239.0 pt monster!

ColdFlame · Post by **ColdFlame** » Mon Apr 19, 2004 5:49 pm

So I got a Tinker! A 239.0 point one. And my filthy Palomino 1.77 Ghz generates well over 1k PPW. Weehahahaa! It seems like Stanford has adjusted their point values already.

Here are the technical details for so oriented:
Protein p639_L939_K12M_nat
Protein Core Tinker
Credit 239.00
Deadline 45.00
Time Per Frame 5 mins, 47 sec
Est. PPD 148.77 (0.62 WUs)
Est. PPW 1,041.41 (4.36 WUs)

haysdb · Post by **haysdb** » Mon Apr 19, 2004 5:59 pm

It's true, the Tinkers will have to be worth some MASSIVE number of points in order to be equally as valuable as something like the 160 point Gromacs.

This is going to tip the scales a bit. Teams with lots of "Tinker folders" are going to get a HUGE boost. People like myself, with our armies of high-end processors, stand to gain next to nothing. I'm not complaining, just making the observation. For example, I have heard of folders with huge numbers of P3 systems. They could see their PPD go up by a factor of 3 or 4.

It's going to be a new playing field.

David

unregistered · Post by **unregistered** » Mon Apr 19, 2004 6:04 pm

I wonder what kind of impact this will have on flags. If you have a pentium, a Tinker is a sTinker, but for an AMD it may be a plus. This changes everything!!!!

ColdFlame · Post by **ColdFlame** » Mon Apr 19, 2004 6:20 pm

unregistered wrote:I wonder what kind of impact this will have on flags. If you have a pentium, a Tinker is a sTinker, but for an AMD it may be a plus. This changes everything!!!!

What do you mean? Both Pentiums and AMD support SSE. At least decent AMDs

haysdb · Post by **haysdb** » Mon Apr 19, 2004 6:24 pm

Check out the Points column.

Code: Select all

Proj Server IP       Work Unit Name      atoms  Pref  Final  Points  Frames Code
639  171.64.122.143  p639_L939_K12M_nat   393  45.00  67.00  239.00  400    TINKER	
678  171.64.122.143  p678_L939_K12M_355K  393  46.00  68.00  242.00  400    TINKER
680  171.64.122.143  p680_L939_K12M_414K  393  45.00  67.00  237.00  400    TINKER
687  171.64.122.143  p687_L939_K12M_473K  393  45.00  66.00  235.00  400    TINKER
688  171.64.122.143  p688_L939_K12M_444K  393  45.00  66.00  236.00  400    TINKER
689  171.64.122.143  p689_L939_K12M_503K  393  44.00  65.00  231.00  400    TINKER
690  171.64.122.143  p690_L939_K12M_532K  393  44.00  65.00  231.00  400    TINKER
692  171.64.122.143  p692_L939_K12M_int   393  46.00  68.00  241.00  400    TINKER	
698  171.64.122.143  p698_L939_K12M_600K  393  41.00  60.00  214.00  400    TINKER
699  171.64.122.143  p699_L939_K12M_700K  393  41.00  60.00  213.00  400    TINKER

Will there still be an advantage in -advmethods?

David

unregistered · Post by **unregistered** » Mon Apr 19, 2004 6:34 pm

What do you mean? Both Pentiums and AMD support SSE. At least decent AMDs

From what I have read pentiums suck at tinkers. SSE doesn't have much affect on a tinker but 3dnow does.

isp · Post by **isp** » Mon Apr 19, 2004 6:35 pm

Good observation David, this team is well educated with the flags and we all are sad to see a tinker slip through onto one of our folders these days...

The ones out there with just as many boxen as us, but with clients not configured for maximum points so to speak...are going to come out of the closet and it will be scary!

edit: Also just noticed p212_villinURE212 was modified to 135 it appears.

ColdFlame · Post by **ColdFlame** » Mon Apr 19, 2004 6:41 pm

It seems the answer to your question is NO David

And I got thru 4k barrier today

(on a 2k PPW farm)

If not for buggy FAHLogStats that calculates PPW incorrectly at times I'd show you a screenshot.

Mayhem Mayhem!!!

haysdb · Post by **haysdb** » Mon Apr 19, 2004 6:45 pm

We will just have to be very proactive in understanding the new rules of the "game" and take advantage of them as best we can. It may be that there will BE no advantages to be gained with the flags, but if there are, we need to be on top of it. For example, if it happens that -advmethods is actually a DISadvantage for some machines, then we need to get the word out to people to remove it.

I'm glad this change is happening. I was getting a bit bored with the status quo.

David

1911user · Post by **1911user** » Mon Apr 19, 2004 7:13 pm

I just checked rig#1. It has a p680 tinker worth 237 points. The frame times are about 4 and half minutes. That works out to ~1300 points per week

-advmethods is gone!

Should I also take out -forcesse? It helps with gromacs, but tinkers???

EDIT: I checked the log file and nothing is mentioned about SSE being enabled with the tinker core, so -forcesse stays for now. If all of the AMD machines start getting tinkers, I might get a darned gromac on occasion

haysdb · Post by **haysdb** » Mon Apr 19, 2004 7:52 pm

Definitely KEEP -forcesse. -advmethods is a different story. Gromacs and Tinkers should now be worth the same, so in theory there should be no advantage or disadvantage to using or not using this option, but I'm waiting see see how close theory mimics reality.

David

We may have to stop calling them sTinkers.

sonofdbn · Post by **sonofdbn** » Mon Apr 19, 2004 8:11 pm

haysdb wrote:We will just have to be very proactive in understanding the new rules of the "game" and take advantage of them as best we can. It may be that there will BE no advantages to be gained with the flags, but if there are, we need to be on top of it. For example, if it happens that -advmethods is actually a DISadvantage for some machines, then we need to get the word out to people to remove it.

This might be even more important in getting older PCs into the team. One thing which might put some people off folding is the low score after hours/days of grinding away on an old PC. If they can now get a decent score they might be encouraged to put them to work.

I have PCs with a PII, a Celeron (is that specific enough? - it runs at 1.4GHz) and an AMD-K6-2 not doing anything very useful. Can anyone recommend what flags I should be using for each of them under the new scoring system?

haysdb · Post by **haysdb** » Mon Apr 19, 2004 8:21 pm

sonofdbn wrote:I have PCs with a PII, a Celeron (is that specific enough? - it runs at 1.4GHz) and an AMD-K6-2 not doing anything very useful. Can anyone recommend what flags I should be using for each of them under the new scoring system?

In this case, the new scoring system doesn't change my recommendation for the flags since this cpu doesn't have SSE, so -advmethods would never have been recommended anyway, since Gromacs would not have been worth more than Tinkers.

You need no flags at all for this cpu, unless you want to see lots of detail in the log, in which case you can use -verbosity 9

David

mas92264 · Post by **mas92264** » Mon Apr 19, 2004 9:05 pm

Oh. I've got a few of the 239 pt Tinkers running now. Only worth >1,100 ppw on my amd boxes, err, boards. P4's quite a bit slower

but still doing >1,000 ppw running 2 (Tinker & Gro) at a time.

Why the heck doesn't XP let me in the program files directory of a networked box when W2k will? I've got the whole friggin drive shared along with the program files directory and folding directory - still can't get in.

Have another dead dedicated folding box. Gets to the Windows 2K screen with the progress bar, then flashes a bsod for about 1 sec with some text that goes by too fast to read then re-boots. Guess the os is corrupted.

Sheesh.

Michael_qrt · Post by **Michael_qrt** » Mon Apr 19, 2004 11:53 pm

I'm going to suggest that all AMD machines would now be better off folding tinkers in the long run. It takes my stock 2500+ about 48 hours to complete the average tinker which means about 120 ppd. On the other hand the only gromacs that give me that sort of ppd are the new 160 pointers, most are around 70-80 ppd.

With that in mind I'm going to try turning off the -advmethods flag and see if I can get some tinkers. It seems that since the new reference machine is a P4 and P4s are a bit slow on tinkers the points on tinkers now heavily favour AMD machines.

haysdb · Post by **haysdb** » Tue Apr 20, 2004 12:20 am

So, now we need a -Tinkers option, or maybe -oldmethods

Michael, from what I am seeing, you may be right. I have NEVER seen 1K PPW from ANY of my XP's, but right now TWO of them are doing 1K PPW on Tinkers according to LogStats, but only with 237 and 239 point units. Two blades running 128 point Tinkers are showing only 500 PPW. The frame times on these two groups of proteins are nearly identical, leading me to wonder if there has been a mistake in the point values.

Something else that doesn't add up for me is WHY Tinkers would run better on Athlons?

I want to give it a few days and see how things shake out before making any blanket recommendations. Tinkers are now worth more points, but P4's and T'breds and Bartons are still optimized for Gromacs, meaning more "science" gets done with them working Gromacs, even if the points are the same.

If I were betting though, I would "bet" that -advmethods will prove to be "obsolete" in light of the new points system.

David

Stevo@ARM · Post by **Stevo@ARM** » Tue Apr 20, 2004 12:41 am

haysdb wrote:
...Something else that doesn't add up for me is WHY Tinkers would run better on Athlons?...

David

Athlons have a far superior native floating point unit, and a much shorter instruction pipeline to boot, plus a larger L1 instruction cache. Therefore when executing traditional FP instructions (read non-SSE / non-3DNOW!) all AMD processors easily outperform their P4 counterparts by a fairly wide margin. And Tinkers are written in 100% original floating point code - so based upon the new points system (if it stays in place as it is now), you could say that Tinkers just might be an Athlon's best friend.

Stevo

haysdb · Post by **haysdb** » Tue Apr 20, 2004 12:51 am

Stevo@ARM wrote:based upon the new points system (if it stays in place as it is now), you could say that Tinkers just might be an Athlon's best friend.

Ironic, isn't it?

Thanks for a definitive answer to my question Stevo.

Now, seriously, how does one "attract" Tinkers? Removing -advmethods is obviously a good start, but as I have mentioned elsewhere, I need more data before I am prepared to welcome Tinkers with open arms. It's a little like being told that an old enemy is now your friend.

David

samwc912 · Post by **samwc912** » Tue Apr 20, 2004 1:05 am

I've got a P4, should I try running tinkers? Or go on with gromacs? How long would a tinker run on a P4? Same amount of time as those 160 points gromacs?

haysdb · Post by **haysdb** » Tue Apr 20, 2004 1:16 am

samwc912 wrote:I've got a P4, should I try running tinkers? Or go on with gromacs? How long would a tinker run on a P4? Same amount of time as those 160 points gromacs?

In theory, since both Tinkers and Gromacs are benchmarked on a P4, all work units should be worth the same PPD/PPD on a P4 system. I am leaving -advmethods on all of my P4 systems since I would rather run Gromacs on them, because I like the gratification of uploading completed work every day or even several times a day, as opposed to only one every 2 or 3 days. Too many eggs in one basket with Tinkers, assuming they aren't worth any MORE than Gromacs.

In practice? We will see.

David

Wrah · Post by **Wrah** » Tue Apr 20, 2004 1:28 am

haysdb wrote:Now, seriously, how does one "attract" Tinkers?

Try switching from fah to gah. I read somewhere that they're sending out small tinkers to gah boxes and I started using that at work where I've got some slooow p3's folding that need up to 2 weeks for regular WU's. And it seems to work so far. Strangely though the projects are rarely listed on stanfords site. The pc I'm writing this on is now working on a 697, again not listed. I have no idea how many points it is but it can't be more then 10. So that's the downside.

Michael_qrt · Post by **Michael_qrt** » Tue Apr 20, 2004 1:44 am

At this stage I think I'm not going to try to do anything funky with the system. I'll just experiment with removing the -advmethods flag and see what they give me.

also anyone with older thunderbird core athlons should be cheering now as they lacked SSE support and hence were much slower on gromacs than machines with SSE. Some older machines could come alive and start producing at a decent rate now.

I'm just thinking how to work these changes in our favour in the race against EOC. I think it's probably pretty safe to turn off the -advmethods flag on AMD machines as you may still get gromacs and these new tinkers seem to be etter in terms of ppd. But I guess the safest thing is to not change what's been shown to work previously untill there's evidence one way of the other. What I'm wondering is whether the -advmethods flag affects which gromacs you get?

Also, I know its against the wishes of the Stanford group but one way you may be able to get more tinkers isto run multiple clients on the one machine, hence lowering the performance fraction. Though personally I have no interest in this as it only complicates matters and isn't good for the science. Just a thought.

haysdb · Post by **haysdb** » Tue Apr 20, 2004 2:03 am

Michael_qrt wrote:What I'm wondering is whether the -advmethods flag affects which gromacs you get?

I don't believe so, no.

Also, I know its against the wishes of the Stanford group but one way you may be able to get more tinkers isto run multiple clients on the one machine, hence lowering the performance fraction. Though personally I have no interest in this as it only complicates matters and isn't good for the science. Just a thought.

In order to get a PF below .8 on a fast machine, you'd have to run more than two clients. Let's take the following project as an example:

Code: Select all

639  71.64.122.143  p639_L939_K12M_nat  393  45.00  67.00  239.00  400 TINKER

With a 67 day deadline, you would have to return it in no less than 14.4 days to avoid a PF > .8
Lets say your cpu can complete one of these in 2 days. You'd need to run maybe 6 or 7 clients in order to keep the PF of all of them under .8. I'd rather work one WU quickly than 6 or 7 very slowly, and this is better for the F@H project also.

David

geordie · Post by **geordie** » Tue Apr 20, 2004 3:40 am

I guess the new points system will mean people folding with the google toolbar plugin will get a massive points boost too. Tinkers all the way with google.

I don't suppose a huge number of google users hack the registry to fold for their own team, even though the toolbar is a simple way of borging machines. I bet more than the odd corporate network has been assimilated

trodas · Post by **trodas** » Tue Apr 20, 2004 3:53 am

So, the conclusion for below average user, like myself is - don't need to run folding with any special switches, right now, yes?

haysdb · Post by **haysdb** » Tue Apr 20, 2004 4:06 am

trodas wrote:So, the conclusion for below average user, like myself is - don't need to run folding with any special switches, right now, yes?

-forcesse is still recommended for AMD Athlon XP processors with SSE (Palomino, T'bred, and Barton).
-advmethods is no longer recommended for Athlon processors and may be obsolete for everyone.

David

dukla2000 · Post by **dukla2000** » Tue Apr 20, 2004 2:06 pm

I agree with most of what is said above:
- if you have an Athlon keep the -forceSSE (in case you get a GROMACS)
- if you have an Athlon ditch -advmethods to improve chances of Tinkers
- Athlon FPU is much better than P4

Stunningly, while my Athlon ppd is up for Tinkers, the power consumption and temps are down (because SSE is not in use). But equally I agree there is probably a transition period with plenty of chaos: at the top of the thread Coldflame is > 1k ppw on a 1.77GHz Palomino. My 2.167GHz Tbred B is on p697 for 238 points at about 960 ppw. While this is way over the 500ppw I used to get on Gromacs, I want one of those WU Coldflame has!

In a week or so we will need another thread comparing performances to see what is really shaking out.

unregistered · Post by **unregistered** » Tue Apr 20, 2004 4:06 pm

I'm waiting for their server to upload my 239pt tinker now! They are better at dishing them out than taking them lately. Queue is looking good, but My PPD is Zero. My PPW ain't much better either.

Maybe I'll have a 500+ day tomorrow, maybe there will be free beer tomorrow??????????

I used to be BLUE

Putz · Post by **Putz** » Tue Apr 20, 2004 5:47 pm

You say the new reference system is a P4... but Pande requests that users run only a single client on each, right? Have there been any indications that running two clients on a HyperThreaded machine skews the reference Tinker vs. Gromacs points ratio?

haysdb · Post by **haysdb** » Tue Apr 20, 2004 6:24 pm

Putz wrote:You say the new reference system is a P4... but Pande requests that users run only a single client on each, right? Have there been any indications that running two clients on a HyperThreaded machine skews the reference Tinker vs. Gromacs points ratio?

Running two clients on an HT processor will increase your points, but deliver the work to Stanford just over half as fast. More work gets done, but in a less timely fashion, so I suppose if the Pande Group had their druthers, they would probably prefer to see one client per cpu, regardless of whether it's HyperThreaded, but they are fighting a losing battle on that front and they know it.

David