Always Getting Crappy SMP WUs on F@H; Could Something Be Wrong?

[LOL_Wut_Axel]

Alright, so the background: I made the desktop in my sig around a month ago, and one of the things I made it for apart for gaming and video encoding/transcoding is folding.

I changed my username, so I had to start all over. For the first 10 SMP WUs I wouldn't get bonus points; only base. And that's fine, since it applies to everyone. After those 10 WUs I started getting good ones: many would get me over 5-6K points. At first I was folding for Tom's Hardware and was able to amass 112K points in around two weeks, but since I barely post there I decided to change to here, Anandtech (198). Around the first five WUs I got while folding for this team were much the same as I got earlier, which is to say good. Even with smp4 I was getting 3K+ point WUs.

As of around a week or so, however, for some strange reason all I'm getting nonstop are crappy 8004 project WUs for smp8 and 8001 project WUs for smp4. The base points for these are around 89 points and all I get with the bonus is 750 points each time I complete one, which is crap. Because the WUs are much smaller my PPD estimates have gone down, too. Before I would get 25K+ PPD with smp8 and the larger WUs, while now I'm barely able to break 20K.

Now, could there be a reason for this? I don't really want to do GPU folding anymore. It consumes almost twice the amount of power and normally gives me only 1/2 the PPD, so it's only a quarter as efficient.

 

[Mr. Pedantic]

You're just unlucky?

It happens. All people get a bad WU once in a while, it's not much of an extrapolation to say that it's also possible for someone to get a single run of lots of bad WUs.

 

[LOL_Wut_Axel]

You're just unlucky?

It happens. All people get a bad WU once in a while, it's not much of an extrapolation to say that it's also possible for someone to get a single run of lots of bad WUs.

But getting way over 20 bad WUs? That just sounds like something is wrong to me.

I'm now finishing an 8004 WU. I'll go into Configure and dig around. Before all I did was go into Slots>SMP and set it for 8 threads and all would be fine. Perhaps I should set the -smp8 flag in Expert>Extra core options? It is using 99-100% of my CPU currently, though.

EDIT:

Got yet another 8004 WU. Goddamn it!!!

 

[GLeeM]

A while back they announced they were going to only send bigadv to 16 core or better systems.

Other than that, Folding has always had problems setting WU points.
They only benchmark the first few frames and if the other 95 frames take longer then you get lower ppd.
If, as the WUs get deeper into the cycle, they start taking longer then you get lower ppd.

There are many different proteins they are working on, if they put a high priority on the bad ppd ones then you get mostly those. You can sometimes use different flags to try to get better ones.
Some people used to see bad ppd WUs as a virus and blocked the server.

If you complain about points at the forum, they lock the thread.

Don't let it get to you - this is like a hobby you should enjoy.
You just have to enjoy crunching what they send ... or have fun crunching a BOINC project (or DPAD or SOB).

 

[LOL_Wut_Axel]

A while back they announced they were going to only send bigadv to 16 core or better systems.

Yeah, that's why I decided to do SMP only.

Other than that, Folding has always had problems setting WU points.
They only benchmark the first few frames and if the other 95 frames take longer then you get lower ppd.
If, as the WUs get deeper into the cycle, they start taking longer then you get lower ppd.

In my experience the first frame always gives me an estimate around 5K PPD lower than at the end. After the 5th frame or so it stabilizes.

There are many different proteins they are working on, if they put a high priority on the bad ppd ones then you get mostly those. You can sometimes use different flags to try to get better ones.
Some people used to see bad ppd WUs as a virus and blocked the server.

Which ones? I need to know. These 8004 WUs are driving me mad! I've been folding for almost a week for our team and have yet to break 40K points.

If you complain about points at the forum, they lock the thread.

Where, here? Why would they do that?

Don't let it get to you - this is like a hobby you should enjoy.
You just have to enjoy crunching what they send ... or have fun crunching a BOINC project (or DPAD or SOB).

Well, yeah, but the competition and rising in numbers is fun too! If I always get crappy WUs how am I supposed to better my numbers? And the crappy WUs aren't as good when it comes to research value, either.

^^

 

[Mr. Pedantic]

But getting way over 20 bad WUs? That just sounds like something is wrong to me.

I'm now finishing an 8004 WU. I'll go into Configure and dig around. Before all I did was go into Slots>SMP and set it for 8 threads and all would be fine. Perhaps I should set the -smp8 flag in Expert>Extra core options? It is using 99-100% of my CPU currently, though.

EDIT:

Got yet another 8004 WU. Goddamn it!!!

Why not? It's got to happen to someone, at some point in time.

 

[Sunny129]

If you complain about points at the forum, they lock the thread.

Where, here? Why would they do that?

no, not here - i believe he was referring to the Folding@Home forums.

 

[VirtualLarry]

F@H's point system has always been a bit screwed up. Especially since you mention that CPU folding is 4x better PPD/watt than GPU folding, which doesn't match up to reality. In reality, GPUs have much higher FLOPS/watt. That's why they are building supercomputers with Tesla cards, instead of massive CPUs.

Personally, I'm all about GPU computing. I mostly run PrimeGrid these days, but I am confident that GPUs are more efficient in processing than any current CPUs.

Possibly AVX and AVX2 might help CPUs catch up, but that will take some time.

 

[LOL_Wut_Axel]

F@H's point system has always been a bit screwed up. Especially since you mention that CPU folding is 4x better PPD/watt than GPU folding, which doesn't match up to reality. In reality, GPUs have much higher FLOPS/watt. That's why they are building supercomputers with Tesla cards, instead of massive CPUs.

Personally, I'm all about GPU computing. I mostly run PrimeGrid these days, but I am confident that GPUs are more efficient in processing than any current CPUs.

Possibly AVX and AVX2 might help CPUs catch up, but that will take some time.

That may be true for some other projects, but not Folding@home.

You need a GTX 590 or 2x GTX 560 Ti to match a 2600K at ~4.6GHz in smp8. Those two will also mean a system power consumption that's around 3x higher, plus they cost more. Even taking into account you can put up to three or four of those graphics cards in each motherboard, the cost of three or four 2600K systems is lower and, again, you get much lower power consumption. GPU folding makes the system a lot more sluggish even for very simple tasks, too.

 

[Sunny129]

F@H's point system has always been a bit screwed up. Especially since you mention that CPU folding is 4x better PPD/watt than GPU folding, which doesn't match up to reality. In reality, GPUs have much higher FLOPS/watt. That's why they are building supercomputers with Tesla cards, instead of massive CPUs.

Personally, I'm all about GPU computing. I mostly run PrimeGrid these days, but I am confident that GPUs are more efficient in processing than any current CPUs.

this.

as i've mentioned in previous threads, Folding@Home is the exception to the rule when it comes to CPU PPD vs GPU PPD, and i can only think of 2 possible reasons for this:

1) F@H computing really is more efficient on CPUs than it is on GPUs. of course that would mean that the characteristics of F@H data would have to be VERY different from the data characteristics exhibited by most other DC project data. that is to say, F@H data must not be massively parallel in nature at the instruction level, or else it would never be as efficiently crunched on a CPU as it could be on a GPU...and the current points system shows the opposite trend.

2) F@H data really is massively parallel in nature, and GPUs do crunch the data much more efficiently than CPUs, but the points system must be so fudged that it heavily favors work done by CPUs with a high core/thread count. if this is truly the nature of F@H, then this would mean that some tasks which earn less points may in fact contribute more to the science of F@H than tasks that earn more points.

EDIT - what we're getting at is that while it may be better from a "PPD/watt" point of view, we're not 100% confident that a higher PPD actually guarantees that you've contributed to the science of the project any more than someone w/ a lower PPD. i'm sure on average, a higher PPD does actually mean a higher contribution to the actual science, but not always...

 

[LOL_Wut_Axel]

Why not? It's got to happen to someone, at some point in time.

Well, yeah, but the point is trying to minimize when it happens. You mentioned some flags could be put on the client to help minimize me getting those pesky 8004 WUs. What are they?

 

[Desslok]

Alright, so the background: I made the desktop in my sig around a month ago, and one of the things I made it for apart for gaming and video encoding/transcoding is folding.

I changed my username, so I had to start all over. For the first 10 SMP WUs I wouldn't get bonus points; only base. And that's fine, since it applies to everyone. After those 10 WUs I started getting good ones: many would get me over 5-6K points. At first I was folding for Tom's Hardware and was able to amass 112K points in around two weeks, but since I barely post there I decided to change to here, Anandtech (198). Around the first five WUs I got while folding for this team were much the same as I got earlier, which is to say good. Even with smp4 I was getting 3K+ point WUs.

As of around a week or so, however, for some strange reason all I'm getting nonstop are crappy 8004 project WUs for smp8 and 8001 project WUs for smp4. The base points for these are around 89 points and all I get with the bonus is 750 points each time I complete one, which is crap. Because the WUs are much smaller my PPD estimates have gone down, too. Before I would get 25K+ PPD with smp8 and the larger WUs, while now I'm barely able to break 20K.

Now, could there be a reason for this? I don't really want to do GPU folding anymore. It consumes almost twice the amount of power and normally gives me only 1/2 the PPD, so it's only a quarter as efficient.

What??? I thought folding via GPU was more efficient than via CPU???

 

[LOL_Wut_Axel]

What??? I thought folding via GPU was more efficient than via CPU???

GPU folding has always been anywhere from 1/2 to 1/4 as efficient. F@h is better optimized for CPUs than GPUs, but ironically floating point performance matters more than integer. And, as we all know, GPUs excel at floating point. So yes, F@h is kinda weird when it comes to optimization.

Speaking of which, I wonder how the new Opteron 6272 would do on -bigadv under the new rules. You need 16 threads now as a minimum to run -bigadv, but Folding@home shouldn't care that the Opteron has 16 integer cores and 8 floating point units/cores. Perhaps it wouldn't be enough to finish in the set deadline? I haven't seen any info on the web regarding that.

 

[Mr. Pedantic]

What??? I thought folding via GPU was more efficient than via CPU???

It depends what you mean. FLOPS/W, GPU folding is much, much more efficient. Points/W, though, goes to CPU for the time being. I don't really know why this is; maybe the CPU projects are more urgent, or the flexibility/scale of CPU folding gives it something that is worth more?

 

[blckgrffn]

If they awarded points based on actual work done, then all of the CPU projects would likely see a huge exodus. They need people to fold based on projects they have already started.

You can see this in PrimeGrid or other GPU BOINC projects where the point totals make CPU crunching a worthless endeavor if points are why you crunch.

My 5870 got 1M in MilkyWay in one week. With 12 CPUs, I hope to get 1M in WCG in one year. Does that mean I should stop doing WCG since its a waste of power? No, that seems silly to me.

In that regard I think that F@H does a good job balancing SMP and GPU point totals. -bigadv really knocked things out of whack IMHO. It was a great way to get to 12M in F@H in two months though. I hope to break 20M by the end of the year, although that is looking a little doubtful with my current BOINC focus. I want to get 1M in WCG, Malaria Control and Correlizer this year and the GPUs I have left folding are worth "only" 13k per day.

 

[LOL_Wut_Axel]

If they awarded points based on actual work done, then all of the CPU projects would likely see a huge exodus. They need people to fold based on projects they have already started.

You can see this in PrimeGrid or other GPU BOINC projects where the point totals make CPU crunching a worthless endeavor if points are why you crunch.

My 5870 got 1M in MilkyWay in one week. With 12 CPUs, I hope to get 1M in WCG in one year. Does that mean I should stop doing WCG since its a waste of power? No, that seems silly to me.

In that regard I think that F@H does a good job balancing SMP and GPU point totals. -bigadv really knocked things out of whack IMHO. It was a great way to get to 12M in F@H in two months though. I hope to break 20M by the end of the year, although that is looking a little doubtful with my current BOINC focus. I want to get 1M in WCG, Malaria Control and Correlizer this year and the GPUs I have left folding are worth "only" 13k per day.

If I keep getting all these crappy WUs I'll be at 1M F@h points in half a year. :$

Do you know about those flags Mr. Pedantic mentioned? I've only broken 40K in a week. :|

 

[blckgrffn]

If I keep getting all these crappy WUs I'll be at 1M F@h points in half a year. :$

Do you know about those flags Mr. Pedantic mentioned? I've only broken 40K in a week. :|

Hmm... I've had my share of crappy WUs... are you at the command line or using the beta client? You could try specifying "large" WU's during setup on the command line, you should be able to do that somehow in the beta client too. These WUs are simply larger when it comes to downloading and uploading them and I think they are are skewed for more points. I would give that a go. Are you in Windows or linux?

 

[LOL_Wut_Axel]

Hmm... I've had my share of crappy WUs... are you at the command line or using the beta client? You could try specifying "large" WU's during setup on the command line, you should be able to do that somehow in the beta client too. These WUs are simply larger when it comes to downloading and uploading them and I think they are are skewed for more points. I would give that a go. Are you in Windows or linux?

Using Beta client... version 7.1.43, in Windows 7.

If there's an option for specifying something regarding WUs, it's probably in Configure>Expert and either on Extra client options or Extra core options. But IDK what to put there. :/

 

[Mr. Pedantic]

Well, yeah, but the point is trying to minimize when it happens. You mentioned some flags could be put on the client to help minimize me getting those pesky 8004 WUs. What are they?

Me? Really? I think you give me too much credit

But you could try -advmethods and -forceasm, they might be what you're looking for?

 

[LOL_Wut_Axel]

Me? Really? I think you give me too much credit

But you could try -advmethods and -forceasm, they might be what you're looking for?

Thanks a lot. Looked both terms up and found this article on it. I'll try it with smp8 once the 8001 project I'm currently on finishes and we'll see how it goes. :thumbsup:

EDIT:

Am I doing this right? Here's what the log says once I put the options in:

04:56:07:Saving configuration to config.xml
04:56:07:<config>
04:56:07:  <!-- FahCore Control -->
04:56:07:  <core-priority v='low'/>
04:56:07:
04:56:07:  <!-- Folding Slot Configuration -->
04:56:07:  <extra-core-args v='-smp8 -advmethods -forceasm'/>
04:56:07:
04:56:07:  <!-- Network -->
04:56:07:  <proxy v=':8080'/>
04:56:07:
04:56:07:  <!-- User Information -->
04:56:07:  <passkey v='********************************'/>
04:56:07:  <team v='198'/>
04:56:07:  <user v='LOL_Wut_Axel'/>
04:56:07:
04:56:07:  <!-- Folding Slots -->
04:56:07:  <slot id='0' type='SMP'>
04:56:07:    <cpus v='8'/>
04:56:07:  </slot>
04:56:07:</config>

 

[theAnimal]

Don't worry about a relatively small ppd difference like that. My bigadv WU vary by about 80-90k ppd.

 

[blckgrffn]

Ah, just like 20k ppd to 25k ppd? I had missed that - yes, that variation is "normal" and seems to be exacerbated in Windows when you use the machine in way that Linux handles better. It seems like Windows is really quick to unload F@H threads. I think there is a way to have it run at a slightly higher priority? That might be a good idea...

 

[salvorhardin]

During the race in december I found this page useful in configuring the beta client. When you start the client (running under advanced) go to configure, choose the slots tab, click on edit for the smp slot, scroll to the bottom and click on add under extra slot options. In the dialog window that pops up under name put client-type under value you can put advanced or bigadv. then ok it and add another slot option, this time using max-packet-size for the name and then either small, normal and big. this last option lets you choose wu size (~5, ~10 and ~500).

 

[blckgrffn]

During the race in december I found this page useful in configuring the beta client. When you start the client (running under advanced) go to configure, choose the slots tab, click on edit for the smp slot, scroll to the bottom and click on add under extra slot options. In the dialog window that pops up under name put client-type under value you can put advanced or bigadv. then ok it and add another slot option, this time using max-packet-size for the name and then either small, normal and big. this last option lets you choose wu size (~5, ~10 and ~500).

Nice, that is helpful!

 

[GLeeM]

-forceasm I think is obsolete. It was needed because if the computer got turned off hard or crashed (think pre WinXP days!) the client, thinking it was the cause of a crash, would normally start without using the advanced assembly code. This flag forced assembly code on all restarts.

CPU is more flexible, it can do many types of things. GPU can do one sort of thing very fast. They find that many fast CPU cores with lots of ram can return certain types of computations quickly. Fast return is KEY to their research because the next in line WU is created from the returned WU. If they DON'T send that type of WU to a computer that will take 10% longer or more then they can get through the research before the semester (or year) ends. If it doesn't finish before they graduate, it doesn't do them much good!
They give higher points to what they need more of, trying to get people to buy that hardware. After they get too many of that hardware then they give higher points to something else. Which is why experienced crunchers are careful where they invest.

GPU folding has always been anywhere from 1/2 to 1/4 as efficient. F@h is better optimized for CPUs than GPUs,

Not always, before SMP and -bigadv. GPU ruled - by alot!
If you want to get a taste of what it was like, try the uniprocessor client on a Pentium 3 where we used to get 25 ppd, or if you were lucky you could get a WU that used SSE and get 50 ppd

-advmethods flag may give different WUs or no flag at all, come to think of it that may only work with uni client?

You just have to enjoy crunching what they send ... or have fun crunching a BOINC project (or DPAD or SOB).