Folding Forum

csvanefalk wrote:Are you sure your client has your username and passkey properly set? If yes, what are your CPU and memory specs?

Yeah, my username and passkey are set. My CPU is an i7 2760QM @ 2.4 GHz (3.5 GHz turbo) with 4 physical cores (8 threads) with 6M cache. My memory is 16 GB of DDR3 RAM @ 1600 MHz. I also have an Nvidia GeForce GTX 675M GPU which I'll use when the Linux GPU client comes out.

bruce wrote:SMP performance generally runs as slow as the slowest thread. Depending on the type of game(s) you play, you may need to Pause/Finish FAH before you start the game -- and even if you accept the performance of your game with FAH running, the FAH performance will suffer quite a bit. There may also be a case for running FAH with a reduced number of cores with your game if you can find the proper balance to maximize performance of both but you'll have to experiment with that.

The only games I really play are Touhou Project and Visual Novels (VNs use very little CPU). Other than that the only things that take up the most CPU time are (currently) KTorrent at 7% and Xorg at 1%. Overall FAH is taking up about 90% of the CPU.

Right now the only applications that are open are Chromium, KTorrent, and FAH and I'm getting 3.5k PPD.

do you run 24/7? or even for 12-15hrs a day?

That 7% Ktorrent will be a killer for SMP folding. that 7% x8 cores means that it will be using 56% of a single core. Another way of analyzing it is that if folding is at 90% then 10% x8 is 80% (of one core) is non-folding activity.

The SMP client runs one thread per core. It is designed so that if some other application needs to use the CPU it suspends that thread. Unfortunately, It is also highly synchronized so that if one thread is suspended, the other cores just sit in a loop waiting for it to return. If you use 80% of a single core then there is only 20% left for SMP.

I would strongly suggest that you drop the number of threads the SMP core uses from SMP:8 to SMP:6 (SMP:7 will cause some WU's to fail) leaving enough CPU for the other applications. You should find that the SMP productivity should skyrocket.

patonb wrote:do you run 24/7? or even for 12-15hrs a day?

24/7

P5-133XL wrote:That 7% Ktorrent will be a killer for SMP folding. that 7% x8 cores means that it will be using 56% of a single core. Another way of analyzing it is that if folding is at 90% then 10% x8 is 80% (of one core) is non-folding activity.

The SMP client runs one thread per core. It is designed so that if some other application needs to use the CPU it suspends that thread. Unfortunately, It is also highly synchronized so that if one thread is suspended, the other cores just sit in a loop waiting for it to return. If you use 80% of a single core then there is only 20% left for SMP.

Then what's all the extra CPU usage and heat from? Overhead?

P5-133XL wrote:I would strongly suggest that you drop the number of threads the SMP core uses from SMP:8 to SMP:6 (SMP:7 will cause some WU's to fail) leaving enough CPU for the other applications. You should find that the SMP productivity should skyrocket.

I'm doing that. How much of a speed-up should be expected?

With the extra 2 cores that are no longer folding, would it be best to give one (or both) of them uniprocessor cores so that even if they are effected by KTorrent, it'll still add some PPD? I'd like to put some of that extra idle time to work if I can.

Best to leave them not folding. The virtual cores provided by HT do not contribute as much processing. So adding an uniprocessor slot will cause the same interference to the SMP slot as Ktorrent.

Stonecold wrote: Then what's all the extra CPU usage and heat from? Overhead?

There is some Windows overhead and it is hard to measure in that the task manager does not report it. That which is reported (like Ktorrent) you can see. If you are talking about folding then even if the folding threads are just checking to see if the suspended core is up and running yet they are still doing something and that generates the task manager usage numbers and lots of heat. It is not as if your machine can power-down because it is not doing anything: There is CPU activity.

Stonecold wrote:I'm doing that. How much of a speed-up should be expected?

You'll see for yourself very shortly but I would not be surprised at a 5x or even more increase in PPD. Don't forget with the SMP client the vast majority of points come from the quick return bonus.

P5-133XL wrote:There is some Windows overhead and it is hard to measure in that the task manager does not report it. That which is reported (like Ktorrent) you can see. If you are talking about folding then even if the folding threads are just checking to see if the suspended core is up and running yet they are still doing something and that generates the task manager usage numbers and lots of heat. It is not as if your machine can power-down because it is not doing anything: There is CPU activity.

I'm not using Windows, I'm on Kubuntu 12.10. Does the "top" command act the same way as Windows' Task Manager?

Sorry, I know little about Linux ...

Stonecold wrote:

P5-133XL wrote:I'm not using Windows, I'm on Kubuntu 12.10. Does the "top" command act the same way as Windows' Task Manager?

Yes, the top command usually acts just like Task Manager with the CPU% filter applied. It will bubble "to the top" of the list your top resource / cpu usage threads. 1% for XOrg is around normal for reporting for Linux. Animated Icons and local search indexing ( beagle comes to mind ) can slow SMP processing as they consistantly interrupt the threads. Not sure what Kubuntu uses for search ( hard drive ) indexing.

PinHead wrote:Yes, the top command usually acts just like Task Manager with the CPU% filter applied. It will bubble "to the top" of the list your top resource / cpu usage threads. 1% for XOrg is around normal for reporting for Linux. Animated Icons and local search indexing ( beagle comes to mind ) can slow SMP processing as they consistantly interrupt the threads. Not sure what Kubuntu uses for search ( hard drive ) indexing.

Kubuntu uses Nepomuk to index the hard drive (although I have it set to only index filenames, I don't like the thought of it remembering the contents of files I delete). And what do you mean by "animated icons"?

It's very likely that SMP:6 will be more productive than SMP:8 since you're saying that 20% of the time one SMP thread will be suspended. There's a good chance that 6 cores can get their 75% with very few interruptions and the other two virtual cores can handle the 20% you're spending on other tasks leaving a 5% idle cushion so what would otherwise be interruptions will often find a free CPU.

Top and TaskManager are very similar. The only obvious difference is that Top adds everything up to 100% of a CPU with each thread being 12.5%. Task Manager adds 100% per CPU up to a total of 800% on an 8-way CPU like your i7. It just 100% of two different things and you can divide/multiply by 8 when talking about the other way of reporting it.

I set my client to SMP 6 but I'm not seeing much PPD increase. It's currently at 5k PPD, but that's within the range I normally got with SMP 8. Is there anything I'm missing?

You can expect the numbers to stay in the same general range unless some other task preempts one of FAHs threads much of the time. That needs to be avoided but we can't tell if you have that problem or not.

With nothing except FAH running, on your i7, changes between SMP:4 through SMP:8 will be small but will favor larger numbers (except for odd numbers which should be avoided)
.
With some other task hogging any one or more of the CPUs that FAH is trying to use, performance will pretty much go down to one half. If the interruptions from other processes happen to be equally distributed across all SMP threads, performance won't suffer like that but will so down to about what you'd logically expect. All interruptions carry an overhead so sometimes it cheaper to let some CPU time be wasted/idle rather than trying to keep all CPUs busy, ESPECIALLY CPUs 5 through 8.

Without actual data from your machine, predicting whether the other tasks will distribute their interruptions uniformly or not.

Stonecold wrote:I set my client to SMP 6 but I'm not seeing much PPD increase. It's currently at 5k PPD, but that's within the range I normally got with SMP 8. Is there anything I'm missing?

The only thing that I haven't seen addressed is the fact that you said "Gaming Laptop" and you haven't addressed laptop power saving features and you haven't mentioned the FAH client setting for when you are on battery power. You probably need to look at your power save modes and how you configured the client since you are on a laptop.

PinHead wrote:The only thing that I haven't seen addressed is the fact that you said "Gaming Laptop" and you haven't addressed laptop power saving features and you haven't mentioned the FAH client setting for when you are on battery power. You probably need to look at your power save modes and how you configured the client since you are on a laptop.

All power saving features are off because I never unplug my laptop unless I'm moving it somewhere. I never fold on battery.