donors, will have any control over.

With FAH, you definitely have the larger ups and downs.

[kasson]

The idea of BA in the first place was indeed for these "extra-demanding" work units where 1) speed is important because the systems we simulate are larger/slower, so we need more computation to get useful results and 2) the larger nature of the simulation meant an increased use of system resources that we might not explicitly benchmark for (memory, upload/download bandwidth, etc.). The goal of having a separate bigadv flag was to let donors opt in to these more resource-intensive work units that could cause problems on slow links, etc. And the goal of the BA bonus was to have a ~20% boost to recognize the demand on resources that was not explicitly reflected in benchmarking.

The BA vs. SMP points yield comes down to a few things. 1) BA-eligible machines are usually fast-end, so they are getting higher QRB whatever they're running. 2) large, not just long, simulation systems tend to scale better (e.g. 4x the cores -> 4x the performance) than very small simulation systems. SMP work units are a mix of sizes--we have some regular SMP projects that should scale essentially the same as BA through at least 128 cores if not 256--but some won't scale quite as well. 3) because BA work units are more resource-intensive, the benchmarking has been harder: BA work units run more slowly than one would expect on our standard benchmark machines, and some of the SMP work units that don't scale as well would run more slowly than expected on good BA machines. That's why the points assignment has been less trivial than we'd like. It is possible (likely?) that this has caused a greater points discrepancy for BA vs. SMP than originally intended. There are a number of reasons (not just core count but NUMA configuration and more subtle things) why not all work units/projects will scale the same between different machines. We'd really like some more consistent way of assigning points, but we haven't come up with something that we like better, think is simple and easy to understand, and think is reasonably secure.

Hope this helps a bit with some of the technical questions.

[7im]

So even at the very low end of BA, there is always going to be at least a 20% cliff drop in points from BA down to SMP because of the way the BA bonus was added.

Seems like there needs to be a better bonus curve formula so BA points line up with SMP points. And then to make it fair for those people running other operating systems with big iron, a way to add a BA like bonus for their additional resources donated that are not easy to benchmark. BA bonus for Linux only is not exactly hardware agnostic.

[Napoleon]

Adak meet in the middle 80%+

There are quite a few helping with the Kill The SMP Backlog drive. I am pleased with the effort.
We should make it a contest between teams. Who can fold the most SMP WUs in 5 days.

Good idea, albeit not a new one... viewtopic.php?f=15&t=20852.
If some awards/rewards could be arranged to lure teams/donors into it, all the better.

[mdk777]

well, let me repeat again that smp work units are a red herring.
The real issue is that BA work units are just no longer required or have projects lined up to use the unique capability.
If you look at the total for Linux, and assume it is ALL BA:

You have 13 out of 17032 current TFLOPS.

The diligent silence, the refusal to address any substantive issue raised on this thread indicates that PG is not worried about the loss of those 13 TFLOPS.

If they(BA folders) are moved to smp, or if they ALL simply quit, the difference to PG is negligible.

This of course is very harsh news to those who invest and pay utility bills to run these WU.

edit:
humor redacted voluntarily.

[Viper97]

Oh.. I don't know mdk777 I thought your post on our forum was hilarious and oh so true.

Meanwhile if indeed this is all as they say unprocessed cow feces then I'd say PG and FAH are toast donor wise. It isn't just the BA they would lose but rather the other poor souls who begin to realize that for them this is a massively unappreciated donation on their part. We'll see how this plays out but right now... FAH was at 234K machines connected and as of today there are 226,224. When I first started to get interested in the loss of machines a month ago, the machine count was in excess of 240K... 14K is not an insignificant amount to lose.

What is interesting a few days ago it was 230K... so an additional ~4K have been lost since then.

I'm thinking if I were counting on us to help I'd be doing a far better job of PR. Just saying.

[Adak]

well, let me repeat again that smp work units are a red herring.

I would agree with you that it's not the primary reason for PG to change the BA threshold.

The real issue is that BA work units are just no longer required or have projects lined up to use the unique capability.

That is possible, but unlikely, in the long term. Having a resource that will turn out results that rival supercomputer clusters, but at a lower cost, makes the BA capability, a valuable part of FAH. Perhaps THE most valuable part of FAH.

If you look at the total for Linux, and assume it is ALL BA:
You have 13 out of 17032 current TFLOPS.

You can't measure the value of BA folders, by just TFLOPS. Their value includes a much faster turn around speed, an ability to handle much larger work units, and a tolerance for the larger up/down loads those larger work units require.

The diligent silence, the refusal to address any substantive issue raised on this thread indicates that PG is not worried about the loss of those 13 TFLOPS.

You do remember that:

1) Stanford is closed for the holidays. There will be wok done on FAH, throughout this period, but it's not a full staff, as many will have scheduled vacation time, etc., during the shut down of the University.

2) Dr. Kasson just responded, above. PG has not hired their donor relations representative yet. Why make substantive changes now, when they're still lacking that asset, and need to study the issues?

3) One of the main issues is the current point disparity between what a BA rig earns folding BA wu's, and what the same rig's ppd is when folding regular SMP work units. That requires a careful adjustment.

If they(BA folders) are moved to smp, or if they ALL simply quit, the difference to PG is negligible.

I strongly disagree, because nearly every BA folder has other non-BA folding rigs, as well. BA may make up "only" 13 TFLOPS, but the rigs BA folders run, add a lot more to that figure.

This of course is very harsh news to those who invest and pay utility bills to run these WU.

The rising electric rates are bad news, but most folders will respond by turning off the smaller rigs, and leaving the BA rigs running as long as possible. They do the most for FAH (per rig), and they do the most for our points.

FAH must run with whatever hardware the donors have, today. Tomorrow, our hardware may be made obsolete by the newest computer hardware. (Imagine a Quantum Computer), and BA by new GPU's with thousands of cores, but FAH will have to make those changes after the new hardware becomes available, not before. FAH itself, may become obsolete, if a faster and more accurate method is found to study folding proteins, that doesn't use a distributed computing project model.

There are no guarantees here. Not for FAH, and not for folders either. Both FAH and it's donors, do what we can with what we have, today. For tomorrow, we need to stay flexible, and expect change - because change will surely be there.