Low SETI credit (specifically on GPU)

[trigggl]

I realize that credit is per task and is shared between CPU and GPU tasks, but it seems like tasks at SETI receive about a tenth of the credits that they should. I'm using the nvidia opencl linux app (AstroPulse), but I wonder if there's a way to streamline/optimize/multiply tasks that get done to maximize credits. Does running multiple tasks on a GPU help? Is there an optimized linux app out there somewhere? 15k a day using 3 half-decent GPU's just doesn't cut it, especially since I can get roughly 400k a day on Collatz between two of them when the project is up.

My wife isn't complaining about slowdowns, so I'm thinking the GPU app isn't using enough of the GPU. Also, the office doesn't feel any warmer when these apps are crunching than when there's no GPU app going.

Any thoughts?

[zombie67]

Yes, you should run more than one task per GPU. Assuming a powerful GPU like a 7970, I run 2x AP at a time, reserving .5 of a CPU thread each, or 3x of the regular S@H tasks reserving .333 of a CPU thread each. CUDA uses more CPU, and requires reserving more CPU threads. Adding any more tasks at a time doesn't add any more load to the GPU in my experience. Take a look at your GPU load to see what is best for your mode.

Yes, there are optimized apps for linux, both CPU and GPU: http://lunatics.kwsn.net/index.php?module=Downloads;catd=1

[Maxwell]

Well, you should be getting more than that, I'd say...

I have a 6950 that can get 15-20k/day by itself. Yesterday was 34k, which was a bit of an anomaly. But you do have to run it for quite a while for pendings to catch up to the new tasks you're completing. I'm running 4 WUs at a time on that card, reserving only 1 thread for those four WUs.

I've also run SETI on NVidia cards. I've found that you have to reserve an entire thread per WU, which was just a bit unacceptable for me, so I stopped doing that. And the NVidia cards didn't seem to do as well as the AMD cards.

And I've been running stock apps. No optimization.

[zombie67]

And I've been running stock apps. No optimization.

Well that's just crazy. If you are going to run it, run it right. Go get the all-in-one opt-app installer for windows. It does all the work for you. Then just add an app_config.xml on top of it to easily manage the number of tasks per GPU and threads per task.

http://www.arkayn.us/forum/index.php?PHPSESSID=9b10b7271446ac228a13a864238031 d9&action=tpmod;dl=cat3

[Maxwell]

Well that's just crazy. If you are going to run it, run it right. Go get the all-in-one opt-app installer for windows. It does all the work for you. Then just add an app_config.xml on top of it to easily manage the number of tasks per GPU and threads per task.

http://www.arkayn.us/forum/index.php?PHPSESSID=9b10b7271446ac228a13a864238031 d9&action=tpmod;dl=cat3
I know I should. But you just don't understand how lazy I can be... =))

[zombie67]

Have you tried the all in one installer? Seriously lazy-friendly. Download and click a few buttons. That's it.

[Al]

Have you tried the all in one installer? Seriously lazy-friendly. Download and click a few buttons. That's it.
Z is not lying, it's easy and the only way to run SETI.

[zombie67]

Well that's just crazy. If you are going to run it, run it right. Go get the all-in-one opt-app installer for windows. It does all the work for you. Then just add an app_config.xml on top of it to easily manage the number of tasks per GPU and threads per task.

http://www.arkayn.us/forum/index.php?PHPSESSID=9b10b7271446ac228a13a864238031 d9&action=tpmod;dl=cat3

I just realized that you already have the app_config.xml, based on your earlier post. So...just run the damn app.

[trigggl]

I'll start reserving a core probably after I get to 250k in Cosmology.

I don't think I'm using much CPU at all with the Linux app.

[trigggl]

Yes, you should run more than one task per GPU. Assuming a powerful GPU like a 7970, I run 2x AP at a time, reserving .5 of a CPU thread each, or 3x of the regular S@H tasks reserving .333 of a CPU thread each. CUDA uses more CPU, and requires reserving more CPU threads. Adding any more tasks at a time doesn't add any more load to the GPU in my experience. Take a look at your GPU load to see what is best for your mode.

Yes, there are optimized apps for linux, both CPU and GPU: http://lunatics.kwsn.net/index.php?module=Downloads;catd=1

Didn't see any gpu astropulse apps there. The released one from the project apparently comes from Lunatics anyways.

I had one box with a free core, so I tried running 2 per on the GTX 465, but it doesn't appear to be gaining me anything. The tasks just take twice as long.

Edit: Saw Maxwell's statement about needing a full CPU core per task, so trying that now.

[Shadowlurker]

There are a ton of people with really slow computers running SETI so the deadlines are quite long, it will take you at least 2 to 3 weeks for your pending credits to stabilize, so when you first start out it is hard to tell how much credit you are actually getting.

[denim]

I'm just happy with my 1500+/- RAC on SETI just from the opencl Intel HD I just got fired up. Need to vastly improve that team rank, slowly.

[FourOh]

I noticed that the latest Lunatics Intel GPU app for MultiBeam v7 has increased GPU utilization to the point that my i7 laptop fans spin up quite a bit to keep up. This is good news for the efficiency of the app, but bad news for me on this machine. That laptop sits in my office, and I use ThrottleStop during the day to adjust the CPU multiplier to a level where the fans aren't too loud. Unfortunately, I'm unable to underclock the Intel GPU (I tried using the Intel Extreme Tuning Utility, but Graphics Tuning is locked) so I can't really run the SETI Intel GPU app 24/7.

The Einstein Intel GPU app grants more credits and runs a bit cooler, but uses more CPU... I'll stick with that on this machine.