A64 and P-M and QMD work.

[muddocktor]

Since I have been getting an abundance of really crappy point Gro and also a lot of Tinker work on my A64's and my P-M laptop and desktop rig, I decided to do a little experimentation with QMD work on them. I d/l'ed some QMD work on 1 of my P4 machines, then transferred them over onto my P-M laptop, my desktop P-M machine and to 1 client on my X2 4400 system I just got running last month. After convergence, all 3 are turning in much better PPD totals than with either Tinker or those crappy small Gro's or PMD's I've been getting so many of lately on my P-M and A64 machines. So far, the P-M laptop is averaging 242 PPD (single channel memory is limiting bandwidth and performance, but still is better than with a Tinker), the P-M desktop is averaging 347 PPD (higher processor speed and DC memory making a big difference) and most surprisingly, the client on the DC A64 rig is averaging 322 PPD without using the SSE2 assembly loops, which is much, much better than what it was averaging doing Tinker or PMD or crappy small Gro work.

So, if you have access to a P4 machine and have a P-M or A64 machines folding, this is a viable way to increase points production without being a WU dumper with the present series of WU's being assigned by the assignment server. Just d/l a QMD with the P4, then transfer the queue.dat file to a client on the machine you want to do the QMD on, along with the work files to the work folder and then restart the client with the -pause flag set so it doesn't automatically d/l a new, non-QMD unit after finishing the WU. It involves a bit of monitoring to do this, but the production increase can make a difference for you, since Stanford still seems to be dragging their heels on getting the version 6 client and still won't let an AMD machine natively get QMD work.

The way I feel about this, if Stanford doesn't like the fact that in order to get decent production from my machines that I have to move work around to utilize the machines for max production, then they should see about fixing the assignment server logic. I'm damn tired of seeing my high performance A64 and P-M machines drawing such an amazing amount of low-PPD wu's when electricity is costing me so damn much lately. I looked on last month's electric bill and they are now charging me an $0.11/KWH surcharge for a fuel adjustment and at this present rate, I'll probably shut down the majority of my farm when temps start warming back up and I'm using the cental air to cool the house. I just won't be able to afford the extra electricty then, especially for a bunch of sorry return WU's. :mad:

[WarriorII]

I hear ya Brother !

Found out one of the local electrical suppliers wants a 20% rate hike here. (AZ)

I hope they choke on it.

Good job on moving your assigned WU's around for best production. :thup:

Ya gotta keep thinking of the Big Picture!

:attn:

[TollhouseFrank]

I hear ya Brother !

Found out one of the local electrical suppliers wants a 20% rate hike here. (AZ)




Dude... be glad ya don't live here then. We get 20-30% increases on our utilities about every 3 months. Then the water/power companies wonder why people disconnect.....

[pscout]

i have been running my laptop that way since i got hooked on qmds ... it only gets about 200 ppd and even with the down time when it waits for me to give it a new one it is still way better.

[ChasR]

Yeah Muddoc, it really is pretty good for the points. I swapped WUs between my FX-53 and a P4C. The fX is getting about 354ppd. That's better than everything but DGromacs and p147x/81.

The logic you'd like to see introduced would likely limit assignment of those precious bonus WUs in a way many wouldn't like. Where would you draw the line? Is a XP Barton @ 2.0 fast enough. What about a P4A @ 1.5 GHz? It has to be random assignment, based on the flags set, regardless of cpu speed, to avoid disenfranchising a large group of folks who don't have the latest greatest hardware.

[muddocktor]

Chas, since I drew a p2352 wu from server 150 using the -advmethods flag today on my DC Opti system I can now talk about what I was referring to about the assignment server being screwed up. That server until really recently was only assigning to the "special" flag and I couldn't talk about it in the general public, but now I can. ;) Looking at the server status page, that server is showing over 3400 WU's available and all WU's on that server are showing a weighting of 200 or higher except 1, which is a 10 weighting. But it's almost impossible to get an assignment from that server. I've tried for hours, starting a client manually by using the -advmethods flag or the special flag and never get assigned to it, but rather to another server that has much lower weighted work. Something is wrong with this assignment server picture to me and my understanding of how the projects are assigned. And that is trying both with bigpackets=yes and bigpackets=no set in the client. I was mostly getting assigned to servers 133, 136, 149 and another I don't remember. Now, I wasn't letting the client download from those servers when getting assigned to them; I just killed the client instead (checking to see how hard it was to get assigned to server 150). The reason I like server 150 is that it has those humongous DGro's on it and my DC procs do well with them, plus it helps cut down on results transfer time on my sat connection, which is real upload bandwidth limited and can be a real bottleneck with the low point Gro work (9 machines on a 56K uplink sucks with those 40 and 50 point wu's).

I finally got lucky today and got the 2352 assigned to 1 client and the other client on my Opti system has a 600 pointer, so that machine is happy and busy, but I don't have the time to keep banging away trying to get decent work for the rest of my machines. Hence, trying the QMD's. I can just d/l a bunch of them every morning and then put them to the machines that will need a wu during the day this way.

[AlabamaCajun]

I'm with you on this. I makes no sence why heavy duty rigs don't get these heavy WUs. There's plenty of fruit on those servers to pick but yet we get 40 point sTinkers on a Opti/P4(8xx). It looks like a load balancing act to keep work flowing with some priority settings to push on some servers more than others. What effect does a large number of machines out there with Timeless Set have on the overall assignments.
I would say all heacy units go to Dual Core, 64 bit and systems above 3G.
Middleweight units to the AMD 754s and XPs above 1.8G, P4 Norths below 3G and Cellys above 2G.
All the low stuff get the Timeless.
I can see people setting Timeless on middle machines that are crappy or for laptops and dialups. I'm sure they could adjust the granularity on that a little.

[Joe Camel]

[comic relief]
just think how MAD some people will be after they build (buy) fast (expensive) P4 farm layers...when QMDs are gone!
[/comic relief]

a LOT of this is due to the existence of QMDs; without them, the "playing field" will be much more level...there would still be "crappy" WU's out there, its just that the P4's would have a chance of getting them TOO.

ehh, FOLDing is just like life, sometimes it aint fair and you just got to suck it up... learn to adapt (and/or take advantage) to the way things are.

:beer::beer:we'll get through this my friends :beer::beer:

PS i assume we ALL agree that we all know this is for science and points are totally meaningless in the big picture. its just...more "fun" this way? :shrug: :p

EDIT:
damn, i forgot to say:

great work muddocktor!!

although this must take quite a bit of time, the payoff seems quite nice.
wish you could have the client draw new WU's from a file (that you would need to fill/refill) but still send finished ones in automatically...sort of how they were talking in the lappy-FOLDing in Iraq thread

PS my buddy is in Baghdad right NOW (11.5 more months) he's basically the Army's network admin...but i cant get him to BORG (yet)

[Rpkole]

I decided to do a little experimentation with QMD work then transferred them over onto my P-M laptop. So far, the P-M laptop is averaging 242 PPD (single channel memory is limiting bandwidth and performance, but still is better than with a Tinker)

:santa: thank you for the Idea now my laptop is putting out 240ppd and a strick diet of QMD's and i found that if i use the -oneunit Flag it will upload the WU then shut down the core at which point i can give it a new QMD

[muddocktor]

Actually it won't take up that much time, Joe. Since the QMD WU's are generally so big (I've only gotten a handful of the 125 point variety), you are only looking at tranferring a new WU to a client about every 1 1/2-2 days to the machines that can't draw them natively. If Stanford will get off their behinds and roll out the next version client, the problem with P-M machines drawing a QMD will be fixed and maybe they will allow some kind of flag like -qmd or something so that you can opt in on QMD work with an AMD machine too. I put in a request for something like that at the community forums at least. We will see (maybe :rolleyes: ) whether Stanford will answer back about letting AMD procs draw QMD work with the next version client.

[ihrsetrdr]

It's a strange coalition we have, scientists teamed up with thousands of people such as ourselves...enthusiasts<?> The scientists get their kicks in pursuit of the goals of the project. WE get our kicks by being able to contribute to the pursuit of such a beneficial cause. OUR reward? Points. Points satisfy our primal needs(**cough** unless there's not plenty of them!) of competition, recognition and conquest. Several months ago, when recruitment and revitalization of interest in folding was a common topic on forums with folding teams, I PM'd Vijay on the Folding-community.org forum and asked him(naively, on my part) if he had any thoughts on effective ways to recruit/retain team members. Here's a paraphrased excerpt of his response..."People come and people go, yet the project as a whole has been growing steadily. I think it's good for people to try out other d.c. projects and will hopefully come back to us" Ain't that a scientist for ya? ;)

Heavy hitters like OC-X on 2CPU.com Folding@Home team have managed to "nudge" Stanford in the right direction. But, until AMD gives permission for Stanford to use their code library, there's not alot of hope for natively feeding our Opty's the WU's they deserve.

[ChasR]

@muddoc
The assignment order is what makes it so hard to get the Dgromacs. .150 is 26 out of 28 active servers.

@ac
To ensure that our high power rigs get the choice assignment would require that lass than high power rigs not get them. As I said before, where are you going to draw the line? I'd be pi**ed as all get out if it were drawn where you suggest and exclude my fleet of XPs from the highest producing WUs. Plenty of others would be too. Which of us would be willing to drop all the flags and bigpackets on our middle weight machines so they would no longer get bonus Work units so as to imporve the odds of our fastest rigs getting them? Not many I'd bet.

[LandShark]

[comic relief]
just think how MAD some people will be after they build (buy) fast (expensive) P4 farm layers...when QMDs are gone!
[/comic relief]

:mad: :bang head http://landshark.damnserver.com/icon/rocketQ3.gif

[ChasR]

For jollies, I transfered a QMD to my daily work machine, sig rig #3. Preconvergence it was almost unsuable. The performance hit made it act like a 400 MHz P2. After convergence it was better but noticeably laggy. I'll see what it does when idle tonight. EMIII is currently reporting 251 ppd, 100 ppd less than an A64 at the same speed will produce.

[TollhouseFrank]

I've done the same with my A64 3200+.

I dont' know when convergance is... but around 6% done, this thing became bloody unusable for about 15 minutes... and now it is perfectly fine.

[gulp35]

You do not need to "sneaker-net" them around...a much easier way is shown right on the FAH forum...http://forum.folding-community.org/viewtopic.php?t=14118

We tried it on our dual opteron...works like a charm. We did it to both and we ended up at about 240ppd per CPU.

We got about 315ppd running one QMDs , with a a crap WU on the other CPU.

For us, its one QMD to a customer! (we are running single channel memory, NUMA enabled, 2Ghz per processor).

[TollhouseFrank]

yah... that's the method i used. I'm gettin' 295 or so PPD on this QMD's i've got.

[the garynator]

i think adding the -qmd flag might be a good idea, but it needs more control so that point hungry folders don't abuse it. so the line is going to have to be drawn somewhere... i'd say clock speed(different for each type of cpu, such as a64, p4, etc) and a ram cuttoff... also possibly make them worth, i know i'm going to get lynched for this, but make them worth less so people don't get as ****ed off that they are unable to get them. There's always going to be those people that are right below the cuttoff line, but it'd be better for stanford if they put qmd's only on the top tier of systems vs. any p4's(where half or more of those are slow as dirt). I dk, just my thoughts.

[TollhouseFrank]

good thoughts... and i'm sure they may include a -qmd flag in thier next generation of the software.

[Max0r]

Wow, just when I thought I couldn't optimize things further... mauhahahah

[TollhouseFrank]

yup. I have a copy of the 'optimized' client and core... but that is for myself... it IS against the EULA and if someone else wants to break it, i'll let it be up to them... Besides... I'm sure that exploit will be patched in the next version.

[ChasR]

If violation of the EULA becomes commonplace, I expect the response from Stanford will be to update the core regularly. Folks just might get tired of patching it.

[LandShark]

If violation of the EULA becomes commonplace, I expect the response from Stanford will be to update the core regularly. Folks just might get tired of patching it.
I agree and that's the most easiest way for Stanford to stop patched client. or they might eventually built in some sort of validating method in it to be use the result is from a non patched client or no point will be issue. patched client was no news from seti, and berekely did clear things up quite good included delete the user account and/or reduce the point on patched client account.

I would against of using a patch/mod'ed client that isn't support/allow by Stanford. unless it's like seti/boinc that berkeley allow some optimized client for certain type of cpu.

just my 0.02.

[TollhouseFrank]

honestly... that is how this is likely gonna end up Landshark. If this becomes a problem, then Stanford will jsut update the client, and if you are right, add in a validation protocol.

[Max0r]

I'm sure as long as they don't have any potential legal troubles and the results they're getting aren't screwed up it won't really be an issue.

[ChasR]

THere are really two issues involved hacking FAH to get QMDs. One is hacking the client so AMDs will get QMDs. I don't think Stanford views that as serious. Distribution of a hacked client would be viewed as a very serious, actionable violation of the EULA. Hacking the core so that AMD cpus use the Intel library in violation of the Intel ULA would be viewed as serious and distribution of a hacked core as very serious and actionable.

[muddocktor]

Heh, I clicked the link that gulp35 posted there and get directed to a screen that says "Sorry, but only users granted special access can read topics in this forum.", so I can't read what he was talking about. Since I am on the beta testing team already, they must have moved the thread to a place I can't see, unfortunately. What are they talking about and how do you hack the client? Please PM me the way to do this if you don't want to publish this methods in the open. Otherwise, I will be shutting down most of my farm next time I go to work offshore, since I'll be gone for 2 weeks then and won't be able to manually transfer the work.

About the slowdown speed-wise at preconvergence, I haven't had any problems with my main rig, but it has 2 GB of ram too. The others are dedicated folding rigs, so no problems either.

[LandShark]

Heh, I clicked the link that gulp35 posted there and get directed to a screen that says "Sorry, but only users granted special access can read topics in this forum.", so I can't read what he was talking about.
that is a very good indication that Stanford doesn't like the idea of hacking their client/core!!