If I understand properly how this works copying the appropriate file Jeroen linked to into my BOINC directory and changing the coprocessor count 0.5 should allow 2 CUDA tasks to run.
Not quite. The app_info.xml file (and all other files mentioned therein) go(es) to the appropriate project directory (einstein.phys.uwm.edu in this case), which is located in the projects subdirectory of the BOINC data directory.
Gruß,
Gundolf
Computer sind nicht alles im Leben. (Kleiner Scherz)
If I understand properly how this works copying the appropriate file Jeroen linked to into my BOINC directory and changing the coprocessor count 0.5 should allow 2 CUDA tasks to run.
Not quite. The app_info.xml file (and all other files mentioned therein) go(es) to the appropriate project directory (einstein.phys.uwm.edu in this case), which is located in the projects subdirectory of the BOINC data directory.
Gruß,
Gundolf
Thanks,
I'll give it a try to see what happens. The more I think about the more I think I'll probably run without it in the long term.
This system has only been up for a few days and hasn't got a stable RAC yet but it looks like almost 20K credits/day, no overclocking. I'm guessing another GPU task will add 2 or 3K.
I still think running more then 2 instances on any gpu is not worth doing.
The only gain you have with running 3 is that it actually do the first 2 and when the gpu has some free time spends that on the 3th unit.
But basically it can only do 2 as far as i know nothing has been changed in the current hardware to support more then those 2 instances.
It does not matter if its a ati/nvidia card or even if its a gts250 as long as it has enough video memory and is not fully loaded you can add a second process.
Even my little GT440 with 512 Mb gddr5 can do 2 instances, nevertheless this ofcourse is not worthwhile doing
Only the medium and fast cards really benefit from 2 instances.
As Steve already said the gains can be minimal on projects if the project needs much contact with the cpu like einstein
Projects like milkyway, primegrid and collatz show a much bigger gain cause they hardly use cpu, they simply give the calculation to the gpu with some checks on how far it is.
Else they hardly have to interfere with the process as soon as it has been feeded to the gpu. Milkyway even crashes if you stop it while running.
It simply needs to end the work before control returns to the cpu.
I still think running more then 2 instances on any gpu is not worth doing.
Probably, it is right for more slow cards but just some statistics from GTX 560 Ti card with 2Gb memory:
1. only one BRP4 Wu - about 38 minutes
2. 6 BRP4 Wus samulteniously - about 1 hours and 30 minutes so 150/6= 25 minutes for one Wu.
[ADDED]
2 hours 30 is 9000 seconds, how much time takes 1, at a time?
I'm running a Q6600+GTX470 & a X9650(@3.51GHz)+GTX480.
1 WU on each GPU, 2000 seconds on the 480 and ~2500 seconds
on the 470.
Since the 480 has 1560 MByte DDR5 and the 470 1280MByte RAM, so 2
is an option, IMHO, if compute time doesn't increase, too much!
I did run 3 SETI MB WUs on the 480, changed that to 2, cause of lack
of work and too much heat. (It now runs 'case-less' a 'modular- build')
RE: If I understand
)
Not quite. The app_info.xml file (and all other files mentioned therein) go(es) to the appropriate project directory (einstein.phys.uwm.edu in this case), which is located in the projects subdirectory of the BOINC data directory.
Gruß,
Gundolf
Computer sind nicht alles im Leben. (Kleiner Scherz)
RE: RE: If I understand
)
Thanks,
I'll give it a try to see what happens. The more I think about the more I think I'll probably run without it in the long term.
This system has only been up for a few days and hasn't got a stable RAC yet but it looks like almost 20K credits/day, no overclocking. I'm guessing another GPU task will add 2 or 3K.
Joe
Would be there a real gain on
)
Would be there a real gain on BRP4 CUDA tasks with GTX260?
RE: Would be there a real
)
no, only on Fermi like GPUs. (there was a better performance with BRP3 tasks, but not with BRP4)
I still think running more
)
I still think running more then 2 instances on any gpu is not worth doing.
The only gain you have with running 3 is that it actually do the first 2 and when the gpu has some free time spends that on the 3th unit.
But basically it can only do 2 as far as i know nothing has been changed in the current hardware to support more then those 2 instances.
It does not matter if its a ati/nvidia card or even if its a gts250 as long as it has enough video memory and is not fully loaded you can add a second process.
Even my little GT440 with 512 Mb gddr5 can do 2 instances, nevertheless this ofcourse is not worthwhile doing
Only the medium and fast cards really benefit from 2 instances.
As Steve already said the gains can be minimal on projects if the project needs much contact with the cpu like einstein
Projects like milkyway, primegrid and collatz show a much bigger gain cause they hardly use cpu, they simply give the calculation to the gpu with some checks on how far it is.
Else they hardly have to interfere with the process as soon as it has been feeded to the gpu. Milkyway even crashes if you stop it while running.
It simply needs to end the work before control returns to the cpu.
its my
)
its my app_info:
einsteinbinary_BRP4
Binary Radio Pulsar Search
einsteinbinary_BRP4_1.00_windows_intelx86__BRP3cuda32.exe
cudart_xp32_32_16.dll
cufft_xp32_32_16.dll
db.dev.win.3d35195e
dbhs.dev.win.3d35195e
einsteinbinary_BRP4
100
windows_x86_64
0.200000
1.000000
BRP3cuda32
6.13.0
einsteinbinary_BRP4_1.00_windows_intelx86__BRP3cuda32.exe
cudart_xp32_32_16.dll
cudart32_32_16.dll
cufft_xp32_32_16.dll
cufft32_32_16.dll
db.dev.win.3d35195e
db.dev
dbhs.dev.win.3d35195e
dbhs.dev
CUDA
0.5
einsteinbinary_BRP4_1.00_windows_intelx86__BRP3cuda32.exe
314572800.000000
whats false?
failure: couldn't start No main program specified
RE: failure: couldn't start
)
Refer to the app_info documentation
The format for directives like that is
[pre][/pre]
- a self-closing tag on its own, not an empty open/close pair.
Likewise ,
RE: I still think running
)
Probably, it is right for more slow cards but just some statistics from GTX 560 Ti card with 2Gb memory:
1. only one BRP4 Wu - about 38 minutes
2. 6 BRP4 Wus samulteniously - about 1 hours and 30 minutes so 150/6= 25 minutes for one Wu.
RE: RE: I still think
)
Sorry, TWO hours and 30 minutes.
[ADDED] 2 hours 30 is 9000
)
[ADDED]
2 hours 30 is 9000 seconds, how much time takes 1, at a time?
I'm running a Q6600+GTX470 & a X9650(@3.51GHz)+GTX480.
1 WU on each GPU, 2000 seconds on the 480 and ~2500 seconds
on the 470.
Since the 480 has 1560 MByte DDR5 and the 470 1280MByte RAM, so 2
is an option, IMHO, if compute time doesn't increase, too much!
I did run 3 SETI MB WUs on the 480, changed that to 2, cause of lack
of work and too much heat. (It now runs 'case-less' a 'modular- build')