New Folder - PPD Question

adi67

adi67

adi67

Associate
Joined
15 Jun 2008
Posts
143
I'm using a Q6600 at 7x400 for 2.8GHz and a standard clocked ATI 4850. I'm running the GPU client and the SMP client.

FahMon tells me my PPD with this combination is a whisker over 2.5k.

Is that right? I've read some posts on various Folding related forums with people getting many thousands more PPD on a single machine than this - so I want to check I'm not doing something wrong.

Thanks.
 
It's definitely low. What project is each client is running and what's their individual PPD? I'm going to hazard a guess that your SMP client isn't picking up SMP WUs.
 
The SMP client is running project 2665 (R0, C571, G34). It was started with the -smp option - is there something else I need to do to make sure it's getting the right type of WU.

The PPD is split 942 for the SMP client and 1565 for the GPU client (which is running project 4726 (R3, C403, G2)).

Is that the right project info? Taken from FahMon - point me in the right direction if not.

Thanks for the help.
 
The people getting many points per day tend to be making a lot of effort.

e.g. My Q6600 is overclocked to 3.2 (pretty mild by many standards). It runs two Ubuntu Virtual Machines (VM) under VMWare Server (the main OS is WindowsXP). Each VM gets around 512MB RAM and two cpus. Each VM runs a Linux SMP instance. That machine makes just over 4000ppd alone (GPU is a 2400 so not worth running the GPU client).

My E6600 is overclocked to 3.2 and has a single VM. The Ubuntu install has two cpus but I use Windows Task Manager to limit it to core 0. That gets around 1000ppd. The 8800GT in that machine runs the GPU2 client and is getting around 5000ppd. The GPU client eats one of the cpu cores.

By using the Linux SMP client under a VM it's easy to optimise the output and the NVidia GPUs are achieving much higher ppd than any of the ATI cards.
 
Why do the nvidia cards make so much more PPD than the ATI ones? Just for interests sake.

I have no problem using Linux VM's (have done so frequently in the past) if that's the optimal way to run the SMP client. Do I understand correctly then that the Linux SMP client is much more efficient / productive than the Windows one?

Or are points weighted to encourage the take up for, for example, of folding on nvidia cards such that the points awarded don't always reflect just the work involved but also an incentive to fold on particular clients?

I know my overclock is so mild as to be practically not worth mentioning...but in essence my question is: does 2.5k PPD sound right for the setup I outlined in the OP under Vista 64bit and with 4Gb RAM (guess they're all the most relevant specs).

Thanks for the info and help chaps.
 
The official line is the nVidia guys have optimised their client for speed, and are now working on making it stable. The AMD guys have started out with a stable client and are now working on making it quicker. To be honest though, the nvidia GPU2 client is very stable as it is. The AMD client isn't working their latest hardware very well either. The 4850 has something like 800 cores iirc, yet is only using about 300.

The Linux SMP client being better is more to do with the Operating System itself actually. Windows just doesn't like multithreaded communication and the software engineers are finding it a pain in the butt - whereas Linux/OS X was pretty much built for a multi-processor environment and has been for years. And it shows in the client performance basically.

Hope that enlightens you a little.
 
adi67 said:
The SMP client is running project 2665 (R0, C571, G34). It was started with the -smp option - is there something else I need to do to make sure it's getting the right type of WU.

The PPD is split 942 for the SMP client and 1565 for the GPU client (which is running project 4726 (R3, C403, G2)).

Is that the right project info? Taken from FahMon - point me in the right direction if not.

Thanks for the help.
Click to expand...
2665 is an SMP WU but a particularly crap one. My Q6600 at 2.7GHz only gives about 1800 PPD if the client has the PC to itself. If I run two I get about 1400PPD each.
A 3850 on its own gives me 1850 PPD from a 4726.
It might be worth Task Manager to fix the cores used by each client. Give one core to the GPU and the other three to SMP. 2665 is a really horrible WU that drags everything down with it.
 
My quad was at 3Ghz and I got about 9.5kppd with an 8800GT. Running VMWare SMP has two advantages, really. The first is the obvious one - the Linux client is just quicker than the Windows one, for the reasons theheyes pointed out. The second is that running VMs handily avoids all the rubbish WUs, like the 2665 you've picked up. Under VMWare you'll pick up a steady stream of 1760-point WUs - at 2.8Ghz, you can expect 1800-2000ppd for every two cores which you give to SMP.

There's not really much you can do about the ATI client, apart from wait for the ATI guys to optimise it a lot more. The ATI client also seems always to use a whole core, where you can get the nV one down to 15-20% of a core - this is a bummer because it affects your SMP points quite a lot. Even running VMWare SMP, you're gonna have to give the GPU client its own core, which means you'll have an SMP client running on one core only - slow, but still faster overall than Windows SMP.

I reckon you should be able to get 2000ppd on the VMWare SMP client on cores 1 and 2, 1500ppd on the GPU client on core 3, and about 1000ppd on a second VMWare SMP client on core 4. It will probably need a bit more effort to keep everything running smoothly, but you'll almost double your current PPD!

Hope this helps! How many points you can get depends on how much inconvenience and tinkering you can tolerate, really.
 
Last edited:
Guys, that's helpful stuff to a novice folder - thanks. If you can stand it, I have a few more questions...

1. Why does running VMs avoid the WU's like 2665 that generate lower PPD?
2. I thought (from what I've read) that the SMP client was optimized for 4 cores...so what's the deal with running multiple VMs with 1-2 cores each? (ie. 2 x VM instances running the SMP client on 2 cores or the example above where 1 core is devoted to the GPU).
3. My second machine is a creakingly ancient Athlon 64 X2 4200+. Once it had finished it's initial CLI client WUs, I swapped it to the SMP client. As it's running XP 32bit with 2Gb of memory perhaps this would be better reverted to the CLI client?
4. What about I create a native Linux install (probably Fedora or Gentoo as they are the distros I have most familiarity with), either on an HDD or perhaps on a USB drive) and run 1 instance of the SMP client? Would this do better or worse than 2 x VMs under Vista on my quad core?

Sorry for the list, and thanks once again for the input. Hopefully I'll find the best space for my machines to fold in very soon and the questions will cease.
 
adi67 said:
1. Why does running VMs avoid the WU's like 2665 that generate lower PPD?
2. I thought (from what I've read) that the SMP client was optimized for 4 cores...so what's the deal with running multiple VMs with 1-2 cores each? (ie. 2 x VM instances running the SMP client on 2 cores or the example above where 1 core is devoted to the GPU).
3. My second machine is a creakingly ancient Athlon 64 X2 4200+. Once it had finished it's initial CLI client WUs, I swapped it to the SMP client. As it's running XP 32bit with 2Gb of memory perhaps this would be better reverted to the CLI client?
4. What about I create a native Linux install (probably Fedora or Gentoo as they are the distros I have most familiarity with), either on an HDD or perhaps on a USB drive) and run 1 instance of the SMP client? Would this do better or worse than 2 x VMs under Vista on my quad core?
Click to expand...

1. I think that is because people usually allocate a smaller amount of memory to the VM, which triggers the client to avoid the WUs with larger memory requirements. Which just happen to be worth less points.

2. "Optimized" is a bit of a catch-all term for the WUs being benchmarked on a 4 core machine. Also, since it was in beta for a long time, a 4 core machine was the preferred test-bed but people chose to ignore it.

3. If it finishes in time then no, stick with it.

4. If you have enough RAM for the Windows +2VM, then there probably isn't a lot in it. For maximum ppd you would want to run 2 instances under Linux anyway.
 
theheyes said:
1. I think that is because people usually allocate a smaller amount of memory to the VM, which triggers the client to avoid the WUs with larger memory requirements. Which just happen to be worth less points.
Click to expand...
That's a good theory which I shall test when I get home next weekend - currently on holiday in Devon.
 
Well with a change of WU for the SMP client and the CPU speed upped to 3.2GHz I'm now slightly north of 3.5k PPD.

Better. Have to see what I can squeeze out of the old Athlon 64 X2 now.
 
You must log in or register to reply here.
Back
Top Bottom