Gamma-ray pulsar binary search #1 on GPUs

Richie
Richie
Joined: 7 Mar 14
Posts: 597
Credit: 1,687,986,369
RAC: 3,032

Jonathan Jeckell wrote:Crap. 

Jonathan Jeckell wrote:

Crap.  I was wondering why my Linux machine with an NVIDIA GTX 960 wasn't getting any GPU tasks. 

 NVidia device (or driver) doesn't support OpenCL

I had success in preparing my Mint host and GTX960 for both OpenCL and CUDA work by running basically these commands in terminal:


sudo add-apt-repository ppa:graphics-drivers/ppa

sudo apt update

sudo apt install nvidia-375 nvidia-settings

sudo apt-get install nvidia-opencl-icd-375

sudo apt-get install libcuda1-375

sudo reboot

sudo apt-get install boinc-client-nvidia-cuda

sudo apt-get install boinc-client-opencl


Maybe there's something you could use.

Jonathan Jeckell
Jonathan Jeckell
Joined: 11 Nov 04
Posts: 112
Credit: 765,111,554
RAC: 561,863

So after panicking I double

So after panicking I double checked, and the NVIDIA driver I have on Ubuntu has OpenCL 1.2, which I pretty much figured.

 

So why am I getting this error, and why won't it download any FGRPopencl-nvidia tasks? 

I can't get any BRP4 tasks, so it looks like I am going to Winchester on GPU tasks in about an hour.

 

Jonathan Jeckell
Jonathan Jeckell
Joined: 11 Nov 04
Posts: 112
Credit: 765,111,554
RAC: 561,863

Thanks!  I'll give it a shot

Thanks!  I'll give it a shot

Richie
Richie
Joined: 7 Mar 14
Posts: 597
Credit: 1,687,986,369
RAC: 3,032

Jonathan Jeckell wrote:why

Jonathan Jeckell wrote:
why won't it download any FGRPopencl-nvidia tasks?

This project probably doesn't have any work for Linux + Nvidia available at the moment. "Task to send" that you see at server status page https://einsteinathome.org/server_status.html ... those FGRPB1G's might all be for AMD GPU or CPU right now.

If you want to test your system's abilty to run OpenCL, there are still '1.12 (FGRPopencl-nvidia)' tasks of 'Gamma-ray pulsar binary search #1' available for Linux + Nividia ... at Albert@Home :

https://albertathome.org/

https://albert.phys.uwm.edu/apps.php

http://albert.phys.uwm.edu/server_status.html

Jonathan Jeckell
Jonathan Jeckell
Joined: 11 Nov 04
Posts: 112
Credit: 765,111,554
RAC: 561,863

Thanks.  I'll check it out

Thanks.  I'll check it out when I get home tonight.  I've been panicking because my GPU did Winchester and now my average credit is plummeting. 

Bernd Machenschalk
Bernd Machenschalk
Moderator
Administrator
Joined: 15 Oct 04
Posts: 3,962
Credit: 203,684,186
RAC: 29,169

FGRPB1G was relabeled from

FGRPB1G was relabeled from "Beta test" to "production"

CPU app versions were deprecated

I intend to release a preliminary Windows version (not yet as efficient as the others) later today as "Beta test" versions

BM

Mumak
Joined: 26 Feb 13
Posts: 325
Credit: 1,902,909,184
RAC: 612,459

Great news, can't wait for

Great news, can't wait for the Windows version ;-)

-----

Mike Davies
Mike Davies
Joined: 12 Mar 11
Posts: 10
Credit: 20,209,110
RAC: 18,425

Don't know if this is an

Don't know if this is an appropriate comment for this thread, but I've not had any GPU work for several days. Now today, every GPU task is failing after about 27 seconds.

Something to do with compiling OpenCL code.

14:17:47 (5875): [debug]: Flags: X64 SSE SSE2 GNUC X86 GNUX86
14:17:47 (5875): [debug]: glibc version/release: 2.21/stable
14:17:47 (5875): [debug]: Set up communication with graphics process.
boinc_get_opencl_ids returned [0x2d7c610 , 0x7fe013b1e430]
Using OpenCL platform provided by: Advanced Micro Devices, Inc.
Using OpenCL device "Juniper" by: Advanced Micro Devices, Inc.
Max allocation limit: 134217728
OpenCL compiling FAILED! : -11 . Error message: "/tmp/OCLxfWROs.cl", line 10: error: identifier "double2" is undefined
  __kernel void test( __global double2 *vec) {
                               ^
Full grim details at :- https://einsteinathome.org/task/595005534

 
Mumak
Joined: 26 Feb 13
Posts: 325
Credit: 1,902,909,184
RAC: 612,459

Running the first batch of

Running the first batch of FGRPB1G Beta v1.15 on Windows.
GPU usage is heavy fluctuating - 0/100% on Fury X, 0/30% on GTX 1050 Ti with a few seconds on CPU, then a short burst on GPU. On a Tesla K20c the GPU spikes seem to be even more rare. So really not very efficient, looks like most work is done on CPU.
No WU has finished yet, but the estimated run time on a Fury X could be ~ 4500 s, about 4000 s on GTX 1050 Ti (faster CPU).
At least they don't crash..

-----

Oliver Behnke
Oliver Behnke
Moderator
Administrator
Joined: 4 Sep 07
Posts: 854
Credit: 25,165,115
RAC: 9

As Bernd said, the Windows

As Bernd said, the Windows version isn't yet as efficient as the Linux/macOS versions. While the latter do almost all computation on the GPU (if it supports double precision) the former only runs one (essential) part (FFT) on the GPU. We we do need to do a few more tests to ensure stability before we issue the full-fledged version for Windows on a large scale. Please bear with us.

Cheers,
Oliver 

 

Einstein@Home Project

Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.