PG announcement about Big Advanced Projects

[7im]

I doubt "we all" have a $1000 to spare right now.

Again, you don't need to raise the bar. If there are more people trying to fold BA WUs than there are BA WUs, then the extra people get regular SMP WUs.

And if "we all" had followed the Best Practices FAQ from the start, when BA was 8 cores instead of using 8 threads, they probably wouldn't have need to raise the bar 2 times in the first place. That's your catch-22.

[codysluder]

I think the main issue here is a marketing one. I don't blame PG for wanting to change the hardware requirements for certain types of science.

They had a couple of basic choices;

1. make changes that will reduce the PPD available to certain people (ie. the bigadv folders), or

2. they could have increased the PPD available to other folks (ie. GPU and SMP)

I disagree. You are suggesting that they needed to change the points system. They specifically said they were not changing the points system. Any changes that they might make to the points system would have provoked many times as many complaints.

It's impossible to make any change without 20% of the people thinking it's a good idea and 20% thinking it's a bad idea. By changing the BA deadlines, they might have reduced the 20% numbers to 15%. Changing the points system would have changed the 20% numbers to 40% and the emotional level would have been 10X as strong.

[codysluder]

tighter deadlines won't do it either as reducing deadlines enough to shut out say a 3.6g Hex core Cpu will shut out lower end server class rigs like dual quadcore 2 Ghz Xeons that meet the hardware requirements

Realistically, though, shouldn't it be about performance rather than about cores? Forget about changing the hardware requirement from 8 to 16. Announce that you're going to change the deadlines and then do it. With enough warning (like maybe by mid-January) enough people will know about it. You have been adequately warned that your hardware may not meet the new deadlines. If the 3.6 Hex core cpu can meet the deadlines, let them request BA assignments. Same for the dual quad Xeons. If it fails to meet the deadlines the owner can stop requesting BA assignments.

The points incentive works both ways, and for those who shoot for maximum PPD, they'll figure it out. Why should this be the responsibility of the assignment servers?

[Dead Things]

Realistically, though, shouldn't it be about performance rather than about cores... [snip]

Overall a smart idea and in-line with existing scenarios. For example, a 2C/4T Atom is technically eligible for SMP but would never manage to meet any deadlines, hence people have learned not to fold SMP with them.

The issue though, goes back to one of the impetuses for the change. That is, the set-it-and-forget-it crowd are missing too many deadlines. Narrowing the deadlines would only exacerbate that issue, which is why there also needs to be a stricter eligibility requirement that can be handled at the AS level. Since a thread-counting mechanism already existed, it is obviously the easiest and most expedient to implement. I agree with the much-repeated sentiment in this thread, though, that some form of benchmarking should replace the thread-counting mechanism at the earliest possibility. One based on past performance of entire units would be my preference, so as to avoid the pitfalls of tpf benchmarks that fail to take into account machines that do not fold at 100% 24/7.

[Jester]

I doubt "we all" have a $1000 to spare right now.

Again, you don't need to raise the bar. If there are more people trying to fold BA WUs than there are BA WUs, then the extra people get regular SMP WUs.

And if "we all" had followed the Best Practices FAQ from the start, when BA was 8 cores instead of using 8 threads, they probably wouldn't have need to raise the bar 2 times in the first place. That's your catch-22.

True,
But those following or exceeding those guidelines by building a true 16 core rig (with 32 cores as a bonus) are the ones Pandegroup are targeting with new big Wu's and fast returns,
then having "marginal" 16 thread and/or lesser rigs "tricked up" to squeeze inside the deadlines how can they assure that the big Wu's go to those rigs that are most able instead of simply
falling back to smp while the "marginal" rigs see an equal share,
Which I feel is at least a factor in the new harware requirements and tighter deadlines, going for anything other than the 16 core minimum could well see it's days as a Bigadv rig numbered already.

[Leonardo]

the bar would need to be raised higher still to avoid the current situation of demand outstripping suppy....

If Pande Group continues to have classes of work units that are being neglected at the expense of limited-edition advanced projects, then sure, I would assume there would be future raising of the bar. This applies to all of us - every hardware upgrade we make is risky, with respect to Folding@Home. Although I haven't always placed my bets optimally, it's a risk I am willing to accept. If one can't accept the risk, settle for for less risky, older, stable project series. I'm not throwing this at you, Jester, as I too have to consider risk/reward every time I contemplate a hardware reconfiguration. Bad bets on my part: purchasing a whole platoon of video cards just as the -bigadv QRB for CPU-SMP was announced.

updated for clarification

[mike.million]

I see the comment "the new requirements are not about the cores, but the time limit", then why should we make people run a hack to fold them? Plus why only use the best machines to the fullest, stratify all levels to a two day completion time.

To start off a hard 2 days for bonus, + 18 hours for base points, for every smp and bigadv work unit. Then with enough settings we can pick the best match for our hardware. People are not going to select a setting they won't get any bonus points for! Then, if necessary, Stanford can adjust the completion times up or down to stream line the process.

What we need is a mechanism to pick the right work units for our rigs (you can have beta testers work up some suggested time frames for different hardware or just ask for input from folders currently folding, prior to going live). Lots of settings like -smp, -regsmp, -bigsmp, -supersmp, -adv, -bigadv, -superadv, some like super ones could be future use, while the current lot of smp and bigadv could fit easily into the others settings. For example 6901 and 6902 type wu's would be -adv, and that flag would get you 6901's and below only, 6903's and 6904's -bigadv. I don't fold regular -smp but seems that they are all over the place on tpf and points, stratify.

Also the points should be reset, with lowest points at current -smp levels, then increase from that base with higher points for each step up the smp ladder, and the -bigadv need to be adjusted down a bit as they are a little out of hand.

To show why stratify
...................Xenon 5650's
My................SR-2 at 4 ghz........PPD...........Days.....Quad 6174........PPD.........Days
6901..............10m 14s.............150,505......0.71......6m 19s...........310,310.....0.44
6903..............26m 51s.............190,561......1.86.....14m 18s...........490,284....0.99
6904..............36m 31s.............196,706......2.5......19m 20s...........510,618....1.34

You can see the Quad looses a lot just moving from the suggested new -bigadv to -adv, (I have folded smp on it once and seems it was around 150,000 ppd on those). Currently the SR2 level of performance is almost even, loose a few points and get them every day, rather than every 3 days.

Much better than making someone with the SR2 level of performance fold SMP because they can't make the deadline on 6904's. Which is the same thing that you are doing making a I7 980x @ 4 ghz fold the same smp's that a I7 920 not oc'ed does.

[gwildperson]

There has always been a debate about whether a donor should have thousands of specific settings to pick the best WU for his hardware or whether centralized decisions should be made to best optimize all assignments to generate the most science.

Mike.million, you come down squarely on the side of the former statement. You sound like you're technically able to optimize whatever settings are provided and you would spend the time and energy getting the best possible PPD out of your hardware. There's nothing wrong with any of that.

Not everybody is like you. Stanford has to have centralized policies that make FAH simple to run and it has to make good assignment decisions for the millions of less technical folks who will not spend the time and energy maximizing their PPD. FAH has to have a policy that works for them, too.

Blending those two approaches has never been easy, especially if FAH needs to make changes that improve the overall throughput in ways that reduce the productivity of a limited number of selected machines.

[It's not unlike trying to get the Republicans and the Democrats to agree on a fiscal policy for the USA]

[mike.million]

SMP folders are not average users, I think it would get around. If you don't pick you automatically get the smaller smp wu's. They have to pick smp in the first place. Would be better if the system reset all smp and bigadv to smp and we had to change back, but that probably won't happen.

However, after I posted the above chart I finally figured out, "I" should be the one worried about loosing a boat load of points. So for the SMP folder in relationship to points vs the high end, this is a big boost in production.

As bonus points are based on how soon you complete vs the time allowed X a bonus, so my points will be going down big, Big, BIG, time (unless some adjustments are made). Points for one day vs three days of bonus will be Humongous! That is right? Right.

Some how I just figured with all the noise being made that it was the mid level machines that were getting it, not the high end. May be better to fold 64 / 4 = for 16 smp clients, brother.

[gwildperson]

I can't make any rational statements about how much your PPD will change, but you're right about one thing. In the marketplace, what used to be a "high end" machine has changed and FAH is making a corresponding change to their policy. When machines with 6 or 8 or 12 cores were rare, they were legitimately considered high end machines and they got BA assignments. Now that "everybody" can afford that type of machine, they're now mid-range machines, not high end machines. At the same time, the supply of server-class machines has probably increased much faster than the supply of bigadv projects.

Welcome to the middle-class.

[Jester]

SMP folders are not average users, I think it would get around. If you don't pick you automatically get the smaller smp wu's. They have to pick smp in the first place. Would be better if the system reset all smp and bigadv to smp and we had to change back, but that probably won't happen.

However, after I posted the above chart I finally figured out, "I" should be the one worried about loosing a boat load of points. So for the SMP folder in relationship to points vs the high end, this is a big boost in production.

As bonus points are based on how soon you complete vs the time allowed X a bonus, so my points will be going down big, Big, BIG, time (unless some adjustments are made). Points for one day vs three days of bonus will be Humongous! That is right? Right.

Some how I just figured with all the noise being made that it was the mid level machines that were getting it, not the high end. May be better to fold 64 / 4 = for 16 smp clients, brother.

That's something that most appear have missed in this thread so far, or at least not really commented on,
When the new hardware requirements take effect and the deadlines are reduced there were no comments on ppd reductions
as it was stated at the same time that there would be no changes to the points system, but as you've pointed out, reducing the deadlines will have a big impact on ppd
for faster machines, and the faster the machine the bigger the impact.

[codysluder]

FACT: The points system is not changing.
FALSE ASSUMPTION: PPD is not changing.

The same points per WU and Kfactors will still apply to all projects. The same formula will be used to calculate points.

Which WUs will be assigned will change, and that's nothing new. In fact, it's constantly changing. The only difference here is that there has been an announcement of a future change to assignment policy. Some will benefit from that change. Some will suffer a loss. Some will see no change. Some won't care if the change affected them or not.

Those that suffer a loss will understandably complain, but VijayPande has announced the change because it benefits science (and there seems to be little doubt about that), recognizing that it needs to be made even though some people won't like it.

[jimerickson]

@codysluder +1

[MtM]

Print screen, put's a picture frame around it, hangs it on the wall... goes back to couch under the blankets ( I'm sick ). Nicely said.

[Jesse_V]

Well said indeed. I doubt even Bruce could have said it better. +1, a Like, or Thumbs Up, however you want to measure the fact that I'm impressed.