(Yet Another) Which Client to Use Thread?

[Zithras]

I'm, into the software installation stage (next comes customization and overclocking) of the new computer. I'd like to run Folding@Home on it (like on all my other computers) and would like advice on the best clients to run.

I don't want to install any of them as a system service, since I want to manually start them when not using the computer otherwise, and stop them as needed (They'll probably be running most of the day, every day).

I'd also like 1 (and only 1) 'pretty' GUI-type taskbar client, so I can stare at it/screensaver it occasionally (although this is optional if the best clients to run aren't available with this). The other running clients will probably be running in open command windows.

My computer has the following cpu/gpus:
980X processor (6 cores with hyperthreading, so 12 virtual cores)
3 x ATI 5870 cards linked in CrossfireX

What clients do I want?

I assume I'll need 1 advanced CPU client running configured for 6 cores, and some number of gpu clients (not sure which ones/how many/if Crossifre or multiple cards make a difference)...

Thanks!
Zithras

[toTOW]

Did you read this : viewtopic.php?f=60&t=15446 ?

[7im]

Welcome to the Zocalo, Zithras.

This is the quick answer, but you should do more reading (see below)...

For best production, you will want to install the SMP console client, and 3 GPU console clients. No tray icons and no screen saver. We don't have a screen saver, and the viewer that comes with the GPU client doesn't work well. Set the GPU clients to run at a higher priority (inside the client setup, NOT in windows). Then search the forum for info on how to configure environment variables to make your ATI cards fold faster.


Read this? Client Install Guides

In particular, this section? Which client is right for me? Then read the SMP and GPU install guides.

Also check out this section on the client download page What are the different types of clients? and then get the clients you want.

[patonb]

Why no systray gpu? Honestly they are the best option for anyone who uses their machine.
Easily paused when needed for gaming, or other gpu intense apps.

Everyone hates on the systray, but why? Its the same friggen client!

I wish i could get my smp to be systray...

[7im]

3 GPUs, that's why. Would you have someone new to the project try to configure 3 systray clients, or mix systray and console clients? Not me.

3, er 4 console clients was the easiest solution, IMO.

But 2 console and 1 systray client for the GPUs, plus one console for SMP is just fine.

Who knows, maybe he only wants to fire up 1 GPU client? 1 Systray for that, plus one console for SMP would work well also.

P.S. I like the systray client, but it is different. Only the fahcores are the same, and they do all the work.

[Zithras]

Hmm I should have clarified a bit, but figured I'd just get a quick 1 line confirmation answer to my question...yay for overly helpful people! I'm not new to the project, and I've configured console and systray clients to work together before.

However, there's several reasons why I wanted to ask about my particular setup:
1) I wasn't sure how multiple graphics cards were currently being treated - never had to configure a multi-gpu system for F@H before (and likewise didn't know if Crossfire would make them be seen as a single card - apparently not)
2) wasn't sure if 6-core chips were supported by the cpu client or if I'd have to use 2 of them (I've only run the advanced cpu client on 2 and 4 core chips, usually with hyperthreading disabled) For some reason I thought the advanced cpu client only supported (up to) 4 threads/cores. I couldn't find any documentation that said this, and dunno where I heard it, so thought I'd better check.
3) wasn't sure if ATI cards had problems (NVidia seems better supported by the project, and all cards I've run before have been nVidia)
etc.

It sounds like my original intuition was almost correct - 1 CPU client will work, but I'll have to use 3 GPU clients (not 1, like I originally thought)

Could someone with a newer ATI card chime in on the stability of the current system tray client? If it's pretty stable, I'll run 1 Systray GPU client and 1 console CPU and 2 console GPU clients.
Also, just to confirm, with hyperthreading enabled, the SMP client should be running 12 threads, not 6 right? (I know if I was just running the classic client, I'd want 6 copies, but the SMP works on a single unit at a time, so I want the virtual cores running their own threads?)

However, on every system with an nVidia card I've installed the systray GPU client on, it crashes occasionally, especially if I actually open up the graphical representation mode/screensaver bit.

Thanks for clarifying,
Zithras

[7im]

Systray client is very stable, as is the console client.

As I said above, the viewer for the Systray client is not that great. Viewer is actually a separate program used by the client. It in no way should reflect upon the actual stability of the client, although it can be misinterpreted that way.

The SMP client (advanced CPU client?) can go above 32 cores, so you're safe for now. Just use the -smp switch with no number, and the client will use all of the available cores, including the HT threads.

[PantherX]

Welcome to the F@H Forum Zithras,

1) XFire/SLI is too slow for F@H. You can set-up multiple GPUs by using -gpu 0 -gpu 1 -gpu 3 flags. (more details in my GPU Guide; link in sig)

2) 1 instance of SMP2 v6.30 can support more Cores/Threads then your current set-up (IIRC, someone tested it on a 32 core system and it worked fine) just run with -smp 10 so you can 2 threads free to feed your ATI GPUs. You can enable HT if you run a single instance of SMP2 Client.

3) ATI GPUs fold slower than Nvidia GPUs (their counter-part) but if you use Environment Variables, you can get a decent PPD

Using the GPU3 BETA Client, you don't need the -forcegpu flags so the setup is marginally easier.

For your Nvidia GPU, its the viewer as it is known to be buggy and may cause unexpected behavior of the F@H Client and/or the system. Avoid it and use FahMon or HFM.NET. If you use SMP2 WUs, HFM.NET supports the bonus points calculation while FahMon doesn't.

[Zithras]

o_O I never even thought about leaving a core (2 threads) on the CPU free for the GPUs - seems obvious in retrospect - thanks for pointing that out! I'll also poke around for the environment variable guide

Edit: Is this the environment variable guide: http://en.fah-addict.net/articles/artic ... iables.php
It's not hosted on this forum, but I found several forum threads linking to it

[PantherX]

Yep, that is the link. FAH-Addict is officially recognized by PG (Link). Also my suggestion about 2 threads for 3 ATI GPU is based on reading posts on this Forum. It is best if you experiment with your setup and post your results here. Sometimes the CPU Usage can be very low so a free thread isn't really needed while in other cases, you may need more than 1 thread to feed the GPUs. Remember that under certain situations, it is better NOT to use a large prime number for -smp N example:
-smp 12 will be stable on your system
-smp 11 may be unstable on your system
-smp 10 will be stable on your system
It will need some tweaking to ensure that you have optimized your F@H contribution.

[7im]

I'm not sure what he means by "officially" but that site was introduced in Vijay's blog. It does have some good info, but instead of adding that info to existing resources like the FAH WIKI or the FAQs, or even this forum, it's just one more place we have to go dig for answers.

And I'll have to disagree with PantherX. When you set the GPU clients to run at the higher priority, you DON'T need to reserve 2 whole processor cores to drive the GPU clients, especially after setting the EnVars.

The one part I do agree with is that experimenting with your setup will allow you to fine tune your settings to match your hardware, and computing environment. For example, if the system is dedicated to folding, you might run smp on all 12 cores. However, if you plan to use the PC, you might consider running smp with 11 or 10 cores so that some PC resources are available for other tasks, surfing, gaming, whatever.

[PantherX]

I'm not sure what he means by "officially" but that site was introduced in Vijay's blog. It does have some good info, but instead of adding that info to existing resources like the FAH WIKI or the FAQs, or even this forum, it's just one more place we have to go dig for answers.

I wanted to reassure Zithras that the website is well maintained with latest info rather than some other sites which may appear in Google Results and then complain that the "guides" aren't updated (we had a fairly recent example)

And I'll have to disagree with PantherX. When you set the GPU clients to run at the higher priority, you DON'T need to reserve 2 whole processor cores to drive the GPU clients, especially after setting the EnVars...

I mentioned 2 Threads [= 1 core] which isn't the same as 2 Cores [= 4 Threads] if HT is enabled. Also one may yield better TPF on the SMP2 WU if they have 1 thread free to feed the ATI GPUs with data. However, I would like it if I were proven wrong as it would mean that more resources are spend on F@H As Zithras is using 3 ATI GPUs, there might be more overhead when compared to 2 ATI GPUs, if by using Environment Variables, the CPU Usage can be reduced to 2% (Link) then it would be almost equivalent to Nvidia ones hence there won't be any need to have a free thread. The only way to find out is for Zithras to test his system and if needed, provide the necessary support

[Zithras]

Hmm what's the easiest way to determine which setup is more efficient (12 CPU threads, 3 gpu clients or 10 CPU/3 GPU)?

The best way to seems to be to calculate PPD: How do I do this?
a) Simply running the programs for an hour and checking total points on the site would work, except that 1) I have 5 other machines also running F@H, and would prefer to leave them running and 2) this assumes I will actually turn in enough results during the runtime for an accurate calculation.
b) The best way that I can see is if I could run both setups for an hour and have them calculate and display their own estimated PPD: is there any way to do this? I see no obvious commandline options that would have them display PPD at regular intervals...(nothing on the systray clients either)
c) run the GPU clients, but not the CPU clients, and report back on the forum asking for advice on how many CPU threads to run (10. 11. 12). Problems with this are a) I'd have to ask you all for more help (not really a problem, since the forum posters are quite helpful ) and b) with a multicore CPU, calculating the load while it is distributed across more cores than it would normally have with the CPU client running may introduce unnecessary error/might not show up well on the task manager

Problems with all these approaches seem to be
a) The clients may use different resources when the computer is overclocked, and I don't know what overclocking I'll be able to get out of the CPU and GPU (probably a TON; this is a (over-)watercooled system, and temps right now are rock bottom)

Should I just wait till I overclock it to install and setup the clients? (I'd prefer to deal with them now, since I'm almost done installing all the other programs, but....)
Which method (or another one I haven't thought of?) should I use to calculate the most efficient folding setup? - I've never had to do this before either.

[PantherX]

The best way to calculate the PPD is by installing a third party application like HFM.NET and configure it. Also you may want to select the calculate bonus points box so that your SMP2 Client's PPD will include bonus.

Since your system is stock, I recommend that you run all your Clients with -oneunit flag and capture the WU: (From my bigadv Guide)

2) If you want to experiment with your Overclock, you can "capture a WU" and let your system fold it. Do remember that you will not get any points for it. You can capture a WU by copying the entire SMP2 directory while the WU is being folded. You will have to reconfigure the Client and set this option (in advance configuration):

Ignore any deadline information (mainly useful if
system clock frequently has errors) (no/yes) [no]?

Set it to Yes. You can then hold this WU indefinitely and use it to test your system. In the Work folder, you can keep only wudata_0X.dat and wuinfo_0X.dat files and delete the rest so it will fold from the start. (This may vary from version to version and I tested it with v6.30). I also choose to block the Client from internet connection to ensure that it won't connect to the Servers.

Once they have finished the WU, they can exit and you can begin your OCing session. However, please make a note that in certain situations, your OC may appear stable but when you run F@H, it will give you errors. Hence you will have to tweak your system to ensure that it is "F@H Stable". Please note that I highly recommend that once you reach your OC, to run Stress CPU and MemtestCL to ensure that your system can produce scientifically valid data. Once you are satisfied with your OC, you can then fold the captured WU to ensure that they fold without any errors. Also with this method, you can see how much performance you have gained for F@H with your OC.

Regarding how the CPU resources are allocated, Windows 7 does a good job of loading the physical cores first than the virtual ones.

The way I test what value of N in -smp N is better for my system is like this:
Step 1: I first run the SMP2 Client only with -smp flag
Step 2: I note the TPF for the WU to fold from 0% -> 1% by looking in the Console Window (Value 1)
Step 3: Terminate the SMP2 Client and delete all the files in the Work folder except the wudata_0X.dat and wuinfo_0X.dat files so that I have "reset" my WU to start from 0%
Step 4: Run the GPU Client(s) and the SMP2 Client with -smp flag
Step 5: I again note the TPF from 0% to 1% (Value 2)
Step 6: I compare Value 2 to Value 1. If there is a significant difference, then I lower the N value by 1 in -smp N (i.e in your case, it will be -smp 11) and redo from Step 3 -> Step 5 until I get a Value 2 that is close to Value 1.

I hope that I have explained it in an easy-to-understand manner so you hopefully won't have any difficulties.

[Zithras]

Thanks for the instructions - I'll try it out! I prefer not to use HFM.net, and will try running the console clients directly instead (noting the 1% times, as per your instructions)