Aha! Linux clients DO get more Tinkers

[haysdb]

In a thread at folding-community.org, Rebalancing toward more Gros, I commented that my P4/Windows machines with -advmethods seem to ALWAYS get Gromacs, whereas my Athlon/Linux blades with -advmethods get a fair number of Tinkers. I wasn't complaining, just observing, and "wondering out loud" whether this was a Intel/AMD thing or a Windows/Linux thing.

Something else I have observed is that some proteins seem to be assigned ONLY to my P4/Windows machines. There are enough samples of some of these work units that it goes beyond "chance."

Turns out there are proteins which are not assigned to Linux clients due to an issue with the "PME implementation" in the Linux version of Gromacs. This reduction in available work units in turn causes proportionately more Tinkers to be assigned.

There's nothing earthshattering or sinister going on here, and nothing for Linux users to be "alarmed" about, it's just me "observating."

According to Vijay, Gro 4.0 is supposed to address this issue and is supposed to be out "soon" (he said this last October) but it'll be awhile after that before it's integrated into F@H. Slowly but surely, the "Linux disadvantage" is erroding. For example, it used to be that running the Windows client under WINE (A Windows emulator) was faster than the native Linux core! That is no longer the case, but because of this ""PME implementation" issue, Linux remains a second class citizen. But maybe for not much longer.

David

[Stevo@ARM]

In a thread at folding-community.org, Rebalancing toward more Gros, I commented that my P4/Windows machines with -advmethods seem to ALWAYS get Gromacs, whereas my Athlon/Linux blades with -advmethods get a fair number of Tinkers. I wasn't complaining, just observing, and "wondering out loud" whether this was a Intel/AMD thing or a Windows/Linux thing.....

David

Well as you guys know it seems to me like ARM has been picked by Stanford to be the official poster child for Tinkers So you just know I had to jump in on this topic.

I have been checking our clients and have come up with the following observations. The overwhelming number of Tinkers we get are on our CLI clients, our Windows GUI clients get a lot more gromacs it seems and tends to keep getting them consistently. Our CLI clients get Tinkers over and over again without letting up.

Case in point, all 5 of my systems at home are Windows GUI clients and have been working Gro's pretty much all week. Yet at the shop, where we run the CLI clients for ease of installation and management we get boat loads of Tinkers and we even tried a bunch of Linux clients as well today and yep 20 separate Linux CLI clients ALL got handed Tinkers in a row and these clients are running on reasonably muscular 2.8GHz P4 HT Xeons with RedHat 9. Our other Xeons like the quad and some dualies are running Windows with CLI clients and they have never gotten a Gro that we know of even once in the last 14 days since we started.

Heck, I am wondering if certain CPU types might be flagged by the Stanford servers as good places to work Tinkers and they just wind up cluster-bombing them there. Today we fired up 18 separate XEONs running a total of 36 distinct clients all running simultaneously and EVERY single one got Tinkers, not one Gro in that bunch of 36! They are all P4 generation HT Xeons with SSE2 - 6 of em are 1.80GHz and the other 12 are 2.80 HT's.

Don't get too excited yet about all those points however, becuase I lost tonights production on 10 of the 2.80's when I fired up a P4 1.70 which happend to be on the same circuit as the 5 Dual 2.8 Xeons and tripped the breaker, which brought those 5 crashing to a halt.

The really sad part however, is that I am not a Linux guy, that is Pat's domain here, and he had already left to start his weekend and I didn't know what login passwords he setup on those babies, so they are sitting there idle not folding until tomorrow when I can go back into the office and Pat can guide me through getting them back up and Folding again. Also, these 5 Dual 2.8's are only temporary Folders, they'll have to be shipped out to customers early next week, but Folding 24x7 for about 3 - 4 days is a good burn-in test I say

So does any of this Xeon/CLI client Tinker theory jive with what you guys have been seeing? What kind of WU mix do you guys get for GUI vs CLI vs Screensaver?

Stevo
----------------
Tinkers?... No problem... BRING 'EM ON

[ColdFlame]

This is a punishment for using Linux

This is really interesting philosophically. People volunteer their CPU time yet they b!tch about scheduling, etc. (me included). I hope in near future we'll have more computing projects so we will be able to choose.

Reasoning is that I think Stanford folks are lagging behind in their scheduling algorithms and definitely in their algorithms in general (I guess I'm saying they are far from being efficient - this is just a guess though, based on academic experience).

[haysdb]

...ARM has been picked by Stanford to be the official poster child for Tinkers...

That's not even hyperbole. There is something seriously foobar somewhere in the assignment servers...

Code: Select all

IF $E(USERNAME,1,11)="ARM_Systems" {
  Give_Them_A_Buttload_of_Tinkers
}

or maybe the condition is

Code: Select all

IF $L(USERNAME)>X ...

You have tried, right, to change your username to just ARM_Systems and see what you get? :shrug:

AFAIK, nobody in the Wintel camp is bitching overly much about anything. They don't have the PME issue that effects Linux and they don't have the "SSE bug" that effects Athlons. For certain, if there were people getting the number of Tinkers that you are, they'd be raising bloody hell over at folding-community.org, where they can chew the ear off Vijay Pande himself.

David

[roym@arm]

...ARM has been picked by Stanford to be the official poster child for Tinkers...

That's not even hyperbole. There is something seriously foobar somewhere in the assignment servers...

Code: Select all

IF $E(USERNAME,1,11)="ARM_Systems" {
  Give_Them_A_Buttload_of_Tinkers
}

or maybe the condition is

Code: Select all

IF $L(USERNAME)>X ...

You have tried, right, to change your username to just ARM_Systems and see what you get? :shrug:

AFAIK, nobody in the Wintel camp is bitching overly much about anything. They don't have the PME issue that effects Linux and they don't have the "SSE bug" that effects Athlons. For certain, if there were people getting the number of Tinkers that you are, they'd be raising bloody hell over at folding-community.org, where they can chew the ear off Vijay Pande himself.

David

STEVO, Patrick and I spent a couple of hours after work Friday night and put on-line the following:

1 - P3 Celeron 500 - WIN XP - got Gromack
1 - AMD XP2100 - Win XP - got Gromack
1 - P3 500 - Win XP - got Gromack

We thought, WOW! Our luck is finnally changing

Then
1 - P4 1.7Ghz - Win XP - got sTinker
2 - P4 3.2Ghz - Win XP - got 4 sTinker clients
5 - P4 2.8 Dual Xeons - Linux RH - got 20 sTinker clients

We also currently have a QUAD Xeon running Win2003 ADV Server w/ 8 sTinker clients, 2 DUAL Xeon processors running Win2003 Servers w/ a total of 8 sTinker clients.

We noticed that only our slowest folders are getting Gromacks. Our faster machines are running sTINKERS.

I think STEVO is right. This is a conspiracy against ARM Systems. Maybe one of the admin guys at Stanford is a MacOS folder.

Well, we will just have to keep adding more firepower.

Roy

[ArtCubed]

-Advmethods doesn't really do anything IMO. I've been getting alternating Tinkers and Gromacs.

Macs, on the other hand, don't get tinkers at all. I loaded F@H on my iBook to see if I would get a tinker or not. And it got a Gromac, as did my Dual G5. The G4 has been getting Gromacs the whole time.

And we wonder why Team OS X got past us...

[Zyzzyx]

Hmm... I don't really remember the last time I had a Tinker. I think maybe a couple of WU ago on my P3-566. I'd still like to specify Tinks for my AMD w/out SSE code, it does better on them.

For comparison, my setups:

Home
1 - 1.4 AMD Palomino
1 - 1.4 AMD T-Bird (no SSE)
1 - 1.46 AMD T-Bred
1 - P3-566

Work
1 - P4 2.2
2 - P4 1.8
2 - Cel 2.2

All of these are running Windows 2000 with the command line client installed as a service.

Actually, thinking back, the systems at work I just added (the two pair) got Tinkers as their first WU, but all Gro since then. But its been a few weeks now.

[ColdFlame]

useful info:
Can you try this experiment - just run a single instance of F@H on a machine that usually gets Tinkers and see what you get? Make sure to clean /work directory, delete queue.dat and Core*.exe. Also make sure to specify -advmethods -forceasm -forceSSE -local in the command line. You'll see what you get immediately.

just some blah-blah:
I very rarely get Tinkers. They seem to come as a plague, hitting 2-3 out of 4 machines that I have. But most of the time I get Gromacs exclusively.

I, however, did this test once. I started one of my clients without an -advmethods flag and immediately got a Tinker. Then I stopped the client, cleaned up /work directory and added -advmethods flag. Result: got Gromacs.

So I my case -advmethods has _direct_ effect on what type of WU you get, most of the time.

Since you are getting Tinkers I bet your -advmethods isn't set properly or something. Maybe your dual CPUs are stomping on each other.

[haysdb]

No shortage of theories.

A check of the log files would reveal any invalid flags, such as if Stevo and company were consistantly typing in -advmetods or something.

David

[Stevo@ARM]

useful info:
Can you try this experiment - just run a single instance of F@H on a machine that usually gets Tinkers and see what you get? Make sure to clean /work directory, delete queue.dat and Core*.exe. Also make sure to specify -advmethods -forceasm -forceSSE -local in the command line. You'll see what you get immediately.....

... In my case -advmethods has _direct_ effect on what type of WU you get, most of the time.

Since you are getting Tinkers I bet your -advmethods isn't set properly or something. Maybe your dual CPUs are stomping on each other.

Thanks for the suggestions ColdFlame. I'll monkey around with some of your ideas and see what happens.

For clarity on our client setups:

This is what I copy onto our systems for the CLI clients, I just made a set of files as follows that gets copied to FAHx subdirectory (x=1 to 8 depending on how many clients we are running on that system):

client.cfg - has the setup info with Team 31574 and our ARM_Systems... Member name: With each Machine ID's set to x .
FAHx.lnk - Shortcuts pointing to the clientx folding directories with -local -service -advmethods -forceasm -forceSSE parameters.
FAH4Console.exe - version 4.0 CLI Folding Client
FahCore_78.exe - Gromacs Core (I figure I might get a better chance of hooking a Gro if I have that core already there when it fires up )

Also great news! I got those 20 Linux RedHat clients on the 5 Dual Xeon 2.80 servers back up and running this afternoon, so they should crank out about 1400 extra points for us sometime late Sunday night or maybe Monday.

Stevo
-----------------
Linux Folding @ ARM... it works when you know the login and password - thanks Patrick

[ColdFlame]

Your setups seems to be correct to my blind eyes Still it is very strange that you get so many Tinkers. When I was running my Linux machine I was getting Gromacs mostly.

This is why I suggested running just one client, preferably from the command line, without a .lnk file. I'm just trying to do a "unit test", try to isolate one particular instance of F@H client and see how it performs in isolation.

Do you have any Windows boxes? How to they behave? If you are getting mostly Tinkers on Windows boxes as well then I can really assure you that there is no mystery but some kind of cfg mistake or something