Now past 75%

[MerePeer]

We passed the 75% mark!

Total work done so far: 7,939,230,253
Goal amount of work: 10,000,000,000

I estimate completion on St. Patrick's day.:|party|:


One sour note is the moo team is pressing hard, averaging about twice our output.

http://stats.zerothelement.com/cgi-...Create+My+Graph

[rofn]

hopefully...

this slow protein is driving me mad... :sleepy: :bang:

[the-mk]

That's why I changed to the RC5-72 Gauntlet

[rofn]

i was thinking about that, too

but don't want to loose the hard earned work i did for closing up on gopher...

perhaps i'll drop over if the next protein is as slow as this one...

i'm in anyway not really happy with the whole df client & project...

client who is messing up filelist.txt...then sometimes the previous generation is suddenly missing when it goes over 150 gens where the points are getting more...
the whole pointage system is crap in my eyes....
no proxy solution...or even a good queuing system...
poor 3rd party programs...

:(

we will see...

rofn

[Moogie]

ugh rofn...sorry you are having so many problems. Maybe one of the folks here will be able to give you some ideas. Have you actuall posted this in a new post?

Hang in there!

[rofn]

actually these are not so big problems for me...

the messed up filelist.txt was already discussed here:
http://www.free-dc.org/forum/showthread.php?s=&threadid=5291
so i think things will get better when the new client is out...i got that error two or three times since i'm crunching df...

what the problem with that previous generation missing is...dunno...
perhaps it had something to do with my oc'd system...where it happend 2 times till now...all on the current protein...even the sys is/was totally stable...things are goin now...

i also got that once on a p4 2ghz laptop from my girlfriend...and since she's always on the run i usually can take a look at it only once a week...so i lost about 230 gens there...purgeuploadlist didn't work...no chance to get the points...

i once thought maybe it was a problem with my harddrive which just didn't save one generation...but since i got my new sata drive and got that again i don't think this was the problem...

i'll have to take a closer look at it when the error comes up again...

anyway...

the whole scoring method is a bit confusing for me...especially when crunchin on slow proteins...where some generations take about 20 mins or more to complete...(xp2200)
i don't understand why you should get more points for e.g. a 240. generation as for the first till whatever...i think same pointage for all generations would be just fine...
and also for that 9999 structure gen where you get almost no points...but take longer to complete then about 20 real gens on a fast protein...
i even dont understand the use of that 9999 gen crunch...

the story about the personal proxy solution like dnet....
i once had that thought in the "chasing pack" thread in trash talk...
so i decided to post it in the official df forum for getting some answers about that idea from the developers...
http://www.free-dc.org/forum/showthread.php?s=&threadid=5768
but like fozzies comment in his plea for more time over the changeover

A simple "sorry no" would have been good manners.
no answer there...
only some talk with the dfq creator...which isn't a real solution for me just a nasty work around...

so you know now why i came up with "poor 3rd party programs"
sure dfgui is really good work, i wouldn't crunch without it...
and dfMon is also good stuff, like DCMonitor...but when you come from projects like seti or dnet you are a bit regaled from the amount of 3rd party programs and the quality of them...

to come to an end...i'd say...i'll race that gopher down anyway :P
greets
rofn

[the-mk]

Originally posted by rofn
to come to an end...i'd say...i'll race that gopher down anyway :P

Just do it :D

Bring it on :thumbs:

[bwkaz]

Originally posted by rofn
i don't understand why you should get more points for e.g. a 240. generation as for the first till whatever Because the later generations take longer to crunch. If you got the same number of points for all the generations, people would be letting their client go to gen 50, then start it over -- to stay within the fast-crunching area -- and they'd get more points than somebody that ran it the way it was meant to be run.

This would be crappy for the science, because there's more value in a full set than in 5 independent copies of the first 50 gens (it has something to do with the "folding movie" that you may or may not have seen links to on the DF site on your personal stats page).


and also for that 9999 structure gen where you get almost no points...but take longer to complete then about 20 real gens on a fast protein That's there to give your client a seed. It tries to improve on that seed in gen 1, and then uses the best structure in gen 1 as the seed for gen 2, and so on until gen 250 is done. Then it chooses a new seed by generating another ten thousand random proteins. It's basically the foundation of the whole process, and it didn't used to take the same amount of time as 20 normal generations. At least, not that I remember anyway.

[rofn]

Because the later generations take longer to crunch. If you got the same number of points for all the generations, people would be letting their client go to gen 50, then start it over -- to stay within the fast-crunching area -- and they'd get more points than somebody that ran it the way it was meant to be run.

sure...but what about that:
i sometimes crunch on late generations about 30 to max. 60 minutes...when they are really slow...

so lets say that gen is # 200
dunno if my dfgui is right on that but for generation 200 i would get 1414 Points
so lets say it takes 45 mins to complete...

when i crunch on generation 100 i would get 1000 points
so lets say we crunch on that 22 mins (half like before)
for gen 101 1004 points
so i would get in the same amount of time 590 points more on that lower crunch...

perhaps the times are not quite accurate...but what i'd like to say is that the scoring method souldnt be only depend on generations...

it sould depend on which generation & time it took to complete (calculated on performance of the box with a benchmark of the client itself) it got that benchmark thing so why don't use it for scoring too?


it didn't used to take the same amount of time as 20 normal generations. At least, not that I remember anyway.

on my xp2200 (@ xp2400 ocd) it takes almost 1:30h afaik (current protein)
so 100 (or 50?) points is a bit less...
with an average time of crunching of 10 min per gen (gen #1 to 9) i would get about 1081 points for the first 9 gens...(if i didn't misscalculated that yet)

all # of points for gens taken from dfgui's calc...dunno if they are right...

[Paratima]

Yer floggin' a dead horse, rofn. This issue has been re-hashed ad nauseum. Re-read kaz's sig line. :D

[Anteraan]

rofn,

While I admit that the current point system is very biased towards the latter generations, I just keep in mind that the low RMS numbers come from the higher gens (the top 10 RMS structures are all gen 250, the great majority of the time) -- so that's enough for me. Howard and Elena want to encourage us to follow through to gen. 250, since that is (as I understand it) where the best science is, so I'm good with that.

I also find that crunching can be highly variable in the time needed to knock out gens of any number. In other words, there are times that one machine will just go nuts on gens 225-240 and knock them out in 5 mins each, while that same machine struggled getting from 75-90, and spent ~20 mins on each.
(yes, I also run CB, and I observed this trend long before I started CB, so I've controlled for that variable)

Does the scoring system, thus, always reward the person with the most GHz on the project? Over time, I believe it does. Could it be better? I think so, but I also find it completely adequate for its intended purpose. If I just wanted a GHz counter, I could figure that out myself. If the points of individual generations bothers you greatly, just try to think in terms of 250 gens equalling 264,270 points. The only time we don't run full sets of 250 is during a changeover, true? By that point, we have run many, many sets of 250 (~36-38 for you on this protein by 3/17), so the leftovers aren't that huge of an issue to me.

I do understand frustrations with the client, especially the loss of buffered generations. It has only happened to me twice, but when it did happen, I was quite annoyed over the lost work. But, in the grand scheme of things, it falls into my "accepted variability", if you will. If it happened often, I'd be ticked, and looking to hammer someone. :bonk: Bugs are part of computing, and as long as I feel that Howard and Elena are continuing to work on eliminating these issues (and maintaining communication), I'll stick with it. Overall, my experience with the client has been much more positive than F@H, for instance.

Once again, these are my opinions only, and I'm not trying to tell you how you should think about things. Consider it an alternate view of the situation. It's the best I can offer you -- well, that and an opportunity to RK Gopher. :jester:

P.S. It looks like you'll hit 10 mil for this protein, so pre-congrats!

[bwkaz]

Originally posted by rofn
so lets say that gen is # 200
dunno if my dfgui is right on that but for generation 200 i would get 1414 Points Sounds right. It's 100 * sqrt(gen #), and sqrt(2) is 1.414, which means sqrt(200) is 14.14, and that times 100 is your 1414.


so lets say it takes 45 mins to complete...

when i crunch on generation 100 i would get 1000 points
so lets say we crunch on that 22 mins (half like before)
for gen 101 1004 points
so i would get in the same amount of time 590 points more on that lower crunch... Assuming that your "half like before" estimate (the time one) is correct. It's not always, and I have a feeling that it's not even correct most of the time.

If it is correct, then yes, the point system favors lower generations. But how much more would it favor them if it was a flat 100 points (or whatever) for each gen?


what i'd like to say is that the scoring method souldnt be only depend on generations...

it sould depend on which generation & time it took to complete (calculated on performance of the box with a benchmark of the client itself) it got that benchmark thing so why don't use it for scoring too? What benchmark thing? The ./foldtrajlite -bench argument? That's a calculation based on one fixed set of data, I think, and IIRC it doesn't even use the current protein (though it might).

If you mean that the client should keep track of the amount of time it took to do each generation, that can't be put into the point calculation because the stats server does the point calculation, not the client. Basic rule of security with network protocols: do as many calculations as humanly possible on a machine that you trust. That means the server, because it's possible to reverse engineer protocols and write a fake client. (This is why so many Quake 1 cheats came out right after the engine was open sourced -- the Quake engine sent too much data to the client, presuming that the clients were all honest. But they weren't, so people got the ability to see through walls, around corners, and the like. Had the server removed all objects that the client couldn't see, these cheats would not have been possible.)

And I don't understand any of what you're saying about "calculated on performance of the box with a benchmark...", could you maybe rephrase that? Or an example? I have a feeling that it's just way too complicated of an idea, but I'd like to know exactly what you mean before I say that for sure. :)