GPU or SMP for the new client?

[DrCoras]

Hi there!

I am a little new to the details of this and everything was moving very slowly for me, but it all seemed confusing as to installations and GPU and SMP clients. I also didn't understand the tags, etc.

I have installed the new beta v7 F@H client and it is working great, but suring setup it asked me if I wanted to run SMP or GPU or Uniprocessor. Now, I was looking for the biggest returns as far as most work done in shortest amount of time, and I chose GPU, but I am not sure how to check what the SMP would do.

I don't know what info is needed, but my basic hardware is:

core i5-2500k [edited]
8GB ram
ati 6870

And while this PC is great for everything I have tried with it so far, I do not know which of the options would be my best bet. I have seen mentions of people running multiple clients, but considering the problems I was having getting just 1 client to run, that option scares me a little.

I am sorry if this is not the right place for this, but I have read tons of threads and since I am not getting any closer to understanding, everything is sort of jumbling together now. Thanks for any help you can give, I appreciate it.

DrCoras

[Zagen30]

Welcome to the forum, DrCoras.

The great thing about the new client is that it can run "clients", now called "slots", for all your hardware at once, rather than before where you'd have to install multiple clients to max out your hardware. Follow the Configure instructions under the Advanced section at https://fah-web.stanford.edu/projects/F ... FahControl (after setting your client to Advanced mode) to see how to set up another slot to run SMP WUs on your CPU. I'd also recommend getting a passkey at http://fah-web.stanford.edu/cgi-bin/getpasskey.py and inputting it for both slots, especially the SMP slot since SMP WUs earn bonus points based on how quickly they're returned if you have a passkey.

[DrCoras]

Ok great, did not know that, so now I don't need the seperate SMP client!

I configured my SMP to the specs that PantherX used in his SMP setup tutorial, so hopefully we shall see where it takes me!

Also, on my older rig with an nvidia 8800 and a dual core I am running seperate GPU3/Uni setups, would I benefit from moving that PC to the new vers 7 setup as well? Other than having 1 program running instead of 2?

Thanks again!

[Zagen30]

Changing the old rig wouldn't speed anything up. The actual calculations are performed by the FAHCore, and the Nvidia and SMP cores haven't been updated since before V7 was released (nothing wrong with that, just haven't needed to be updated in a couple weeks/months), so V7, which uses the same cores, won't show any gains there. In the future the Nvidia and SMP cores will probably require V7 to be able to run, so at that point you'd need to upgrade, but I don't think that'll happen for a while (i.e. months to years) since they work just fine on V6.

If you don't mind running slightly beta WUs, a new ATI core (16) was released last week that is supposed to utilize ATI cards much better than the old core (11). If you switch to Expert mode, follow the instructions on how to add extra slot options, and set your GPU client to client type: advanced, you should pick up the new WUs (availability is limited right now, so you may get Core 11 WUs some of the time). This is not at all necessary, and the core is still in the early stages of open beta (right now the projects are just running old simulations to make sure the new core produces the right values), so feel free to ignore that part. Once the new core has been shown to be stable and valid, the Core 16 projects will be sent out to people even if they don't have client-type:advanced set (probably in a few weeks to a couple months).

[DrCoras]

Ok, thanks for all the great info!

I am actually running GPU3 with Uniprocessor on my older PC(the dual core) just because everything I was reading stated it was better than SMP for anything less than a quad core. Maybe I will upgrade to vers7 soon just for ease, but it is not a priority.

Thanks again for all the info, I appreciate it!

[codysluder]

I agree with Zagen30. There's really no urgency to upgrade your dual+8800 from a working V6 installation to a beta V7 installation. Upgrading the ati 6870 client is MUCH more important.

If you really like beta testing, then upgrade that system ASAP. If you're more cautious, you can start slowly by stopping any V6 GPU client that might be running the HD6870 and installing the V7 client [/u]just for the GPU[/u]. Once that is working and you have time to get comfortable with V7, you can decide when it's right to suspend the V6 SMP client and add a SMP slot to the V7 client.

[DrCoras]

Cody,

I was getting about 1.2kppd with just Smp on my main pc when I wasn't running a gpu. Now that I've updated to v7 and set it up for gpu and added smp I am down to 341 ppd on my smp but the gpu is chugging along like a champ, finishing in less than 3 hours per wu.

Are the 2 incompatible? For my smp to drop so drastically seems strange.

[codysluder]

With beta software, anything can happen. On one hand, the FAHCore that's doing the work has not changed so many are saying that there isn't any performance change.

If you're using the PPD reported by V7, it's very different than something you may have been getting from a 3rd party tool on V6. Don't trust it. The only way to tell if you're really producing less is to look at the TPF. How long did it take to complete (say) 5% in v6 and how long does it take in V7?

WUs started in either V6 or V7 must be completed with that code, and you DO NOT want to run two SMP clients or two GPU clients at the same time. It will work, but will be very slow.

[DrCoras]

A big problem is that before v7 I never really had a steady stable setup, so I'm limited in comparison. As far as tpf, my gpu instance in v7 is averaging 1m50s and my smp is doing 1h20m tpf. The smp is definitely running slower than it did in v6.x I'm not sure if it's the wu but it's giving me a 4 day time of completion! I am not home, but I will be able to provide more information when I get home.

[codysluder]

When you get home, do confirm that SMP is getting nearly 100% of your processor. That's probably not true.

The ATI GPU does use a lot of CPU time. The beta FAHCore has not been optimized yet so it's probably slowing down SMP more than you want it to. WIth SMP, it's convenient to allocate one virtual core to the GPU. You have an i5 so that's not possible.

You can compare two scenarios and figure out which one works best for you:
1) Run SMP on all of your CPUs and give it priority. Let the GPU production suffer by not getting all the cpu time it wants. It will still produce some.
2) Give the GPU access to one full CPU by reducing SMP to -smp 3. Now SMP will suffer but the GPU will produce a lot more.

I can't be certain which option will produce the most.

You do NOT want to run with -smp (4) and with the GPU at a higher priority, which has often been recommended for nvidia GPUs.

[phoFAH]

As I understand it, -smp now requires even numbers, so, no -smp 3. you need to set it to 2 or 4. Since you want the GPU to get all the CPU time it needs because GPU is way way way faster then you want to leave 1 CPU core free for it.

so, for example, if you have a quad core CPU, and one ATI GPU, you would want to run the GPU, 1 SMP (-smp 2) and 1 uniprocessor. that will leave 1 core free for the GPU, 2 cores for an smp, and 1 core for a uniprocessor.

on the other hand if you only have a dual core CPU, it would be 1 GPU, and 1 uniprocessor. the GPU would eat up too much CPU to run an smp client on it.

[bruce]

-smp 3 works. Just type the 3 in the designated field. The even numbers have been found to have fewer EUEs than the odd numbers but if you choose to ignore that recommendation please don't complain. If your EUE rate stays low, fine. If it is high and you find that others are successfully finishing the same WUs that you report as "bad", you can deal with that situation when (and if) it happens.

[DrCoras]

I did not see an option to select 3 in the Client Control, the drop down box under configuration only lets you select an even number, but I am running it with 2 right now to see if that changes anything.

And I also set the priority to low for SMP, so hopefully that will help also.

[bruce]

The option is not on the drop-down list. Type the number you want.

[JimboPalmer]

On the older rig with dual core, using v7 with SMP will yield more points than two uniprocessors.