Folding Forum

Why the

That IS what he is saying

*makes mental note, keep an eye on return times for my linux vm's, they are still returning wu's within 80%+ of deadline remaining but some of the wu's I had in the single windows smp where faster.. Mhm maybe I should take the vm's offline again..

edit: to back 7im up, read the same quote I'm futily throwing into shatterdsillicon's face and which he just doesn't seem to be able to comprehend

So, in essence what you're saying is that I should throw out my 16 GPUs and 6 SMPs and replace them with the fastest available single card and the fastest quad core CPU [...]

On condition you can satisfy "2x2=4" model/equation. Keep that in mind.

One doesn't need to be eligible for MENSA to notice that nowadays there's no single chip [in the consumer market] that can match 16 GPUs.

It's not a black-vs-white thing but yeah, I'd, for instance, happily replace my
2xE5335 with a single 4GHz C2Q [please, don't give me "overclocking" as that's
not what's being discussed here].
Then again, there are (electrical) power and cost factors too and extent of their
application may vary from folder to folder.


o&o,
tear

EDIT: fixed typo

osgorth wrote:

7im wrote:When running a race, the goal is to finish first. NOT to finish just before they close the race course.
...
Folding is relay race, with each WU being returned the quickest is the most helpful to the project. Follow the recommendations or don't, but higher PPD is NOT exactly equal to helping the project more. Fold on.

So, in essence what you're saying is that I should throw out my 16 GPUs and 6 SMPs and replace them with the fastest available single card and the fastest quad core CPU, just so I can return my two WUs as quickly as possible?

Absolutely not what I'm saying. You don't seem to be aware of the Stanford recommendation to run 1 fahcore per CPU core. That's the fastest way to run the race with the hardware you have. No one said anything about changing what hardware you have.

osgorth wrote:I highly doubt you mean that, but that's how it comes across.

In light of the Standford recommendation I mentioned above, this clearly "comes across" differently than how you appear to have taken it. Again, no one said anything about running one and only one client. The recommendation is very different than that.

osgorth wrote:Obviously, the more WUs returned, the better. Given that a quad-core CPU can process 2 SMP WUs in almost the same time as 1 WU, I see absolutely no reason whatsoever why this should be frowned upon.

No, not so obvious. More WUs does not exactly equal better for the science. With the current SMP inefficiencies, 2 clients on a quad probably are MORE helpful at this point in time. Unfortunately, they then add in a GPU client or two, and slow things down even more. Does that now seem like a "whatsoever" reason for frowning? It should. Or when the SMP client efficiencies improve again, and 2 WUs are no longer almost as fast as one? People will keep running 2 clients long after the point where 2 clients are LESS helpful to the project, simply because PPD is still slightly higher. And then we're back to the 2 clients on HyperThreading CPUs argument again.

I can hear the same old same crap excuses again.... Just so long as I get an extra 10% higher PPD, I don't give a crap if the WU times are almost twice as long. As long as I make the deadlines, what's the difference?

Well, we've explained the differences quite clearly now. 2 clients almost as fast as 1 seems fine for the moment. No one frowns much on that. But the other scenarios with much slower clients, and/or loading up many multiple fahcores per CPU core and just making the deadlines is where the frowns start.

Here is a post as prime example: How to run 2 SMP clients in vm while also running 2 GPU clients to maximize points. http://foldingforum.org/viewtopic.php?f=52&t=7016

Stanford set a deadline. Does Stanford say "Don't run this software unless you fold 24/7?" Does Stanford say "If you return this work faster, I'll give you more points?" No, it says make my deadline if you want the fixed amount of points.

I fold 24/7 on my overclocked, water-cooled expensive machines and can run additional load yet still make the deadlines with vast amounts of leeway. The way I do it allows more units to be processed in faster times than all the occasional users who only run one fahcore per cpu. I can return to my single Win-SMp client that fully loads my quad yet produces 1700ppd or I can run the two Linux VMs with a Linux SMP unit in each and achieve >4000ppd. Or I could just return to Rosetta@Home and not have the all the hassle.

7im wrote:No, not so obvious. More WUs does not exactly equal better for the science. With the current SMP inefficiencies, 2 clients on a quad probably are MORE helpful at this point in time.

Ask Dan if he agrees

<self-censored>

That's not the message, the message is faster results are off more benefit then more results

You're the one here who's hitting the wrong tone I feel not 7im. Couldn't we please keep this abit civil, no accusations, no frowning on references and no messages like : get used to it, as we'll do what we want anyway.

Those are not helping, and will not cause any improvements. You say you provide PG with computing power because you believe in the project, why can you not believe that science is better served with faster results then more results?

I'm not going to post everything in my inbox atm, but I can tell you the people who are able to judge what is more valuable have done so on more then one occasion and allot of times in public. If you don't want to listen to them, fine, but don't pretend you don't know what they have said many many many times already.

MtM wrote:That's not the message, the message is faster results are off more benefit then more results

Well, in that case I'm stupid because I haven't understood that.

I didn't start the bad vibes, I just felt attacked for asking a question that I didn't understand. :/
Enough of this anyhow, you're right about that.

Although we are rather OT...

To answer the question on when WUs should be returned, we should question how the researchers compile the data and move on the next set of trajectories. Couldn't find any statement by them, but if we assume that all the trajectories for a run needs to come in before they can move on to the next one, then surely there'll always be people who finish by the deadline, or lost WUs that need to be reassigned after preferred deadline. So say for every WU the last WU comes in 1 day before the final deadline.

Since they need all the clones/gens before they can compile, it means that the last guy is the erm.... rate-determining step? So as long as people finish within a reasonable margin (say 12 hrs after preferred for SMP), no matter how they do it (whether by running multiple SMPs) shouldn't really matter.

HaloJones wrote:Stanford set a deadline. Does Stanford say "Don't run this software unless you fold 24/7?" Does Stanford say "If you return this work faster, I'll give you more points?" No, it says make my deadline if you want the fixed amount of points.

Stanford has mentioned changing the points policy to exactly as you stated, "If you return this work faster, I'll give you more points". As of now they have higher priority items that need their attention before adjusting the points system.

MtM has the correct idea (especially with the high-performance clients like SMP):

MtM wrote:That's not the message, the message is faster results are of more benefit then more results

These statements from moderators may also be helpful:

bruce wrote:Running two SMP clients is not recommended. Yes, it is possible to do, but if you have to shut down your machine, the client will probably discard at least one of the active WUs. The Pande Group recommends one FahCore per CPU-core because adding extra FahCores delays the science.

7im wrote:Unfortunately, there is science optimal, and there is points optimal. Clients develop faster than the points system around here, and the two are not always 100 percent in tune with each other. Optimally, what is best for points would also be what is best for science, but the brazillion different hardware combinations can sometimes reveal grey areas where "optimal for points" is not "optimal for science."

For instance, the project recommendation is to run one fahcore per cpu core. That means one SMP client per quad core processor. However, some people run 2 SMP clients for higher points per day while still meeting the existing deadlines. Points optimal, but not science optimal.

However, all donations are welcomed. So until the previously announced points system adjustment brings this clients more in-line, please understand there is a difference. And each one of you has to ask your conscience which "optimal" choice you will make.

WangFeiHong wrote:Although we are rather OT...

To answer the question on when WUs should be returned, we should question how the researchers compile the data and move on the next set of trajectories. Couldn't find any statement by them <snip>

My quote was from Dan Ensign

MtM wrote:

7im wrote:No, not so obvious. More WUs does not exactly equal better for the science. With the current SMP inefficiencies, 2 clients on a quad probably are MORE helpful at this point in time.

For the record, above statement was true back in core A1 times [core had scaling issues]. A2 scales *really*
well [pretty much linearly], ergo 1xSMP on a quad- should give you results on par with 2xSMP on a quad;
same deal with "-smp 8" vs. "-smp 4"x2 on dual quad... and so on, and so on...

tear

tear wrote:

MtM wrote:

7im wrote:No, not so obvious. More WUs does not exactly equal better for the science. With the current SMP inefficiencies, 2 clients on a quad probably are MORE helpful at this point in time.

For the record, above statement was true back in core A1 times [core had scaling issues]. A2 scales *really*
well [pretty much linearly], ergo 1xSMP on a quad- should give you results on par with 2xSMP on a quad;
same deal with "-smp 8" vs. "-smp 4"x2 on dual quad... and so on, and so on...

Not true. It only scales well on an otherwise idle machine. Throw some other process into the mix (e.g. the GPU client), and the SMP client (a2 core) will only run at 4x the speed of the slowest FahCore thread.So if you have a process eating 50% of one core, you'll only have the SMP client running on 50% of each core. Under this scenario, on a 4-core machine, you'll have 1.5 cores sitting idle. If you're using your folding rig as a desktop/workstation/server, the imbalance will generally get bad enough that any speed benefit of only running just one client will be wiped out pretty quickly.

Well that's a different story. Knowing other workloads one can come up with a tuning [playing with affinity and so on and so on]
that gives desired regular-stuff-vs-FAH balance.

And yes, you're right saying FAH SMP does not play well with other CPU consumers.

Regardless, A2 scales very well, won't you agree?


Cheers,
tear

HaloJones wrote:Does Stanford say "Don't run this software unless you fold 24/7?"

As a matter of fact, yes, they do, in LARGE BOLD print on the High Performance download page. Not in those exact words, but the warnings and the VERY short deadlines easily translate to say run 24/7 or you won't make the deadlines, and so you won't get any points. That's a rather clear statement, run fast or no points.

HaloJones wrote:Does Stanford say "If you return this work faster, I'll give you more points?" No, it says make my deadline if you want the fixed amount of points.

Actually, that is your interpretation of Stanford's use of a deadline, not theirs. Vijay Pande, the project lead, has already stated the recommendation to run one fahcore per CPU core to return the work units as quickly as possible, because that is the most helpful to the project. Vijay has also posted (last year) how they are working on a points system adjustment to better align the points to the speed, i.e. he all but said a new bonus would be tied to how quickly you return a work unit.

HaloJones wrote:I fold 24/7 on my overclocked, water-cooled expensive machines and can run additional load yet still make the deadlines with vast amounts of leeway. The way I do it allows more units to be processed in faster times than all the occasional users who only run one fahcore per cpu. I can return to my single Win-SMp client that fully loads my quad yet produces 1700ppd or I can run the two Linux VMs with a Linux SMP unit in each and achieve >4000ppd. Or I could just return to Rosetta@Home and not have the all the hassle.

Sorry, this isn't about you, or only your machines. It's about all of us working together. You are more than welcome to maximize points, and just make the deadlines, as that is an acceptable contribution, just like the people with slower computers. But you are NOT helping the project optimally. Using the "slow guy comparison" excuse is not a good justification. RATE of WU completion is more important than COUNT of WU completion due to the serial nature of the work units.

Again, all contributions are welcomed, even those who just make the deadlines. This isn't a debate about right or wrong, just about what the project recommends as the optimal way to help the project, not the optimal way to earn points. Unfortunately, those two things are not perfectly aligned yet, but they are working on it.