beta/phase 2 question

[jonnyw]

I'm confused as to what is happening when phase two comes in.

From reading through past threads I understand:

... in phase 2 we will be using the beta client that others have been testing.

Also all the stats will start again for this new phase (all though the old ones will still be available).

the client will pick the best structure from every X structures and then work on that.

Is all this true?

Also how will the new scoring system work?

Finally will the new client replace the old one automatically, or will we have to remove the old one and then d/load and install the new one.

I'd be very greatful if someone could take a few mins of their time to explain (sorry if these have been posted b4).

Cheers peeps

:)

[djp]

Hey, if you hadn't posted these questions, I probably would have later today.
:cheers:

The other question I had was this: When I run the Phase I client and press the "Q" key, it stops folding and transmits pending results. When I do the same thing with the beta client, it just stops folding. If I know that I'm going to have to rebuild my machine, I'd really want to transmit any pending DF work before I reformat the hard drive. How do I do this?

[shortfinal]

Originally posted by djp
When I run the Phase I client and press the "Q" key, it stops folding and transmits pending results. When I do the same thing with the beta client, it just stops folding. If I know that I'm going to have to rebuild my machine, I'd really want to transmit any pending DF work before I reformat the hard drive. How do I do this?

What switches are you using when running the beta? If sounds like you have nonet specified, '-i f' which means Internet set to false.

To just upload results and stop run the client with the switch '-u t' which means upload only set to true.

Shortfinal

[Paratima]

The new (currently beta) client sorta seems to upload results when it gets good & ready to. :p Nothing is lost when you stop. It just doesn't upload immediately.

For the rest, yes, yes, and yes.

The scoring gives credits at the end of each upload and a bonus on the final (250th or thereabouts) upload. IIRC

The PLAN is for the upgrade to happen automatically. Those of us who have been through a few upgrades will be watching closely. You may have noticed some of the old hands saying they'll just shut down & manually upgrade. Experience speaks; take it for what it's worth... ;)

[jonnyw]

cheers, just wanted to get a feel for whats going on etc

[FoBoT]

on the upload question, isn't there still a -u switch ? to force "upload only" ?

i hope dfGUI phaseII will include an "upload" button

[Paratima]

Jah, mein herr. There is still a "-u t" for upload=true. However,
The new (currently beta) client sorta seems to upload results when it gets good & ready to. Nothing is lost when you stop. It just doesn't upload immediately.Here's the deal. While it's working on a "pass" it stores intermediate results as it goes. These results do NOT get uploaded until the pass is done.

If, at the end of a pass, the client finds that (a) it can't talk to the Mother Ship or (b) you have set "-i f" (internet false, I think), then it buffers those results and continues. If you then stop the client and restart it, either with "-u t" set or at least without "-i f", then it will attempt to upload the buffered results from previous passes.

(whew!)

[FoBoT]

thanks Paratima


so, that leads to this question

approximately (minutes, hours, days, weeks) how long is a "pass", on say a 1 Ghz p3 ? (or any WAG is fine, i just want a vague/rough idea)

[Paratima]

Generally, umm, somewhere between six minutes and two days.

Approximately. :D Like, more or less.

[FoBoT]

:cool:

that is close enough

[Brian the Fist]

What they said :cool: The scoring will scale as the square root of the generation number - thus completing generation 100 is worth 10 times as much as completing generation 1, etc.
This is how the beta has been scored and it gives reasonable and more or less fair numbers.

[Paratima]

Originally posted by Brian the Fist
The scoring will scale as the square root of the generation number - thus completing generation 100 is worth 10 times as much as completing generation 1, etc. Aha! I knew it was more points out there somewhere bit didn't remember the details. Thanks, Howard. Good to know. :cheers:

[djp]

What they said The scoring will scale as the square root of the generation number - thus completing generation 100 is worth 10 times as much as completing generation 1, etc.
Wow! That means it will be very difficult for a Stats Ho entering the project late to catch-up with folks who've been folding since the first day, unless SH has one heck of a folding farm.

After 100 generations, the leaders will be earning 10s of points while a newbie is still earning fractional points for the same computing effort. It's good for encouraging people to stay with the program, but it's not for folks in their first week/month.

The stats don't matter quite so much to me (though it is nice to see that I'm making a significant contribution). I've liked the folks at Stanford for years, but I'm folding with DF because the design of the project's computing engine lends itself well to running on a farm that has no Internet connections.

If someone wrote a proxy server specifically for folding.stanford.edu, I'd be left with some tough choices, now wouldn't I? Fortunately, the Stanford researchers have made it clear that they don't want anyone or anything to buffer work units, so I don't think we'll see any dedicated proxy servers for their project.

[Paratima]

Originally posted by djp
It's good for encouraging people to stay with the program, but it's not for folks in their first week/month. Not really. The bias is in favor of (a) faster machines (as always!) and (b) people who stay on a protein longer, period. Faster boxen will be through the first 100 "passes" in a day or two, not a matter of weeks/months. A couple of days' or even weeks' headstart will only look good for a couple more days/weeks.

As for people entering the program after weeks/months being behind... Yep! But persistence pays! :smoking:

[tpdooley]

The beta client usually took about 5 days to get through to generation 250 on my win98se 256meg amd axp1800+ system. After gen 250, it starts over at gen 0, and hopefully picks a better protein to work with. ;) The generations lasted from lows of (15-30 mins?) to a high of 2.5 days.
Identical machines that started off at the same time were usually a few generations different by the time a few days had passed - so now we'll have rough estimates with a large variance for how long it will take a certain class machine to go from gen 0 to gen 250. :)

[djp]

Originally posted by tpdooley
The beta client usually took about 5 days to get through to generation 250 on my win98se 256meg amd axp1800+ system. After gen 250, it starts over at gen 0, and hopefully picks a better protein to work with. ;) The generations lasted from lows of (15-30 mins?) to a high of 2.5 days.
Identical machines that started off at the same time were usually a few generations different by the time a few days had passed - so now we'll have rough estimates with a large variance for how long it will take a certain class machine to go from gen 0 to gen 250. :) Is there a log where you were able to extract these statistics, or were these manual observations? If we've got good logging, I'll be able to compare actual performance over the several machines in my farm (as opposed to just running a -benchmark.) Also if we've got good logging, the dfQ clients can eventually upload their statistics to the dfServer and on to the dfUploader for collection.

[Paratima]

With the great variability between proteins, benchmarking is not going to work well at all. The best tool we have, besides the stats from Dyyryath and Statsman and MacNN and the like, is dfGUI!

[tpdooley]

djp: Those were stats compiled from state of the art seat-of-the-pants data collection - i.e. noting what generation the client was working on when I got to work, and where it was when I left. I had one of my systems take almost a day and a half on one generation - it was another of the beta testers that mentioned 2+ days on one generation.
Benchmarking doesn't do a good job with the new client. Identical systems don't perform identically. And the slow one often gets the better results. :)

[HaloJones]

It will be interesting to see whether with the new client, people decide to kill long generations with higher points in favour of starting over with fast generations for lower points. Beta testers tend to be more interested in the testing than normal crunchers who tend to be more interested in raw points.

[pointwood]

Originally posted by djp
If someone wrote a proxy server specifically for folding.stanford.edu, I'd be left with some tough choices, now wouldn't I? Fortunately, the Stanford researchers have made it clear that they don't want anyone or anything to buffer work units, so I don't think we'll see any dedicated proxy servers for their project. There is a good reason they don't do that and don't want anyone to do that. The reason is that new WU's build upon the previous WU's and when they get some completed WU's back, they are able to mark a lot of WU's as "not neccesary to crunch".

I regards to benchmarking - there will be a benchmark switch in the new client, so we should be able to easily compare performance between the different processors and platforms :cool:

This is almost starting to look like a FAQ for the new client. Maybe Howard will have a FAQ up, otherwise it might be a good idea to create a new FAQ based on this thread and make it a sticky?