So I'm willing to bet at this point if the WU is sent to another CPU my result will be validated, but if it is sent to a GPU the GPU result will be validated.
i have 2-3 such tasks per day (out of 45). It seems, that the CPU and the GPU version under some circumstances (not always cause there are also matching results from CPU and GPU) produce slightly different results and then it depends, who is the refferee application. Is it CUDA, the GPU results win, is it SSE, then the CPU results win.
But at all ~5% of my overall BRP tasks are invalid and they are all matched against SSE tasks. This is independent from the GPU configuration. On some machines I use app_info file with task sharing on GPU, on others I do not. Inavild ratio is always the same and always the same reason.
i have 2-3 such tasks per day (out of 45). It seems, that the CPU and the GPU version under some circumstances (not always cause there are also matching results from CPU and GPU) produce slightly different results and then it depends, who is the refferee application. Is it CUDA, the GPU results win, is it SSE, then the CPU results win.
But at all ~5% of my overall BRP tasks are invalid and they are all matched against SSE tasks. This is independent from the GPU configuration. On some machines I use app_info file with task sharing on GPU, on others I do not. Inavild ratio is always the same and always the same reason.
This SHOULD NOT happen!! The Admins need to fix this, if you send a unit to both types of processors both should validate each other! Normally the problem is that gpu's don't have the math precision that cpu's do, but this should be accounted for when they send units to both types of processors for crunching! Have you pm'd an Admin about this?
This SHOULD NOT happen!! The Admins need to fix this, if you send a unit to both types of processors both should validate each other! Normally the problem is that gpu's don't have the math precision that cpu's do, but this should be accounted for when they send units to both types of processors for crunching! Have you pm'd an Admin about this?
The cross validation problems are being looked at quite actively at present.
Cheers, Mike.
I have made this letter longer than usual because I lack the time to make it shorter ...
... and my other CPU is a Ryzen 5950X :-) Blaise Pascal
Is there a wiki page or other description of what exactly these programs are calculating. I'm just curious about what the input/output data looks like and how it's validated.
Not that I could help, more a question of if I can understand.
Is there a wiki page or other description of what exactly these programs are calculating. I'm just curious about what the input/output data looks like and how it's validated.
Not that I could help, more a question of if I can understand.
Joe
Hi,
you could start reading the papers listed under "Science information and progress reports" on the E@H homepage.
Are these just BRP tasks? If so you might consider unselecting them as being able to be crunched by your machine for a bit, at least until they fix the problem.
Another of mine is
)
Another of mine is inconclusive...
So I'm willing to bet at this point if the WU is sent to another CPU my result will be validated, but if it is sent to a GPU the GPU result will be validated.
Which one is actually wrong? Why?
http://einsteinathome.org/workunit/91246842
i have 2-3 such tasks per day
)
i have 2-3 such tasks per day (out of 45). It seems, that the CPU and the GPU version under some circumstances (not always cause there are also matching results from CPU and GPU) produce slightly different results and then it depends, who is the refferee application. Is it CUDA, the GPU results win, is it SSE, then the CPU results win.
But at all ~5% of my overall BRP tasks are invalid and they are all matched against SSE tasks. This is independent from the GPU configuration. On some machines I use app_info file with task sharing on GPU, on others I do not. Inavild ratio is always the same and always the same reason.
RE: i have 2-3 such tasks
)
This SHOULD NOT happen!! The Admins need to fix this, if you send a unit to both types of processors both should validate each other! Normally the problem is that gpu's don't have the math precision that cpu's do, but this should be accounted for when they send units to both types of processors for crunching! Have you pm'd an Admin about this?
RE: This SHOULD NOT
)
The cross validation problems are being looked at quite actively at present.
Cheers, Mike.
I have made this letter longer than usual because I lack the time to make it shorter ...
... and my other CPU is a Ryzen 5950X :-) Blaise Pascal
Is there a wiki page or other
)
Is there a wiki page or other description of what exactly these programs are calculating. I'm just curious about what the input/output data looks like and how it's validated.
Not that I could help, more a question of if I can understand.
Joe
RE: Is there a wiki page or
)
Hi,
you could start reading the papers listed under "Science information and progress reports" on the E@H homepage.
Michael
Team Linux Users Everywhere
RE: Hi, you could start
)
Thanks Michael!
Now up to 88 ! dunx
)
Now up to 88 !
dunx
RE: Now up to 88 ! dunx
)
Are these just BRP tasks? If so you might consider unselecting them as being able to be crunched by your machine for a bit, at least until they fix the problem.
What happens with the
)
What happens with the following WU http://einsteinathome.org/workunit/91506495 ?
Not send, failure during completion, completed - control inconclusive and completed - control inconclusive.
Send out againd and what if the outcome is equal to mine or someebody else's ?