time needed to compute generation

[wirthi]

HI,

no bug, no feature, just something I wonder about.

On one of my machines (P4, 2.2 Ghz, 512 MB Ram, -rt switch) it usually took about one hour to complete a generation. Now I've reached generation 80, and things are speeding up. Generations are now computed in less than 5 minutes, some even in less than 2 minutes.

I know the time needed to compute a generation varies but it seems a bit strange to me how it varies. 2 minutes compared to 60 minutes IS a difference. I had thought that generations with higher numbers take longer to compute.

Ok, another thing I just saw: the energy levels are getting better and better (what's a good thing of course). The last one was just higher than 6 (6.006). Perhaps proteins with this energy level are easier to compute than with an energy level of 7 for example?

Nevermind .. at least I'm improving in stats now ... hehe

Greets,
Wirthi

[Shaqlin]

Generation time seems to go up and down... Sometimes it takes over hour and sometimes 5min. My averege time is 25min which sure tells how slow am I running this now... And my comp is [email protected], 512mb PC-2700 and Windows XP + SP1

[Louis]

I've actually seen things go both ways.

On my 1GHz 128MB work machine, I'm only at gen 73 after a week. This one's averaging about 2 hours per gen, but of course when I'm watching it takes about 5 hours. Seems like the energy is still trending down, but hard to say since it's so slow.

On my 1.7 GHz 640MB home machine, I went through the first 250-generation cycle in about 4 days, with a whole lot of the 'middle' generations (about 75-175) taking a couple of minutes. The energy graph had a distinct hump in the middle, from around 45 to 150, with a high point near 120ish, but then it went down again. The energy at gen 250 was the overall best.

As always, your mileage may vary - seems like home is taking about twice as long for the second cycle as the first. Oh well.

Louis

[Keller]

On my P3 667 (256 mb ram) i needed 24 hours to complete only 6 gens (92-98)!!! (the average time per gen is 4h ... ) i think this is really to much.
I can´t believe this isn´t a bug, because in this time, there could easily the triple amount of structs be computed ...

[Stardragon]

Just a few reminders and clarifications :idea: on the completion times for each generation:

Keep in mind that proteins are bound to get stuck in any generation, and the likelihood of this occurring is random across generations, so some may be very fast and some may keep getting stuck.

At the same time, once the protein has undergone an extensive amount of getting stuck in one or a few generation, the laxness levels will have increased significantly, and they will be maintained for the remaining generations up to gen. 25. This will usually account for a general speedup trend after an initial slow completion rate.

For more detailed information, please read http://www.distributedfolding.org/phaseiifaq.html under the section "My protein keeps getting stuck"

[pfb]

My Barton @ 2800+ has yet to take longer than 36 hours to do a complete 251 gen cycle - the lowest RMS it has produced during that is ~6.5...whereas the P4 I have is taking twice as long and gotten down to 5.46...

One thing I have noticed is that the laxer the structure, the lower the RMS I'm getting in around 80% of the cases...

[FoBoT]

people don't seem to be getting the difference between phase I and phase II

phase I - produce as many random structures as possible as fast as possible in hopes of randomly producing a useful structure

phase II - produce a "best" random structure during generation 0 to use as a seed for a "smart" (non-random) process that try's to improve the gen 0 structure during the course of the following 250 generations

i.e. - it is about getting a better stucture, not how fast you can produce them. it is ok/expected that your client will get stuck, sometimes for a LONG TIME under phase II

just my $0.02

[wirthi]

Originally posted by FoBoT
people don't seem to be getting the difference between phase I and phase II

phase I - produce as many random structures as possible as fast as possible in hopes of randomly producing a useful structure

phase II - produce a "best" random structure during generation 0 to use as a seed for a "smart" (non-random) process that try's to improve the gen 0 structure during the course of the following 250 generations

...
Hey, that's exactly what I'm talking about. I had thought that better structures (with a lower energy level) take longer to compute.

Right now I'm monitoring the same thing as Louis: Energy level dropped til Gen. 80 and started to rise again at about 90. Now structures take longer time to be computed.


i.e. - it is about getting a better stucture, not how fast you can produce them. it is ok/expected that your client will get stuck, sometimes for a LONG TIME under phase II

just my $0.02
You are right; but you forget that most people just care what their position in the stats is and thus, how many points they gather (I include myself). Right now only the number of generations count when the points are calculated - so it's quantity that matters, not quality. Of course that's not what the project designers want - I think the quality should somehow be included in the point-calculation-formula - you should get more points when the quality of the protein is better.

just my ¤0.02

[pfb]

Originally posted by wirthi
Hey, that's exactly what I'm talking about. I had thought that better structures (with a lower energy level) take longer to compute.

That's what I'm noticing with mine (as seen with the timings I posted) + the fact that it seems having one lax value at ~50% is also driving down the the RMS...

[Louis]

I don't watch my work computer that close, but it seems that it will try one fold for 15-20 minutes without changing the laxness. But maybe that's just in my head - I took a lot of math in college, so now I can't really tell the difference between the numbers in, say, structure 26 and 27. My mind sees a 2 in there, so they must be the same. And yes, sometimes 7, 27, and 47 are the same too. (A mind is a terrible thing...)

I'm making the switch from 'some is good and more is gooder' to 'size matters, but not how you think it does'. Whatever it takes to get an energy of 6, or 5, or 4, or....

Louis
(who still can't believe 1GHz isn't fast)
(and who is old enough to remember when 66MHz meant 'server')