HyperThreading and Folding

[Gizmo]

Just came across this at Weatherman's site (www.em-dc.com). (BTW, version 3.0.6 of Electron Microscope is out).

This is a direct quote from his site:

"Folks, for many the HT (HyperThreading) issue is one that they just don't understand. While Stanford allows you to run more than one instance of the Folding@Home client on your computer, you are not really helping them if the number of clients you are running is not equal to the number of REAL CPU's you actually have. While running two instances of the F@H client will return two work units more quickly than doing one at a time, that is not what is important. Returning one work unit 70% faster and starting the next generation of tests on the work unit, is what Stanford wants. Speed in working through each protein generation is very, very important. Sometimes quality is better than quantity. Quality in this case is work being returned more quickly. While it is well known that an HT CPU can run 2 work units because the OS treats it like two CPU's, and you can gain a possible 15-30 percent increase in points for that computer, it also means that each work unit is returned more slowly. In simple terms, if the project has 300 generations needed to test a theory on a protein model, and running two instances at once delays the return of work for 1 day each time, you end up with a 300 day delay. That translates to about a 10 month delay in examining your data for the final result. The bottom line is simply this: Run 1 instance for each CPU you have. An HT CPU is not two CPU's, it is one. Let's work to advance the science and spend a little less time worrying about the number of points you get.

This is a direct quote from Dr. Vijay Pande:
1) If you care primarily about points, running 2 procs on HT is still the best bet. We are grateful for all contributions and if people choose to run 2 procs on HT, our approach is that all contributions are welcome.

2) If you care about the science foremost and are interested in our recommendations, then do not run 2 procs on HT, but please just run one process. That won't be best for points, but is best for the science.

3) If your machine cannot make the deadlines, then one should run the timeless WUs.

I hope that clears these issues up and thanks to all for their contributions.

Thanks for listening and please do the right thing.
Larry"

Just something to think about.

[mookydooky]

Now they tell us...

[mookydooky]

I have to say though, the logic just doesn't make sense. I would think volume of work done would be of primary importance. If it takes me an extra day to finish a WU, but that extra day is spent doing another WU, then I'd have two WU's instead of one. I'm trying to go somewhere with this, but lack of sleep is making my thoughts all fuzzy so someone else is going to have to figure this one out.

[Gizmo]

They are saying that turn-around time is more important.

I have to admit that I am a little confused myself. Granted, the turn-around time on one WU would be improved by running a single instance, but it seems like the turn-around time on two would be improved by running two instances. This is no different than batch processing in this regard.

[SteveI]

By the same logic... they should set up a separate platform of folding, only allowing the fastest CPUs running 24/7 to participate. This would guarantee the fastest turnover for projects earmarked as such. A P4 HT can finish 2 WUs faster than the vast majority of PCs folding just 1. Keep in mind that the min config is something like a 300mhz celeron.

Remember, us overclockers running juiced up PCs are a very small fraction of donors out there.

Vijay has always been strongly opposed to running multiple instances. Early on, HT was unstable when running multiple instances. It is still very sensitive to overclocks. I think he is gunshy about HT, and probably has a lot a bad WUs to show for it.

[keithwalton]

It intrigues me as well to see that they want speed over volume. Are they begining to run out of theories / units or somthing. As with tinkers dual processes is alot more effective at getting work done. around 40% in my observation.
Very rarely do my instances run the same type unit either, typically one gromac and one tinker picking up the slack. Why a 3GHz P4 keeps getting tinkers though and yet a non sse athlon keeps getting gromacs is abit silly.

Ditto what gizmo said, turn around on two units in parallel is shorter than in series.

I wonder also what's happend to the dgromacs units, they've appeared to have vanished for now,
1/5 of my rigs supports hyperthreading, 3/5 supports sse2, and 4/5 supports sse.

The sse2 core to come back would be nice, I guess i'm dreaming of the day when folding will support ht cpu's natively and take full advantage of the processing power avaliable

Somthing i dont quite get is the bit about the 10month delay, yes if there was only one cpu working on the entire project and it was taking an extra day per chunk then fine. But they have 170,000 + cpu's churning through data at present *looks dazed and confused* my p4 returns two units in less time than the athlon does 1 even though the athlon is on all the time and the p4 aint. By his logic the athlon is taking to long to complete work so it shouldnt bother, nor should the p3-m which slightly outfolds it.

I allways thought distributed computing was about mass parralisation, i'm a little stumped on why small scale parralisation which by there own points standard is more productive, is wrong.

The more i think about this the more confused i'm getting so i think i'll call it a night after this edit,...
As i understanded it units have deadlines so that work gets done and redistributed if not finished in time. My p4 does units with deadlines of weeks in a matter of a day or less, many days under the prefered time as well. Recently i've done a fair few 872's, its taking it 6 hours todo a units of it (with a tinker churning away in the background). the deadline is 9 days prefered is 6. much bigger units take a day or two, perhaps 4 days tops for the tinkers which have deadline of 68 days and prefered of 46 days

[GrahamGarside]

i suppos its a bit like downloading. We all know that if we download 2 files at once we are more likely to get closer to our max bandwidth and thus make optimal use of it. But for example you have a 1mbit conection and there are 2 350mb videos you want. Downloading them both may give you close to 120kbs and so it would take an hour and 40 minutes to download. Now if you download them seperately you may only get 100kbs on the download, which is faster than the 60 you got for 2 files (120/2) but not full use of the line. So you have the first file now in 1 hour and while watching that video for 45 minutes you only have 15 minutes left for the 2nd one to finish which means you have only been idle for 1 hour 15 minutes compared with 1 hour 40 minutes for getting them both at once
I'm probabally way off but I bet it's something like this

[Logan]

Could a special client be made to run for H/T cpus? How would that work, actually?

[RussianMissile1]

split the WU itself maybe?

[Gizmo]

Originally Posted by GrahamGarside

...which means you have only been idle for 1 hour 15 minutes compared with 1 hour 40 minutes for getting them both at once
I'm probabally way off but I bet it's something like this

It's something like that, I'm sure. Thing is, your analogy breaks down when you scale from the domain of individual downloads to the domain of hundreds (or even thousands) of downloads, all started at random times, which is essentially what we are dealing with here. This problem can be handled pretty effectively by good resource management, or so it would seem.

[Gizmo]

Originally Posted by RussianMissile1

split the WU itself maybe?

I doubt that is practical. If it were, they could simply send it out as two WUs.

[SteveI]

Originally Posted by gizmo

....This problem can be handled pretty effectively by good resource management, or so it would seem.

I think this is the ultimate problem. Pandegroup does not want to actively manage where the WUs are going and trying to determine expected return times. They do care about return times, since each progressive step is determined by previous results. They don't want to alienate those with slower machines and limit total involvement, but they may see HT as one way to improve response time. However, given their arguement... I think if everyone complied, on a global level, their response time would improve by less than 0.1%.

Most PCs do NOT have HT. I honestly think the majority are P3's & celerons (of different flavors).

[mookydooky]

Seriously, it doesn't make sense to me. I'm awake now, so now my thoughts are less jumbled. Anyway, the obvious argument is that it's better to return two WUs than one if doing two takes proportionately less time than doing one. This is exactly what HT does. Taking it further, I would think they would appreciate getting further in two simulations than one.

The only way I can think of that it would be better to speed one up at the expense of another would be if results from one could would cancel results from another. In this case, they could eliminate possible dead end simulations a little faster. I find this hard to believe though.

[mookydooky]

Originally Posted by SteveI

Most PCs do NOT have HT. I honestly think the majority are P3's & celerons (of different flavors).

There is a distinct possibility though, that the simulations run on the fastest machines have the most value. Althought the vast majority of the installed base may run on P3's and slower Celerons, it may be the situation where the work done on faster machines are more important because they're further along.

Again though, I find this hard to believe because protein simulation is such a difficult task. The whole point of FAH is to distribute the work over a large network to get teraflops of computing power. Not taking advantage of HT seems counter to that scheme.

If it is the case that HT is a detriment, then it basically comes down to the client having been poorly written, especially since they don't say anything about seperate CPUs.

Seriously though, what's the difference between a genuine dual CPU or a virtual one? None as far as the client or the network is concerned. What are they going to do when dual core CPU's become the norm?

[mookydooky]

Originally Posted by keithwalton

Are they begining to run out of theories / units or somthing.

This is impossible. There are so many simulation not yet run, with current technology, it would probably take decades to simulate folding all the proteins they're interested in.

[keithwalton]

It still seems a tad strange and illogical to me that they would like donators to reduce there productivity. Strange also that they're still using alot of the old tinker tech stuff. The only reason for this must be that its impossible to convert it to a more modern core. Unless thats what the amber core is meant to be.
As far as i see it, points = productivity. increasing points would therefore indicate putting more effort into donating more