Folding Forum

(Me)
http://folding.extremeoverclocking.com/ ... =&u=176615
03.11, 9pm 137 1


(Folder who is clearly using the SMP client of F@H)
http://folding.extremeoverclocking.com/ ... =&u=304368

03.12, 3pm 1,760 1

Furthermore, this picture should make this problem much more clear.

If you do not know then I can reveal you a "secret" - the FAH points system is set to influence the FAH client usage/deployment trends. It is normal to assume that when ever is there a state of multiple choice (Do I install one SMP client vs 4 uni-proc clients?) the donor will likely get (some) influence from the awarded points. The current trend is to push for the SMP clients within the FAH DC system and that is what you are seeing. Can you or do you want to run the SMP FAH client is still up to the individual donor.

Dude, are you done trolling ???

Your absolutley right, it's so unfair that...........

I have spent a lot of time and money to build dual core systems to fold SMP, with all the nursing of beta software, and the expense of having high power systems turned on 24/7, while you can just download a regular, stable client, that always works, doesn't need any extra hardware from you, doesn't require you to let it run 24/7........ yeah it's so unfair.

Our purpose is to help with the science. If no points were awarded I would still do it.

Hardhead

Of course it's fair. The extra points are given to account for the extra work which has to be done to use the SMP client, and the fact that it is beta software and therefore in its very nature not entirely stable. The code is a lot better than it was but system crashes are still not unheard of, even on stable systems. And as wilding2004 says, you basically have to run any machines running SMP 24/7 or they won't finish in time. The additional expense of the newest hardware and the costs of running it constantly are enormous, so the points totals have been vastly expanded to thank the SMP (and indeed, PS3 and GPU) contributors for their extra outlay, without which the project would not be where it is today i.e. officially the fastest supercomputer on the planet. I'm surprised that you can even begin to think that it is unfair.

As hardhead says, the purpose is to further science. If you think that the points system is unfair/biased, then ignore it and just be pleased that you are contributing to a cause which is at the forefront of protein research worldwide

Yes, it may seem unfair at first glance. However, the FAQ on each of the client types discusses the points rewarded in regards to scientific production, which is the most important consideration.

However, Pande Group also takes many criteria in to consideration when setting the points benchmark. For instance, the SMP client is much more demanding than the CPU client. The SMP requires 2 or more processing cores, and you have to fold almost 24/7 to make the very short deadlines. The SMP client also takes 4 times the memory, more memory bandwidth, larger uploads and downloads, doesn't back off the CPU usage as well, and is also a BETA client.

And if you had read the forum a bit more, you'd see all the posts about the SMP client crashing occasionally and see the complaints about the points lost because of one type of bug or another. On the other hand, the CPU client does not crash and is very stable, set it and forget it, and collect the points. SMP clients have to be watched constantly. Pande Group rewards the people who run the SMP client, and who donate those extra computing resources with more points. I can't say for sure, but the SMP benchmark might also include a few extra points to compensate for an occasional crash and loss of points.

Also, when you do the math, the CPU client and SMP client have almost identical benchmark levels on a Points Per Day, per GHz, per proccessor core basis, with consideration for chip architectures.

7im wrote:Yes, it may seem unfair at first glance. However, the FAQ on each of the client types discusses the points rewarded in regards to scientific production, which is the most important consideration.

However, Pande Group also takes many criteria in to consideration when setting the points benchmark. For instance, the SMP client is much more demanding than the CPU client. The SMP requires 2 or more processing cores, and you have to fold almost 24/7 to make the very short deadlines. The SMP client also takes 4 times the memory, more memory bandwidth, larger uploads and downloads, doesn't back off the CPU usage as well, and is also a BETA client.

So what, the very fact that the high PPD aren't available for the other clients shows that the people with the SMP clients have an unfair advantage. They will NOT allow you to run the SMP client on a single processor, single core system and that isn't fair if you've got a fast single processor system.

And if you had read the forum a bit more, you'd see all the posts about the SMP client crashing occasionally and see the complaints about the points lost because of one type of bug or another. On the other hand, the CPU client does not crash and is very stable, set it and forget it, and collect the points. SMP clients have to be watched constantly. Pande Group rewards the people who run the SMP client, and who donate those extra computing resources with more points. I can't say for sure, but the SMP benchmark might also include a few extra points to compensate for an occasional crash and loss of points.

So what? The FAH6 is a beta client, I'm running it and I'm still not getting the PPD I would be getting with the SMP client.

Also, when you do the math, the CPU client and SMP client have almost identical benchmark levels on a Points Per Day, per GHz, per proccessor core basis, with consideration for chip architectures.

Are you kidding me? When you do the math, the gap between the CPU client and the SMP client in PPD output is enormous. Did you even bother looking at the picture I posted?

If you had a fast single core system, in this situation, no matter what you did, you would always get fewer points than if you had a system with two much slower processors. I believe that if you ran the regular client on a system with a P4 EE 3.73ghz or AMD64 FX 58 overclocked to 3ghz, and the SMP client on a hypothetical Dual Socket system running two 1.5GHZ P4 processors, the PPD on the SMP client would be much higher that of the regular client on either the P4 EE 3.73ghz or overclocked AMD64 FX58 system. (yes I realize that there is no such thing as a dual socket P4 board as intel made dual processors with the P4 not possible)

The discrepancy of 400+PPD is something that can't be ignored. I truely believe that the PPD between the SMP client and the regular client is something to be seriously considered.

The arguement that people who run the SMP client have to keep their machines on 24/7 is a moot point because if they were to release WU that had the same stringent deadlines on the regular client then I wouldn't be complaining. If Stanford were to simply disable the requirement that you have to have at least two processing cores in order to run the SMP client, I wouldn't be making much of a stink. Sure I'd like them to offer the WU with stricter deadlines with the regular client but since you can't run the SMP client on a single core processor, that is what I'm bringing up.


This should clarify your calculation of PPD:
Remember, the points per day are PER CORE. I know, this is easily messed up but you have to understand that it's PER CORE, not per project. You get 1700points for running the project, if you complete the project in one day on ONE processor, it'd be 1700PPD, PER CORE. If you run the project on two processors and complete it in one day, it'd be 850PPD, PER CORE. If you run the project on 4 processors (which I believe isn't recommended anyways) and complete it in one day, it'd be 425PPD, PER CORE. You take the score of the project (how much the project is worth) and you divide it by how many processors are working on it, that gives you the PPD.

imzjustplayin wrote:.If you run the project on 4 processors (which I believe isn't recommended anyways) and complete it in one day, it'd be 425PPD, PER CORE.

The SMP folding is designed for four cores... and yes, if it completed in one day, it would be 425 PPD per core.

imzjustplayin wrote:...

So what, the very fact that the high PPD aren't available for the other clients shows that the people with the SMP clients have an unfair advantage. They will NOT allow you to run the SMP client on a single processor, single core system and that isn't fair if you've got a fast single processor system.

There is one fatal flaw to your statement. Single core processors are older technology, and they will NEVER score as well as the new multi-cored systems, REGARDLESS of which client you run. Fairness has nothing to do with the way technology ages and quickly becomes obsolete. Enabling the SMP client to run on older single core processors helps neither you nor the project. Single core systems do not meet the short deadlines, and delays the scientific results.

imzjustplayin wrote:...So what? The FAH6 is a beta client, I'm running it and I'm still not getting the PPD I would be getting with the SMP client.

Here is that fatal flaw again. How can you expect aging hardware to compete with the latest greatest hardware? And again, the CPU Beta client is much more stable than the SMP beta client.

imzjustplayin wrote:...Are you kidding me? When you do the math, the gap between the CPU client and the SMP client in PPD output is enormous. Did you even bother looking at the picture I posted?

No, not kidding. And no, i didn't have to look. I know. EDIT: Okay, that's sounds a little too conceited. I went back and looked, just in case. Ah, about as I expected. The p2148 in the pictured example is a standard work unit, NOT a bonus work unit. It is NOT an apples to apples comparison, as all SMP work units ARE bonus work units. See further explanation below...

imzjustplayin wrote:...If you had a fast single core system...

Same flawed argument.

imzjustplayin wrote:...The discrepancy of 400+PPD is something that can't be ignored. I truely believe that the PPD between the SMP client and the regular client is something to be seriously considered.

Pande Group already has seriously considered this, and will continue to do so. The SMP client is very important to the project, as several of Vijay's post on the Project News page discuss. You should read them.

imzjustplayin wrote:...You should clarify your calculation of PPD...

Okay.
SMP Benchmark is 1760 PPD, run on 2 Core 2 Duos running at 2.33 GHz. See SMP FAQ
CPU Benchmark is 110 PPD, run on a P4 2.8 GHz. See Main FAQ

Looks like a huge difference, huh? Not really, let's dig deeper.

First, all SMP work units are BigWUs. The CPU client also does a lot of BigWU type projects, and gets bonus (double) points for doing BigWUs. So this is very realistic adjustment, which puts us to these numbers...

SMP Benchmark is 1760 PPD, run on 2 Core 2 Duos running at 2.33 GHz.
CPU Benchmark is 220 PPD, run on a P4 2.8 GHz.

Next, let's adjust Per core... (SMP/4)

SMP Benchmark is 440 PPD/core
CPU Benchmark is 220 PPD/core

Hey, that's not so bad, only double. We could stop at this point, and very few would complain about getting double the points from a beta client using 4 times the computer resources, with tight deadlines, that crashes occasionally.

I could continue digressing in to PPD/GHz and all that, but let us consider a quicker answer. Most work units are Gromacs based work units, which use SSE optimizations to increase speed. And the SSE engine in the Core 2 Duo processor (in the SMP Benchmark computer) is more than twice as fast as in the P4 processor (in the CPU Benchmark computer). So if you cut the PPD of the SMP client in half (as if run on the P4 system) then the scores come out to be.... ? EXACTLY EQUAL!

OMG, that can't be right, can it? Yes, it is.

Life is UNFAIR!!! We are born to die.

imzjustplayin wrote: So what, the very fact that the high PPD aren't available for the other clients shows that the people with the SMP clients have an unfair advantage. They will NOT allow you to run the SMP client on a single processor, single core system and that isn't fair if you've got a fast single processor system.

7im wrote: There is one fatal flaw to your statement. Single core processors are older technology, and they will NEVER score as well as the new multi-cored systems, REGARDLESS of which client you run. Fairness has nothing to do with the way technology ages and quickly becomes obsolete. Enabling the SMP client to run on older single core processors helps neither you nor the project. Single core systems do not meet the short deadlines, and delays the scientific results.

Single core systems CAN score as well or even better than multi processor systems, however the reason they don't is due to them not being ALLOWED to. If they were allowed to run the same SMP client then there would be some that could score better than some multi processor systems. As I pointed out with my hypothetical scenario with two P4 1.5 and one 3.73EE or FX58 processor, the single processor system would easily beat the dual processor system.

imzjustplayin wrote:...So what? The FAH6 is a beta client, I'm running it and I'm still not getting the PPD I would be getting with the SMP client.

7im wrote: Here is that fatal flaw again. How can you expect aging hardware to compete with the latest greatest hardware? And again, the CPU Beta client is much more stable than the SMP beta client.

I believe you have severely underestimated single processor products. Yes I know the core2duo is great and all, curing your grandma's cancer and giving timmy a free college education but to say that an FX58 or P4 3.73EE isn't up to snuff to complete these same projects is incorrect to say the least.

imzjustplayin wrote:...If you had a fast single core system...

7im wrote: Same flawed argument.

No, your logic is what is flawed.

imzjustplayin wrote:...You should clarify your calculation of PPD...

7im wrote: Okay.
SMP Benchmark is 1780 PPD, run on 2 Core 2 Duos running at 2.33 GHz. See SMP FAQ
CPU Benchmark is 110 PPD, run on a P4 2.8 GHz. See Main FAQ

Looks like a huge difference, huh? Not really, let's dig deeper.

First off that is wrong, the SMP benchmark is NOT 1780PPD. One of the SMP projects IS 1760 points. If you run the SMP project on two processors and it finishes in one day, that comes out to 880PPD. The PPD are calculated PER CORE, if you're not going to read what I have to say and the evidence I have provided, then need not reply.

7im wrote: First, all SMP work units are BigWUs. The CPU client also does a lot of BigWU type projects, and gets bonus (double) points for doing BigWUs. So this is very realistic adjustment, which puts us to these numbers...

SMP Benchmark is 1780 PPD, run on 2 Core 2 Duos running at 2.33 GHz.
CPU Benchmark is 220 PPD, run on a P4 2.8 GHz.

Next, let's adjust Per core... (SMP/4)

SMP Benchmark is 445 PPD/core
CPU Benchmark is 220 PPD/core

You're obviously miscalculating once again, twice even. First you say that it's 1780PPD, it's not, it's a 1780 point project and a (at best) 220 point project respectively. Assuming you can in fact calculate both projects in one day, the PPD for the SMP project (on two cores) would be 880PPD and the PPD on the NormalBonus project would be 220PPD because the "Normalbonus" project can only be run on one processor. Run two instances of the "normalbonus" project on two processors instead of one, complete both in same day and you get 440PPD.

7im wrote:
Hey, that's not so bad, only double. We could stop at this point, and very few would complain about getting double the points from a beta client using 4 times the computer resources, with tight deadlines, that crashes occasionally.

I could continue digressing in to PPD/GHz and all that, but let us consider a quicker answer. Most work units are Gromacs based work units, which use SSE optimizations to increase speed. And the SSE engine in the Core 2 Duo processor (in the SMP Benchmark computer) is more than twice as fast as in the P4 processor (in the CPU Benchmark computer). So if you cut the PPD of the SMP client in half (as if run on the P4 system) then the scores come out to be.... ? EQUAL!

OMG, that can't be right, can it? Yes, it is.

And that's where your argument completely collapses. They're not equal, you're getting at the very minimum, twice the PPD of the SMP client. I for one have never gotten more than 137PPD on the regular client even with "bigWU" enabled and advmethods enabled/disabled.

Okay, so I gave the SMP client benchmark 20 too many points at 1780 instead of 1760 PPD. That brings the two scores even closer together. I'll edit the post for accuracy.

I'll also list references to the benchmark numbers in the FAQs so there is no doubt that you are mistaken.

FAQ: CPU benchmark info
And if you read a few more lines down, it discusses bonuses.

SMP FAQ: SMP benchmark info

Again, your fast single core still can't compete with one core of a Core 2 Duo running at even half the speed on normal work units. The SSE engine in your chip just isn't up to the task, it's aging hardware, 2 generations of chips previous to the current set. It will be 3 gens older when the Penryns ship in bulk later this year. Obsolecsence is a fact of life whether you consider it unfair or not, get used to it.

You should also go check out fahinfo.org, and do some apples to apples comparisons of PPD/GHz on that site.

imzjustplayin wrote:First off that is wrong, the SMP benchmark is NOT 1780PPD. One of the SMP projects IS 1760 points. If you run the SMP project on two processors and it finishes in one day, that comes out to 880PPD. The PPD are calculated PER CORE, if you're not going to read what I have to say and the evidence I have provided, then need not reply.

Actually 7im is correct.

All SMP projects are benchmarked such that they will achieve 1760PPD on the SMP benchmark machine.

All regular projects are benchmarked such that they will achieve 110PPD on the regular benchmark machine.

Saying that PPD is calculated per core is total nonsense. PPD is just how many points per day you get for any given project.

PPD is only given per core on fahinfo.org to simplify the data analysis.

imzjustplayin wrote:my hypothetical scenario with two P4 1.5 and one 3.73EE or FX58 processor, the single processor system would easily beat the dual processor system.

That may be true, but neither one is capabile of doing the work that is being assigned to the SMP client.

I have a dual Xeon @3.0 which is roughly double the speed of those examples. It's old technology, but it does run the SMP client. (It's really a dual P4, but that's another story.)

It runs project 3065 in 1h7m per frame or 4.65 days total. The preferred deadline is 2.2 days and the final deadline is 4.4 days so I'm going to have to move this WU to one of my Core2 systems so it can finish on time. To finish this one on a single CPU (if the WU were somehow configured to run in single-cpu mode) would require that CPU to run at 12.7 GHz. You can't buy one of them.

If you think that the fact that Intel/AMD/etc. are unable to make single CPUs that are fast enough to do the science that is being distributed to the SMP client is unfair, take it up with Intel/AMD/etc.