PG announcement about Big Advanced Projects

[codysluder]

So reduce the deadlines by all means, that is what has been decided - however PG need to be cautious as to how far they can go.

As I said above, a change from detected cores>= 8 to cores>=16 seems to be implying twice as much work for the same deadlines or half of the deadline for the same amount of work. The biggest flaw in the whole discussion is that cores>=N is a poor way to measure performance but as has been said above, it's the only game in town.

Using the "twice the cores" and the "twice the work" concepts makes some sense but it's off by a small amount. If 8 threads completes 100 frames by the current deadline, will 16 threads be able to complete 200 of the same frames within double the deadlne? (I expect that the frames will actually be twice as long, not twice as many of them.) The only question is whether 8 cores scales to 16 cores. I think that it almost does, but not quite.

Can somebody with a 16-thread 2P or 4P machine disable half their machine and tell us how the TPF changes?

[Horvat]

...
Some of us will be planing upgrades to existing equipment in order to make the new deadlines comfortably. Xeon and opteron CPU's are expensive and whilst we want to upgrade cheaply we do not want to spend money on faster CPU's only to find that it still not fast enough. e.g 16 fast xeon cores or 24 slower ones

Good answer. Until there is an official statement on that, start your research based on an expectation of halved deadline lengths as before. Use your own team's benchmark numbers to find system configs that fit the bill. Then when there is an official release, you can adjust up or down slightly to make it work. I don't know when PG will be able to comment on new deadlines, but I expect it will be sooner than later.

Halving the current deadlines is going to exclude an awful lot of hardware that fits the bill.

Here is some basic and quick math

Current deadlines are 4 days, 5 days and 5.6 days for 6900/1, 6903 and 6904 respectively.
Halving them gives 2 days, 2.5 days and 2.8 days respectively.
TPF need to be lower than 28 minutes, 36 minutes and 40m15s respectively for QRB to apply

Now my fast 24 thread machine has the following TPF

13m50s for 6900/6901 Fine no problems
28m29s for 6903 Cutting it close
41m20s for 6904 Misses the deadline

Now this machine has 2 x5670 xeon processors running @2.93Ghz + turbo under linux for max performance. This is not exactly what anyone can possibly call slow. To exclude such a machine from 6904 does not bode well for many other bigadv folders running dual socket machines, as there are more than a few (including me) who have slower hex cores in their machines (e.g L5640). It would also exclude almost all 16 thread machines from 6903/4.

So reduce the deadlines by all means, that is what has been decided - however PG need to be cautious as to how far they can go.

I agree. I was just looking at this myself. My dual X5670 system with Linux has a TPF of 46 min. on a 6904 and my dual X5675 system with Linux has a TPF of 32 min. on a 6903. So by your calculations none of my systems would meet that time requirement.

Pande Group. Please take this into consideration when setting deadlines. The changes are hard enough on us, the donors, but don't change them to the point where even dual socket systems like mine are even excluded. That would be very detrimental to your program.

[Grandpa_01]

Don't you guys think you may be over thinking this a little. PG has said raising the minimum to 16 core / threads my guess is they will set the preferred deadline to whatever a 1.6Ghz or 1.8Ghz 16 core machine will do a WU in plus Upload / download time Plus x Factor for any potential problems. Just a guess here but I would say that the Gulftowns will be able to complete them on time 2600k, 2500k and AMD X6 unlikely. Stanford is not going to cut off their nose to spite there face there is still X amount of work that needs to be done. VJ said and I quote I tried to stress the speed needed (16 physical cores), but as before, the key determining issue will be making the deadlines (especially since some people often spoof the number of cores the client detects anyway). I am pretty sure most of you know a OCed Gulftown will beat a 16 core rig at 1.6Ghz to 1.8Ghz. Maybe we should just wait and see what happens.

[7im]

Maybe it is underthinking it, and over reacting? As in the p6904 = BA12, not BA8, correct? So you would not halve (50%) the DL on that one. Try 75%. Same for P6903, IIRC.

[Nathan_P]

Maybe it is underthinking it, and over reacting? As in the p6904 = BA12, not BA8, correct? So you would not halve (50%) the DL on that one. Try 75%. Same for P6903, IIRC.

You said cut the deadlines in half and go from there so i did

[Horvat]

Exactly 7im. Your said cut in half. Now %75, that's more realistic.

[7im]

Maybe it is underthinking it, and over reacting? As in the p6904 = BA12, not BA8, correct? So you would not halve (50%) the DL on that one. Try 75%. Same for P6903, IIRC.

You said cut the deadlines in half and go from there so i did

Yep. If your new BA system can finish BA12 WUs in half the of the total deadline, then your new BA system can easily handle BA16 work units.

As I said, until there is official info, use the half DL as a general guideline, rough estimate, ballpark figure. You and Horvat took 2 minutes, and my 2 cents, and already tried to figure exact numbers.

If you want to do exact numbers, then you must take in to consideration that P6903s and P6904s are considered BA12s. If you want to stay with rough estimates, that "half" still works, because there are BA8 projects (6900) as well (as you said, 13m50s for 6900/6901 Fine no problems). Please don't try to mix rough estimates with exact numbers. You'd have to use exact numbers for everything, as in the BA12s and 75%.

Or as I also said, wait for the official numbers to come, "soon." Please do not assume anything works, or doesn't work, or is detrimental to the program based on a rough estimate. I'm sorry if that "half" rough estimate was misconstrued.

[thebluebumblebee]

Yet F@H keeps growing 6 Petaflops obvously they are doing something right. viewtopic.php?f=16&t=20011

They just might have a surprise on their hands over the next 3 months. My team captain just announced that he will quit in December because of this change. 136,000 Wu's and >105,000,000 points. I would not be surprised to see our biggest contributor quit as well as I don't see him running his SR-2 for SMP. This just seems to be an insane jump in requirements. I could see a 12 thread requirement and shorter completion times, but 16 threads???? I was trying to build a BA system, but I won't (can't) now.

[derrickmcc]

Maybe it is underthinking it, and over reacting? As in the p6904 = BA12, not BA8, correct? So you would not halve (50%) the DL on that one. Try 75%. Same for P6903, IIRC.

You said cut the deadlines in half and go from there so i did

Exactly 7im. Your said cut in half. Now %75, that's more realistic.

And in his previous post 7im said:

Sorry, but the new WUs won't be out until January (see announcement), so no deadlines on the Psummary until then. However, it's a safe bet that if the BA8 deadlines were 4 days, the BA16 deadlines will be about half that, at 2 days.

[Grandpa_01]

Yet F@H keeps growing 6 Petaflops obvously they are doing something right. viewtopic.php?f=16&t=20011

They just might have a surprise on their hands over the next 3 months. My team captain just announced that he will quit in December because of this change. 136,000 Wu's and >105,000,000 points. I would not be surprised to see our biggest contributor quit as well as I don't see him running his SR-2 for SMP. This just seems to be an insane jump in requirements. I could see a 12 thread requirement and shorter completion times, but 16 threads???? I was trying to build a BA system, but I won't (can't) now.

A SR2 reports as a 16 or 24 core / thread machine. And sorry to here your captain is quiting. I will keep checking the #'s so I do not get suprised and you are right I would be very suprised if they went down.

[road-runner]

Well that $1000 12 core 980x was short lived... I almost quit when they changed the points but didnt, I hung in there but dont think I can hang anymore.. This is getting as expensive as trying to bench hwbot, someone up there must work for Intel also...

[derrickmcc]

Just a few points:
Pande Group is an academic organisation with limited resources.
There has been a long running thread about the scarcity of BA WUs.
A limiting factor is server capacity and the time taken to upload/download WUs.
Given that the upload/download time remains the same for a given BA WU, then a server will be more efficiently used if the time between the download and upload is shortened.
Quicker turnaround of WUs allows the complete trajectory to be calculated in a shorter time. A single BA WU is useless by itself.

I believe that the above points are significant in Pande Group's decision to shorten the BA deadlines and increase the minimum core(thread) requirements.

I guess someone could volunteer to fund extra servers, and extra staff to run them and more post grads to analyse the results?

[7im]

Well that $1000 12 core 980x was short lived... I almost quit when they changed the points but didnt, I hung in there but dont think I can hang anymore.. This is getting as expensive as trying to bench hwbot, someone up there must work for Intel also...

Eh, still works well for SMP, so not a total loss.

Intel favoritism? Really? 2P Xeons and boards are expensive. Mangy Cores, or Interlagos seems like the best new BA solution at the moment.

[road-runner]

Well that $1000 12 core 980x was short lived... I almost quit when they changed the points but didnt, I hung in there but dont think I can hang anymore.. This is getting as expensive as trying to bench hwbot, someone up there must work for Intel also...

Eh, still works well for SMP, so not a total loss.

Intel favoritism? Really? 2P Xeons and boards are expensive. Mangy Cores, or Interlagos seems like the best new BA solution at the moment.

True but I could do SMP with a dual core, didnt buy it for that... I was looking at those a bit ago on Newegg, looks like a 4 processor board and 4-Opteron 6212 Interlagos might last till next year?

[Jesse_V]

I thought F@h was all about science. Like doing disease-relevant molecular simulations at the fastest speed possible. With all the different sizes of WUs, and all the different hardware to tailor to, I can see how changes would be difficult to make. Clearly hardware changes, so they have to keep readjusting to keep on target. I really understand that points drive the project, but I don't understand why enthusiastic donors would quit just because their hardware was no longer top-notch. They can still make extremely impressive contributions to F@h, and their powerful machines can really power the science; SMP-8 or SMP-12 remain perfectly fine options. I thought donors were encouraged to donate as much as they could. That's what we can still do. If I actually had the money, I would purchase powerful dedicated folding machines, and give as much to the project as I could. I'm not sure I would refrain from this just because I couldn't qualify for the best-of-the-best WUs. And I would just enjoy the fact that F@h pioneered multi-core processing for distributed computing, and that I'm curing diseases at the same time. Bottom line, I'll do what I can to cure Alzheimer's/cancer/Parkinson's/HIV/etc as fast as I can, and that's really what drives it for me. I'm pretty sure that's what it's all about.