3WU BRP3cuda on a single GPU
4 Jan 2011 20:53:48 UTC
Topic 195553
(moderation:
Only for advanced users. any problems on your conscience. If you do not know what to do - does NOTHING!!!!
Only for windows x86. x64 has not been tested.
1WU uses more than 256M GPU memory.
tested on XP SP3. 260.99 nvidia drivers. BOINC 6.10.58.
9600GT 512M -1WU.
9600GSO 768M - 2WU.
GTS250/GTS450 1024M - 3WU.
All WU passed validation.
BRP3 application runs with low priority is used no more than 30 cuda cores. If you change the priority to "realtime", using no more than 50 cuda cores (run time is reduced by about 2-fold).
to change the priority, you can use Process Tamer (This program is FREE, donations welcome).
Depending on your GPU, you must select the number of simultaneously running WU, and their priority.
The missing files can be downloaded here.
In the project(Data directory)folder E@H (einstein.phys.uwm.edu), you must create app_info.xml
If you want to change the number of WU on a GPU, change the flag . 2WU = 0,5. 3WU = 0.33. 4WU = 0.25.
GC S5HF for CPU + 3 BRP3 cuda. (BRP3 for the CPU is not supported).
listing app_info.xml
einstein_S5GC1HF
Global Correlations S5 HF search #1
einsteinbinary_BRP3
Binary Radio Pulsar Search
einstein_S5GC1HF_3.06_windows_intelx86__S5GCESSE2.exe
einstein_S5R6_3.01_graphics_windows_intelx86.exe
einsteinbinary_BRP3_1.04_windows_intelx86__BRP3cuda32.exe
einsteinbinary_BRP3_1.00_graphics_windows_intelx86.exe
cudart_xp32_32_16.dll
cufft_xp32_32_16.dll
db.dev.win.4330b3e5
dbhs.dev.win.4330b3e5
einsteinbinary_BRP3
104
windows_intelx86
0.200000
0.500000
BRP3cuda32
6.13.0
einsteinbinary_BRP3_1.04_windows_intelx86__BRP3cuda32.exe
cudart_xp32_32_16.dll
cudart32_32_16.dll
cufft_xp32_32_16.dll
cufft32_32_16.dll
einsteinbinary_BRP3_1.00_graphics_windows_intelx86.exe
graphics_app
db.dev.win.4330b3e5
db.dev
dbhs.dev.win.4330b3e5
dbhs.dev
CUDA
0.330000
220200960.000000
einstein_S5GC1HF
306
windows_intelx86
S5GCESSE2
6.13.0
einstein_S5GC1HF_3.06_windows_intelx86__S5GCESSE2.exe
einstein_S5R6_3.01_graphics_windows_intelx86.exe
graphics_app
PS.. HAPPY NEW YEAR.
3WU BRP3cuda on a single GPU
)
This is excellent. Thanks for posting the config.
My king regards to leks for
)
My king regards to leks for this app file. The output of the box is something unbeliveble. I`ve never seen something like this in enshtein before. Thank you again!
The box with GTX580 could run 4 threads, GTX570 only 3 because of the shortage with memory
Thank you very much for the
)
Thank you very much for the posting, leks!
Works like a charm here
Windows 7 Ultimate 64bit
BOINC 6.10.58
Intel
Gigabyte GTX 460 OC 1GB
GPU load with 2 concurrent BRP processes at high priority
between 75 and 80%
GPU Memory used: around 630-670GB
2 more CPU processes at low priority.
That's what I call throughput!
Armin
Holy Higgs Boson.....I can't
)
Holy Higgs Boson.....I can't imagine ever having 167 million wu's finished on my machines.
There is many examples,
)
There is many examples, including those with more than the billion crunched
Shih-Tzu are clever, cuddly, playful and rule!! Jack Russell are feisty!
RE: There is many examples,
)
If you're talking about BRP3 (the subject of this thread) there's a 500-fold difference between the number of WUs crunched and the number of credits awarded.
I'd like to mention two
)
I'd like to mention two things to keep in mind:
1) just being able to run several apps in parallel doesn't necessarily increase productivity. You should closely monitor the crunchtime, at least one volunteer has reported that for his card, there is no increase in productivity
2) If you apply the app_info.xml file and you do get a productivity boost, please keep in mind that this will disable the automatic app update feature of BOINC. You should come back here to check regularly if there's a new app version (not just for BRP3, all Einstein@Home apps will be affected by the app_info.xml).
Yes I know, this thread was explicitly directed to experienced users only, and 99.99% of those will know all this already, but you know how HOW-TOs like this one spread over the internet and will be adopted by novice volunteers in the end.
Happy crunching
HBE
From some testing I did last
)
From some testing I did last night, here is what I have seen so far with the 580 going from running one task at once to three.
1 WU - 33-34 minutes - 33-40% GPU usage - 300-350MB memory usage
3 WU - 60-65 minutes - 73-75% GPU usage - 1GB memory usage
I had all CUDA tasks set to high priority.
RE: 1) just being able to
)
It does increase because of using most of available cores in graphic cards.
But it leads to a higher GPU temperatures and lowers the stability of memory and core. But this sorts out by WU validation process.
P.S. Nice job Leks. Now I will sure be right after you.
Hello, thank you for this
)
Hello, thank you for this file app_info.
For 64-bit is it that I change some things?
What is more to these lines:
db.dev.win.4330b3e5
dbhs.dev.win.4330b3e5
db.dev.win.4330b3e5
db.dev
dbhs.dev.win.4330b3e5
dbhs.dev
and : 6.13.0
I did a test and it takes 900 MB of memory for 70% load GPU ...
Besides, version 1.05 is out today
thanks you