Amicable Numbers GPU App [Archive]

STMahlberg

02-22-17, 04:16 PM

Is anyone else running the GPU app?

Initially, they all say they'll take 00:19:19. One is pending validation which took 1h and 6m. The ones that my (3) HD5870's are running now say their completion time is 4 hours. I also noticed that the wu's are only using from 6% to 10% of the CPU. I'm wondering if it just doesn't like my GPU's... they are 6 years old.

scole of TSBT

02-22-17, 04:27 PM

Bryan

02-22-17, 04:29 PM

I'm running a GTX 1080 and I see times ranging from 94s to 548s. I've only done 50 WU so far so it isn't definitive by any means. I will say that it is blowing the doors off my 72 thread machines.

I'm guessing that it pays about half of what I was getting on Einsten.

STMahlberg

02-22-17, 04:35 PM

A GTX 1080 running at base clock speed will dust one in 7:00 flat.

Now that's just bragging. Let's assume for a minute that I don't have $600+ to spend on a GPU. :)) I guess I do need to do some upgrading though.

02-22-17, 05:07 PM

Now that's just bragging. Let's assume for a minute that I don't have $600+ to spend on a GPU. :)) I guess I do need to do some upgrading though.
How about a $50 used gpu? My 570 turns one in in about 1070 seconds at 825 (732 is the stock clock.) Errored out at 600 seconds at 900, but I'll bet 850 is doable. I'll test that out in 1.5 days. I haven't even gotten to the Tahiti's yet.

STMahlberg

02-22-17, 05:37 PM

How about a $50 used gpu? My 570 turns one in in about 1070 seconds at 825 (732 is the stock clock.) Errored out at 600 seconds at 900, but I'll bet 850 is doable. I'll test that out in 1.5 days. I haven't even gotten to the Tahiti's yet.

So is it the 570 you're willing to part with for $50?

02-22-17, 07:00 PM

So is it the 570 you're willing to part with for $50?
Well...no.

Dave gpu

02-23-17, 06:22 PM

Now that's just bragging. Let's assume for a minute that I don't have $600+ to spend on a GPU. :)) I guess I do need to do some upgrading though.

GeForce GTX 970 (4095MB) driver: 362.00 OpenCL: 1.2 i3-2130 CPU @ 3.40GHz Windows 10 = 1,072.02 Run time

GeForce GTX 980 (4095MB) driver: 362.00 OpenCL: 1.2 Pentium(R) CPU G3258 @ 3.20GHz Windows 10 = 769.77

GeForce GTX 960 (2048MB) driver: 362.00 OpenCL: 1.2 i5-4670K CPU @ 3.40GHz Windows 7 = 1,790.06

GeForce GTX 750 Ti (2048MB) driver: 362.00 OpenCL: 1.2 AMD FX(tm)-8320 Eight-Core Windows 10 = 3,367.13

STMahlberg

02-24-17, 02:20 AM

GeForce GTX 970 (4095MB) driver: 362.00 OpenCL: 1.2 i3-2130 CPU @ 3.40GHz Windows 10 = 1,072.02 Run time

GeForce GTX 980 (4095MB) driver: 362.00 OpenCL: 1.2 Pentium(R) CPU G3258 @ 3.20GHz Windows 10 = 769.77

GeForce GTX 960 (2048MB) driver: 362.00 OpenCL: 1.2 i5-4670K CPU @ 3.40GHz Windows 7 = 1,790.06

GeForce GTX 750 Ti (2048MB) driver: 362.00 OpenCL: 1.2 AMD FX(tm)-8320 Eight-Core Windows 10 = 3,367.13

I definitely need to upgrade my GPU's. Here's the run times on my Radeon HD 5870's... and thus my problem with fixed credits on Amicable Numbers.

8,117.33
13,113.26
20,684.44

02-24-17, 07:30 AM

I definitely need to upgrade my GPU's. Here's the run times on my Radeon HD 5870's... and thus my problem with fixed credits on Amicable Numbers.

8,117.33
13,113.26
20,684.44
I tried a 5970 on it, but it was obvious it was way to slow. AN seems to have a preference for NV. Tahiti's are not terrible, but not as good as the NVs.

Edit: Though the Tahiti's are slower than newer NV's, they are a big improvement over the cpu version.

Dave gpu

04-01-17, 03:00 PM

Has anybody played with Kernel size for NVIDIA GPUs I see the default is set at 20 .

scole of TSBT

04-01-17, 04:25 PM

Yes. Anything specific? The more kernels, the higher the GPU utilization and faster they run. If you have trouble keeping your GPU cool or trouble with video response/lag, decrease the number of kernels.

Fire$torm

04-01-17, 04:58 PM

Sorry, I'm just running an AMD 5830.

Dave gpu

04-01-17, 05:23 PM

Yes. Anything specific? The more kernels, the higher the GPU utilization and faster they run. If you have trouble keeping your GPU cool, decrease the number of kernels.

Thanks for the info. running 960 70 80 all are running cool so I will try Kernel size 21.

KaptainBlaZzed

06-02-17, 11:31 AM

variability in WU size recently.

my gtx 1080 on windows is taking any where from 170 to 600 seconds to complete a single WU.
here is the host in question: https://sech.me/boinc/Amicable/results.php?hostid=2537&offset=0&show_names=0&state=4&appid=

Up until a few days ago it took ~170 seconds on every WU.
I have checked other host and some show this type of activity and some do not.

What am i missing here? Why is this happening?

is it a windows thing, my Linux hosts are still taking ~170 seconds per WU every time.

Bryan

06-02-17, 01:20 PM

It's because the project doesn't like you KB =))

I looked at my 1 GPU machine which is on Win7. It running 8 threads for AN CPU and 4 threads for AN GPU. ALL GPU tasks are taking 170s and the CPUs are taking 1100s. There is NO deviation on either.

Just for the heck of it you might try rebooting that machine ... pure guess and that is what I always got from the HP IT people when I would call with a problem :D

EDIT: looking at the stderr report the only difference I see is that I'm running the default kernel size of 20 rather than the 21 you are setting yours to.

KaptainBlaZzed

06-02-17, 03:22 PM

lol, maybe it hates me.

i started having nvidia driver crash issues 2 days ago.
I performed a clean nvidia driver install and rebooted the machine multiple times.

Maybe it is time for an complete OS resintall. :(

dcushing

06-04-17, 02:24 PM

I finally broke down and got that Amd 3.5ghz v8 I always wanted.

On a Nvidia 1060, 6min15sec for a work unit....

KB make sure U have visual studio 2013 installed?? Or maybe vs2012?... I have both installed.
Its like Einstein, if U dont have a visual studio installed, all hell breaks loose...

KaptainBlaZzed

06-05-17, 02:03 PM

this is now happening with all 4 of my GPUs 1 win and 3 linux.
CPU time is equal to GPU time for all tasks.

project reset did nothing
project detach and reattach did nothing

einstein used to run fine, have not tried it lately.

Switched to collatz for now.

Bryan

06-05-17, 03:56 PM

KB I looked at the 100 WU I have pending and a 100 or so of my last valids and I'm still seeing no time deviation. They are taking 168s +- less than a second.

I'm on Win 7 using driver 369.30. My cache is only 19 WU so I'm taking them as they come. That is a 12 thread machine and I have the GPU using 4 threads (not needed) and 8 threads doing MT CPU. I looked and the GPU is using a little less than 1/2 of a thread.

KaptainBlaZzed

06-08-17, 12:18 PM

With a little help from the admin i figured out why my times sucked so bad.

My XEON CPU's are too slow to keep the GPU feed with just 1 WU and reserving more CPU's does not help as it appears the GPU app is single threaded.

I am now running 3 WU's on my 1070 and 4WU's on a 1080 and it appears to be working correctly now.
each WU gets its own CPU thread.

Bryan

06-08-17, 01:27 PM

That's a really good thing to know KB since I have similar Xeon machines. My single GTX 1080 is running in a 4.2G I7 and a single WU keeps the GPU at 100% load. Another intersting thing is the I7 is PCIe 2.0 and your Xeon MB is PCIe 3 so it should be quicker.

dcushing

06-11-17, 10:15 PM

All 7 of my gpu's on it for the upcomming challenge.... Cant bunker too well, server only gives 20 work units per gpu...

Bryan

06-12-17, 01:10 AM

All 7 of my gpu's on it for the upcomming challenge.... Cant bunker too well, server only gives 20 work units per gpu...

Why don't you throw 20 GPUs in each box? HERE (http://www.setiusa.us/showthread.php?10423-Multi-GPU-hack&highlight=coproc_info.xml)

dcushing

06-12-17, 10:08 PM

Why don't you throw 20 GPUs in each box? HERE (http://www.setiusa.us/showthread.php?10423-Multi-GPU-hack&highlight=coproc_info.xml)

Cause, its just a hobby...

Bryan

06-12-17, 10:14 PM

Cause, its just a hobby...

Did you even bother to look at the link Doug?

KaptainBlaZzed

06-13-17, 10:30 AM

Having to run 8 WU per 1080 to keep usage up now.

FYI my xeons are downclocked to 1.9GHz due to using AVX.

dcushing

06-13-17, 12:46 PM

Did you even bother to look at the link Doug?

Yes, too complicated to build a co_proc, for my feeble mind!

Mumps

06-13-17, 01:41 PM

So, you didn't read to the very last line then Doug. :))

I'm glad to build the file for you. PM me and we'll get you one built

dcushing

06-14-17, 04:54 PM

Thank You for the offer. Next time! Too late now, challenge started.

KaptainBlaZzed

06-16-17, 03:51 PM

double checking app is quite a bit faster than the regular app and pays the same credits!!!!

somanyroads

06-16-17, 04:51 PM

double checking app is quite a bit faster than the regular app and pays the same credits!!!!

Thanks for all the updates!:-bd

dcushing

06-17-17, 01:14 PM

double checking app is quite a bit faster than the regular app and pays the same credits!!!!

Yes, shifted to that in preferences yesterday.. Hope other teams dont catch on!

Ron Shurtz

06-17-17, 10:22 PM

Yes, shifted to that in preferences yesterday.. Hope other teams dont catch on!

Same here. Thanks KB for the HeadsUp.:D

Dave gpu

06-18-17, 06:39 AM

Late but in. :cool:

zombie67

06-18-17, 08:45 PM

Slightly off topic...

Can someone please post an app_config.xml for the CPU app, so I can limit the number of threads per task?

I tried copying one I had for cosmology, but that uses a vbox plan class, and I don't know the right one for this project.

TIA!

Bryan

06-18-17, 09:40 PM

zombie67

06-18-17, 10:12 PM

You don't need one Z, go into your project preferences and set it there.

Wow. That was nice of them. Didn't even think to look. Thanks!

But I need different values for different machines. In any case, Plan Class = "mt". Simple as that.

Bryan

06-18-17, 10:28 PM

Wow. That was nice of them. Didn't even think to look. Thanks!

But I need different values for different machines. In any case, Plan Class = "mt". Simple as that.

Here is one from before the limits being put into preferences. You may have to change the app name:

<app_config>

<app>
<app_name>amicable_2_64</app_name>
<max_concurrent>1</max_concurrent>
</app>

<app_version>
<app_name>amicable_2_64</app_name>
<avg_ncpus>4.000000</avg_ncpus>
</app_version>

</app_config>

BTW, when you 1st attach it will issue 50-75 WU really quick and assign as many threads as you have. Set your ncpus to the number you want to use per WU and attach. When the WU download they will be using the number you set in ncpus. Then install the app_config and read it. Then you can take the ncpus back to the real number that you have and the project will behave according to the app_config.