p638 Tinkers :-(
Moderators: NeilBlanchard, Ralf Hutter, sthayashi, Lawrence Lee
p638 Tinkers :-(
I have THREE of these darn things, one on each of three 3200+ Bartons . I'm getting just 571, 574, and 595 points per week . It might be time to try out the 'swap' script I wrote last week, to swap folders between two blades. It keeps the client.conf and speed.emd (an EMIII file) with the cpu, while swapping everything else including the queue.dat. I haven't used it yet because I'm not absolutely positive it works!
The idea is to work the less "valuable" work units with the slower cpu's, and the higher value WU's with the faster cpu's. Do more points quicker, fewer points slower.
The p638's have been the center of some controversy at folding-community.org, with quite a lot of people observing that they take just as long to do as the 249 point Tinkers, but are worth just 128 . Stanford even re-ran the benchmarks because of all the complaints, but claim they came up with the same results.
What's odd is, I removed -advmethods from three of my slower blades a few days ago, but those blades have gotten all Gromacs, while these three, with -advmethods have gotten Tinkers .
David
The idea is to work the less "valuable" work units with the slower cpu's, and the higher value WU's with the faster cpu's. Do more points quicker, fewer points slower.
The p638's have been the center of some controversy at folding-community.org, with quite a lot of people observing that they take just as long to do as the 249 point Tinkers, but are worth just 128 . Stanford even re-ran the benchmarks because of all the complaints, but claim they came up with the same results.
What's odd is, I removed -advmethods from three of my slower blades a few days ago, but those blades have gotten all Gromacs, while these three, with -advmethods have gotten Tinkers .
David
I've been tinkered too.
- Just when I had gone up to 2 CPUs,
- just when I made it above 75p/d and they change the coloring...
just one of those weeks I guess.
btw, the second CPU is the 1.3 MHz Centrino. And it's CRAWLING on that tinker. There goes the myth of the nimble Centrino. My 1.6 GHz pally (2000+) runs circles around it. Just lacking one gfx card to put it back into service....
David, did you make any simulations on your CPU/WU assignment practice? (giving gromacs to faster CPUs, tinkers to slower ones) daily it would matter of course but over a full week, does it make a difference? My Barton is doing a 240+ point tinker and I would guess it'll bring me more points than Gromacs for the same duration.
- Just when I had gone up to 2 CPUs,
- just when I made it above 75p/d and they change the coloring...
just one of those weeks I guess.
btw, the second CPU is the 1.3 MHz Centrino. And it's CRAWLING on that tinker. There goes the myth of the nimble Centrino. My 1.6 GHz pally (2000+) runs circles around it. Just lacking one gfx card to put it back into service....
David, did you make any simulations on your CPU/WU assignment practice? (giving gromacs to faster CPUs, tinkers to slower ones) daily it would matter of course but over a full week, does it make a difference? My Barton is doing a 240+ point tinker and I would guess it'll bring me more points than Gromacs for the same duration.
I've got p638 running on one of my P4 3.0C's and it's sloooooooow. About 7:07 per frame for 450 points a week when this folder should be doing double that. I was running 2 at a time (the other wu a gromac) and it was so slow I had to stop the gromac 'til the tinker gets finished. Total time should be about 48 hours.
I read thru the thread at folding-community.org that David referenced and got the impression that "they" had a cavalier attitude about the points assigned to p638 and the frustration expressed. Then the thread was locked.
To say nothing about the fact that my 2600+ Barton folds p638 50% faster.
There's something wonky about the way these wu's are assigned as P4's just don't cope with them at all (and, as you can see, neither do I.) Except for one, all the rest of my AMD folders are crunching gromacs. This makes no sense.
[/rant]
M
I read thru the thread at folding-community.org that David referenced and got the impression that "they" had a cavalier attitude about the points assigned to p638 and the frustration expressed. Then the thread was locked.
To say nothing about the fact that my 2600+ Barton folds p638 50% faster.
There's something wonky about the way these wu's are assigned as P4's just don't cope with them at all (and, as you can see, neither do I.) Except for one, all the rest of my AMD folders are crunching gromacs. This makes no sense.
[/rant]
M
FWIW, I have left the Tinkers on the 3200+ Bartons. I send log files to Ohairy1 occasionally, for his graphs, so I don't want to be changing cpu's in the middle. Besides, compared to other cpu's, the Bartons are blazingly fast on these WU's at "only" 36 hours. My frame times are in the 5:30 range.
burcakb, no I haven't "done the math" on whether it would actually make much difference, switching WU's from one cpu to another. Intuitively it makes sense, but I don't know whether it actually would result in more points or not.
David
burcakb, no I haven't "done the math" on whether it would actually make much difference, switching WU's from one cpu to another. Intuitively it makes sense, but I don't know whether it actually would result in more points or not.
David
Looks like the Tinkers had a big impact on us yesterday, as our points were off 10K vs. the day before. That's just a reflection of the Tinker units, which take at least 36 hours, even on the fastest hardware. I expect to see higher than average production on Sunday or Monday.
Looks like TRC-13, bkh, Frank_Condron, dasman, and D_Bailey all got hit BIG TIME with Tinkers yesterday.
Quite a number of teams have shown similar dips in their production:
Team MacOSX
Futuremark.com
SpeedGuide.net
Sudhian Media
David
Looks like TRC-13, bkh, Frank_Condron, dasman, and D_Bailey all got hit BIG TIME with Tinkers yesterday.
Quite a number of teams have shown similar dips in their production:
Team MacOSX
Futuremark.com
SpeedGuide.net
Sudhian Media
David
I don't know what I got, but beware, Tinker is very fast compared to THIS:
...on my Barton 204x11 machine I getting Tinker frame bellow 5minutes (4:50 mostly), but this one is notably slower...
So, it might be sometimes contraproductive to use the -advmethods, because these ones are so so slow... I hope there are enought points for completing this beast
It also seems that using -advmethods -forcesse -forceasm make the folding of this bastard a bit faster, however don't count on it...
...on my Barton 204x11 machine I getting Tinker frame bellow 5minutes (4:50 mostly), but this one is notably slower...
So, it might be sometimes contraproductive to use the -advmethods, because these ones are so so slow... I hope there are enought points for completing this beast
It also seems that using -advmethods -forcesse -forceasm make the folding of this bastard a bit faster, however don't count on it...
I have a few of these 128 point proteins. I don't mind them on my older folders, but my faster CPUs crank out the points usually. I think my Duron running at 2GHz is processing these as fast as my P4 at 2.53. And the P4 can get at least 3 times more points with the 107s. Having had a good run for a few days (while ColdFlame has been down on production), I'm now facing these AND hotter weather. So expect my output to be reducing. The good news is that my finacee now has broadband (well 150k, but 24/7), so the HTPC we have there (P4 2.0A) has done one or two proteins. However there's no way its going to fold 24/7, probably not even close...
-
- Posts: 154
- Joined: Thu Oct 17, 2002 5:53 pm
- Location: Sydney
David, I think that the best way to get the most points with the work units assigned to you is to have each unit folding on the machine that is best at it. That is it is very bad to have a P4 folding a tinker but and athlon is not that bad with them. So if you could switch a tinker from a P4 to an athlon and then move a gromac or even a Dgromac to the P4, there would be an overall advantge.
I suppose you only have 1 P4 blade but if you get a tinker on that you should try a switch to an athlon blade.
As for switching work units between cpu's I don't know if there would necessarily be a points benefit. I was thinking of exampleas and was able to cook up examples where it was beneficial to swap a slow unit to a slow cpu and examples where it would be counter productive! I couldn't be bothered to work out a formula to figure this out but if you're keen perhaps you could have a go. Anyway I can't see this as being worth fiddling with all the time as it would be time consuming for little gain.
Definitly try switching work units between cpu's of different type though, this could net a decent gain.
I suppose you only have 1 P4 blade but if you get a tinker on that you should try a switch to an athlon blade.
As for switching work units between cpu's I don't know if there would necessarily be a points benefit. I was thinking of exampleas and was able to cook up examples where it was beneficial to swap a slow unit to a slow cpu and examples where it would be counter productive! I couldn't be bothered to work out a formula to figure this out but if you're keen perhaps you could have a go. Anyway I can't see this as being worth fiddling with all the time as it would be time consuming for little gain.
Definitly try switching work units between cpu's of different type though, this could net a decent gain.
Use -forcesse on AMD, -forceasm on Intel. From everything I have read, use one or the other, not both.trodas wrote:It also seems that using -advmethods -forcesse -forceasm make the folding of this bastard a bit faster, however don't count on it...
Michael_qrt, all of my P4's run WinXP, all my Athlons run Linux, so switching between architectures isn't an option. I do have one P4 "blade", but it is connected to a hard drive and runs WinXP. I had some problems building an optimized LTSP kernel for this processor, and I just haven't gotten back to getting that figured out.
David
Okay, okay, haysdb i removing the -forceasm from my main Barton machine, however anyone have any clue what is best for VIA C3 Nehemiah?
I think I should start measuring how average frame took on the Tinker the poor C3 folding for pretty long now (and It's not even in half) with waht options.
So far I should try just -forcesse and/or -forceasm, right?
The -advmethods did not affect the folding speed, did they?
I think I should start measuring how average frame took on the Tinker the poor C3 folding for pretty long now (and It's not even in half) with waht options.
So far I should try just -forcesse and/or -forceasm, right?
The -advmethods did not affect the folding speed, did they?
trodas, as far as I know, you have the first C3 folding for SPCR. I would assume the C3 does not have SSE, but it's easy enough to find out experimentally. Just run a couple of frames with and without -forcesse and check the frame times.trodas wrote:Okay, okay, haysdb i removing the -forceasm from my main Barton machine, however anyone have any clue what is best for VIA C3 Nehemiah?
I think I should start measuring how average frame took on the Tinker the poor C3 folding for pretty long now (and It's not even in half) with waht options.
So far I should try just -forcesse and/or -forceasm, right?
The -advmethods did not affect the folding speed, did they?
You will also have to determine through observation, whether the C3 does better on Tinkers or Gromacs. I don't even have an educated guess on that one.
Something worth trying with the C3 is running FAH with the -config or -configonly options and setting the client preference to "GAH". This will cause "TimelessTinkers" to be assigned to this client. These are smaller work units, worth the same as the larger ones, which have no final deadlines.
FYI - For anyone with a cpu which does particularly well with Tinkers vs. Gromacs, this is how you can specify to FAH that you want ONLY Tinkers.
David
Last edited by haysdb on Mon May 03, 2004 11:47 am, edited 1 time in total.
You aren't the only one hit hard by server .108 being down. bkh is another.dasman wrote:Tinkers, combined with power issues at the office over the weekend left most of my boxes shut down for almost 2 days (you're welcome D_Bailey! )
It has been up and down since last night. I was getting the sigs just awhile ago, and now not, again.What happened to the ninja's sig? Getting the red X?
David
Moderator: OK, now this is really weird. I must have edited dasman's post rather than replying to it.
Never replied to myself before -- kinda differentdasman wrote:You aren't the only one hit hard by server .108 being down. bkh is another.dasman wrote:Tinkers, combined with power issues at the office over the weekend left most of my boxes shut down for almost 2 days (you're welcome D_Bailey! )
Actually, what I meant was power issues on my end leaving about 6 of my boxes down for a little over 2 days. Didn't even know a server was down too...
First laugh I've had all day -- and it's been a long oneModerator wrote:Moderator: OK, now this is really weird. I must have edited dasman's post rather than replying to it.
Dave
I have FIVE of these durn things now. Why me?
The p638's are giving me, in PPW:
298 - half of a P4 2.4C
599 - 3200+ Barton
579 - 2600+ T'bred
512 - 2400+ T'bred
430 - 2000+ Palomino
In contrast, three p1110 Tinkers are giving me
754 - half of a P4 3.1C
1043 - 3200+ Barton
744 - 1800+ Palomino
I.e. at least 75% higher PPW.
David
The p638's are giving me, in PPW:
298 - half of a P4 2.4C
599 - 3200+ Barton
579 - 2600+ T'bred
512 - 2400+ T'bred
430 - 2000+ Palomino
In contrast, three p1110 Tinkers are giving me
754 - half of a P4 3.1C
1043 - 3200+ Barton
744 - 1800+ Palomino
I.e. at least 75% higher PPW.
David
-
- Friend of SPCR
- Posts: 96
- Joined: Fri Dec 05, 2003 11:03 am
- Location: Seattle, Washington, United States
I got one of these barstewards completed on my Duron (running at 2GHz), and the server (171.64.122.143) had gone down. Well it's been up sine mid morning at least (as thats when the protein finally got sent back), and no points yet. Grrrrr.
[engage rant mode]
2 More grrr things are with 59 pointers... The Duron has gone onto a p1042_nat96_273, and made it to 99% before stalling. It's nearly an hour since the last log entry, so I hope it can get it complete, or I will be really pi$$ed off.
[edit - of course its clock is out a bit, and so it hadn't stalled, its just completing right now ]
And also today another 59pt protein, a p1046_nat96_337, being folded by my P4, has stopped at 19%, for no reason. I wish I'd checked it this morning because it hung overnight. Of course, my P4 that was happy to fold the 107 pointers at 1.25v and stock (2.53GHz) speed keeps aborting WUs until I got to this one and about 1.35v. It's hot, and undervolting is a tool I need to use to keep the box folding - I'd happily turn it off for 0 points...
[engage rant mode]
2 More grrr things are with 59 pointers... The Duron has gone onto a p1042_nat96_273, and made it to 99% before stalling. It's nearly an hour since the last log entry, so I hope it can get it complete, or I will be really pi$$ed off.
[edit - of course its clock is out a bit, and so it hadn't stalled, its just completing right now ]
And also today another 59pt protein, a p1046_nat96_337, being folded by my P4, has stopped at 19%, for no reason. I wish I'd checked it this morning because it hung overnight. Of course, my P4 that was happy to fold the 107 pointers at 1.25v and stock (2.53GHz) speed keeps aborting WUs until I got to this one and about 1.35v. It's hot, and undervolting is a tool I need to use to keep the box folding - I'd happily turn it off for 0 points...
Here you go David. This should get your blood to boil. A DIRECT side by side comparison of p638 and p1110 on THE SAME machine. It is a dedicated 24/7 Dual Athlon MP 1800. One F@H CLI client per processor. Drum roll please.....
From EMIII with newest protein data file (5/4/04)
p638_L939_K12M_ext 128.0pts
Time Per Frame 00:06:47
Time Per WU 45:13:20
p1110_L939_K12M_nat_min1 249.0pts
Time Per Frame 00:06:53
Time Per WU 45:53:20
It's a crazy, mixed up world we live in! Stanford's benchmark machine must be from bizzaro world!
From EMIII with newest protein data file (5/4/04)
p638_L939_K12M_ext 128.0pts
Time Per Frame 00:06:47
Time Per WU 45:13:20
p1110_L939_K12M_nat_min1 249.0pts
Time Per Frame 00:06:53
Time Per WU 45:53:20
It's a crazy, mixed up world we live in! Stanford's benchmark machine must be from bizzaro world!
When it rains, it pours! Another serendipitous side by side comparison on another machine. This a Dual Athlon MP 1900 running 24/7 for folding only. From EMIII again.
p638_L939_K12M_ext 128.0pts
Time Per Frame 00:06:24
Time Per WU 42:40:00
p639_L939_K12M_nat 239.0pts
Time Per Frame 00:06:39
Time Per WU 44:20:00
Perhaps Stanford should start using Dualies for benchmarking. Just a thought.
p638_L939_K12M_ext 128.0pts
Time Per Frame 00:06:24
Time Per WU 42:40:00
p639_L939_K12M_nat 239.0pts
Time Per Frame 00:06:39
Time Per WU 44:20:00
Perhaps Stanford should start using Dualies for benchmarking. Just a thought.
I posted the same info at Folding Community in the Client-General section with the p638 performance graph, just to see if anyone can explain the difference in points. Usually someone will pick up a question toot sweet, but nobody has commented thus far. Which is very peculiar. I've read all the other posts about how a machine is different from the benchmark machine, yada, yada, yada. But these are IDENTICAL processors on the same machine so there are no variables! Enquiring minds want to know.