Folding @ Home: Stats

[trevinom]

I haven't had time to scan the other forums to see if there are any like-minded folks that can team up with us, but will try this weekend.

 

[Insidious]

HiYa folks.

I see I'm not the only one feeling a bit overlooked with this QMD thing. First of all, let me say I wasn't trying to incite a riot, but it does make me feel better to know I am not alone in this frustration.

I really like being part of this team and have come to appreciate the people who frequent the F@H thread and don't like the idea of leaving very much. So I'm stuck between two evils... some dissatisfaction with F@H and the prospect of bailing on my teammates.

For right now, I'd like to suggest we put the thoughts of leaving on hold for a little while and give the Pande group time to show us more of their plans for future work units. If they really do get the A64 QMDs up and running that would ease most of my gripes about preferential treatment with CPU branding as an advantage/disadvantage. with only the point assignments left to deal with maybe we can make some noise and get them to reconsider the scoring on these or future units.

I plan to put on my flameproof suit and post sometime this weekend at the stanford forum asking for the scoring to be applied in a way that is fair to all (as much as possible). I would suggest that anyone else feeling slighted do the same.

I know from past experiences on other projects that threats to boycott, etc. will not move this mountain. They have plenty of contributors to ignore a disgruntled few that quit. So lets try catching these flys with honey and keep the best attitudes we can while this settles out.

I'm assuming that the 4 or 5 of us that have posted about this issue are in the minority within team Anandtech and I want to respect our silent majority and not cause them embarrassment. If it turns out that we need to make a change we can do it in a way that impacts those that stay as little as possible, but hopefully we are not to that point just yet.

-Sid

 

[Insidious]

I found this at the F@H forum (thought it was illuminating)

youngmin wrote:I own AMD64 personally too. Too bad that AMD and Intel are playing dirty games in my perspective (I stress in "MY" perspective, and thisi is not an official view from F@H). QMD code uses math intensive routines, and instead of coding ourselves, we simply linked with already available highly optimized (I am pretty sure it is done on assembly level) libraries. And what's dirty? From Intel Program, AMD is a "Generic CPU". (You can see this if you run core directly by double clicking it. I desinged so that it pops out a message box with the detected cpu species) I know SSE2/AMD64 improves performance by like 30% when I use AMD library. For AMD library, the license agreement requires me not to run it on Intel cpu. (even though it can use Intel SSE2, as far as I know) And, for Windows core, I have to use non-intel compiler, which I don't really want to do. But, still, AMD64-939 (either winchester or newcastle) has a very good memory bandwidth, and my 2.0GHz 3200+ behaves exactly as if it is 2.0GHz P4 (projected from performance of P4 1.5GHz) even without SSE2.

Anyways, we are considering two possibilities:
1) Make a separate core (AMD64 core): sounds easy but has some license issue with AMD library (distribution of executibles is prohibited when I last checked the agreement)
2) Wait until Intel enables their library to run SSE2 on AMD64

I asked Intel about 2) the other day, and they said the newest library should use SSE/SSE2 on AMD chips, but certainly Core 1.03 showed "Default Processor" on my personal machine.

If any of you have any suggestion about this, it will be really welcomed.

Thanks!

-Sid

 

[JeffCos]

Translation for those of us that don't speak broken russian nerd.

 

[Insidious]

Originally posted by: JeffCos
Translation for those of us that don't speak broken russian nerd.

ROFL :laugh:

/me slips into crooked horn-rims and loads pocket protector

I think he is saying when they wrote the software for the QMD core, they didn't take time to write their own library functions (software modules that perform calculations). Instead, they used library functions published by Intel. Now these Intel libraries don't recognize the A64 as a SSE2 capable CPU and don't use the instruction sets that would take advantage of a 30% increase in performance... so A64 processors would take too long to process a QMD work unit and are excluded those assignments.

he goes on to say that AMD also publishes library functions that work great for the QMD work units, but AMD has a licensing requirement that their functions are not to be run on an Intel processor. So they didn't want to use those and exclude all the P4 machines (which would be in MUCH greater numbers than A64s.)

For the future, they want to use both A64s and Intel machines for these units. Intel says their next version of library functions will recognize A64s as SSE2 capable and that would solve their problem. Also, they have the option of writing a separate core for A64 machines. but there are licensing issues imposed by AMD that make that difficult or impossible.

So obviously, they are hoping the newest Intel libraries will be the fix. (I don't hear any tone of a fast solution here..... probably a long ways out if we look at the past speed of development in F@H)

/me hides dorky glasses and pocket protector under the bushes

Some of the other things I found when I browsed their forum:

QMD work units are expected to be a very small percentage of all WUs released (~1%)

Trying to run more than one QMD on a single core (HT) will lead to one of them craching. Anyone running more than one instance should only put the -advmethods flag on one of them.

Some forum members were complaining about the resources necessary for QMDs and were more interested in how to make sure they didn't get them.

I couldn't find any mention (complaint or compliment) about point values, but it was clear than posts about this issue would be little more than an annoyance to them. They simply don't care about points.... period!

So it seemed to me that the AMD vs. INTEL issue isn't really one raised by F@H, but one created by AMD and INTEL themselves in their quest for a leg up at the expense of the other.

I'm still kind of pissed and wish things were more fair, but we get what we get and I don't think we are going to sway the course of this project. Let's hope they get the AMDs folding in this decade.

-Sid

 

[JeffCos]

According to Gleem if you have a QMD and you run anything else, even a small WU in another instance, you'll be losing points.But this is the only project I participate in so what's the deal with other DC projects and AMD or Intel Processors. Are they treated equally there?

 

[Insidious]

Originally posted by: JeffCos
According to Gleem if you have a QMD and you run anything else, even a small WU in another instance, you'll be losing points.But this is the only project I participate in so what's the deal with other DC projects and AMD or Intel Processors. Are they treated equally there?

Gleem's info agrees with the posts I was seeing at F@H. Highest PPD comes from a single QMD for each core (IIRC)

Other projects.... From reading posts over the years, I think there has always been some difference in effeciency between Intel and AMD processors. Some projects crunch better on AMDs and some projects crunch better on Intel. But this is the only project I know of that will not design their work units to work on either kind of CPU and most certainly the only one I know of that gives highly accelerated points to a given CPU by brand... on purpose.

-Sid

 

[GLeeM]

They simply don't care about points.... period!

This is true. But they do try as much as they can, without cutting into research time, to get as fair a system of WU rating points. They don't like bad publicity - it affects how the DC community thinks of F@H. They do need us to play their game 🙂 Also, I think they have a lose-lose situation with trying to please everyone, it can't be done.

@Insidious
With my info thread on QMD, I was only trying to educate those that might want to know, how to help the TeAm by optimizing their equipment (=donation) as much as possible. (Not that anyone read it anyway 🙁 )

You have to remember that QMD gets the SSE2 boost AND the BigWU double bonus. I don't think anyone would run them otherwise.

@JeffCos
I haven't tried very many different WUs with QMD on HT. There might be a protein WU that crunches well with QMD.

From what I have seen though, WUs or other projects (ex. Predictor) that use the same parts of the CPU as QMD don't do well and slow down the QMD alot. E@H is one that does not slow the QMD as much.
EDIT: The mention of E@H is just informational, (to show that different projects use different resources, from my experience), not to try to get anyone to run E@H. My loyalty is to the TeAm and then F@H. It used to be the other way around, until some problems I had over at the F@H forum. 😕 🙁