Opt ATI app on Linux x64
log in

Advanced search

Message boards : Unix/Linux : Opt ATI app on Linux x64

Author Message
Skip Da Shu
Avatar
Send message
Joined: 31 Jul 09
Posts: 21
Credit: 59,804,098
RAC: 0
Message 9407 - Posted: 30 Aug 2010, 0:56:47 UTC

Environment:
Ubuntu 10.04, x64
BOINC v6.10.58 (from PPA, Debian package install)
ati-driver-installer-10-7-x86.x86_64 (glxinfo broken so I can't prove at the moment)
single HD5830 card

App
collatz_2.01_x86_64-pc-linux-gnu__ati14_r1.5_mgpu.tar.bz2

Think I followed the README but got "download failed". I think it was trying to download the libbrook* libs it found in the app_info.xml.

I found this. Moved the '64' libs to the project directory and changed the 'non-64' libs in /usr/lib64 to be symbolic links instead of copies of the '64' libs. Also changed them in my /test folder and re-ran the test... all works in the stand-alone test.

Just started boinc back up and it downloaded 4 WUs and is crunching on one:
Sun 29 Aug 2010 07:23:02 PM CDT Collatz Conjecture Starting task collatz_1282112751_844469_0 using collatz version 201

So it's working.

I am curious why the '64' libs need to be in the project folder though. Why can't it locate them in /usr/lib64? Also is there some sort of link I could put in the project directory to point it to copy in /usr/lib64 instead?

Next Post "Tuning Opportunities"

Skip Da Shu
Avatar
Send message
Joined: 31 Jul 09
Posts: 21
Credit: 59,804,098
RAC: 0
Message 9408 - Posted: 30 Aug 2010, 1:18:04 UTC

Tuning Opportunities:

I wanna ask about the following (test result) but have to get to work now...be back l8r.

shmget in attach_shmem: Invalid argument
17:51:02 (25346): Can't set up shared mem: -1. Will run in standalone mode.

Running Collatz Conjecture (3x+1) ATI GPU application version 2.01 by Gipsel (Linux64, CAL 1.4 - R1.5 multi-GPU Support)
Reading input file ... done.
Checking 4294967296 numbers starting with 2361185725354183731560
CAL Runtime: 1.4.736
Found 1 CAL device

Device 0: ATI Radeon HD5800 series (Cypress) 1024 MB local RAM (remote 247 MB cached + 875 MB uncached)
GPU core clock: 855 MHz, memory clock: 1000 MHz
1120 shader units organized in 14 SIMDs with 16 VLIW units (5-issue), wavefront size 64 threads
supporting double precision

Initializing lookup table (16384 kB) ... done
Setting BRT to use GPU 0
Starting WU on GPU 0
Copy lookup table to GPU memory (16384 kB)
Initialize step array on GPU (256 MB)
predicted runtime per iteration is 43 ms (33.3333 ms are allowed), dividing each iteration in 2 parts
borders of the domains at 0 2048 4096
Resuming from checkpoint ... done
needed 1516 steps for 2361185725355416900338
2242246260675 total executed steps for 4294967296 numbers

WU completed.
17:51:10 (25346): called boinc_finish

____________
- da shu @ HeliOS,
"A child's exposure to technology should never be predicated on an ability to afford it."

Skip Da Shu
Avatar
Send message
Joined: 31 Jul 09
Posts: 21
Credit: 59,804,098
RAC: 0
Message 9420 - Posted: 31 Aug 2010, 0:11:01 UTC - in response to Message 9408.

My original thought was that this:


Initialize step array on GPU (256 MB)
predicted runtime per iteration is 43 ms (33.3333 ms are allowed), dividing each iteration in 2 parts
borders of the domains at 0 2048 4096
along with a GPU load in the 82~87% range might indicate some tuning was in order but...

I changed the app_info.xml entry that controls how many GPUs to use to 0.5 so it'd run two at a time. WU times went up quite a bit, but not double and GPU is now 93~98%. Was looking good the other day but no more WUs since then. :-(

If somebody thinks there's some tweaking that should / could be done to <cmdline> please advise. I was gonna play with 'f20' setting but no WUs so that's on a back burner now.

Thanx, Skip

____________
- da shu @ HeliOS,
"A child's exposure to technology should never be predicated on an ability to afford it."


Post to thread

Message boards : Unix/Linux : Opt ATI app on Linux x64


Main page · Your account · Message boards


Copyright © 2018 Jon Sonntag; All rights reserved.