Well, in similar thread I got the answer that issues with Linux vs Windows cross-validation are known.
So, project's validator is too rigid (if Windows vs Windows is OK, Linux vs Linux is OK and Windows vs Linux results in validation error one can't talk about scientifically correct or incorrect result, just about little too rigid validator cause DB will be filled randomly with both "Windows" and "Linux" results anyway).
In that view GPU vs CPU validation errors shouldn't amuse.
The question is what overall overhead of re-sends because of that and why it tolerated?
If it's impossible to make different builds agree more and still any of builds deems good one then validator could be tuned to allow little more divergence than currently tolerated.
I have five individual iGPUs working on this project. The first three are all identical, Intel i5-4690 CPUs with HD 4600 GPUs.
TASKS FOR COMPUTER 1001562
All (929) | In progress (7) | Pending (290) | Valid (563) | Invalid (22) | Error (0)
TASKS FOR COMPUTER 1001564
All (908) | In progress (8) | Pending (279) | Valid (550) | Invalid (25) | Error (0)
TASKS FOR COMPUTER 1226365
All (875) | In progress (7) | Pending (260) | Valid (545) | Invalid (23) | Error (0)
The fourth is an Intel i5-4570 CPU, also with HD 4600 GPU
TASKS FOR COMPUTER 8864187
All (932) | In progress (9) | Pending (261) | Valid (589) | Invalid (28) | Error (0)
See how all four HD 4600 GPUs have a comparable ratio of Valid/Invalid results. Also note that 'All' is inaccurate, as status "Completed, validation inconclusive" is excluded.
The machine I linked yesterday was the exact machine we used for that SETI@Home bug-hunt. It has an Intel i5-6500 CPU with HD 530 GPU.
TASKS FOR COMPUTER 12496320
All (1121) | In progress (10) | Pending (297) | Valid (666) | Invalid (65) | Error (0)
A little faster, but significantly more invalids. I don't put that down to a greater affinity to Linux wingmates (though somebody else is welcome to check through them ...)
For the record, all five machines are running Windows 7 / x64
So I tried just now... but where is SVN now?? I still have my super-old (few changed from that time) netbook where most of ocl AP was developed left just in "working" state... but it can't connect SVN.
?
)
?
(No subject)
)
Or https://einsteinathome.org
)
Or https://einsteinathome.org/host/12496320/tasks/5/19
Those are mine, not his.
Some years ago, Raistmer coded, and I tested, a similar problem at SETI@Home.
The final conclusion, IIRC, was that an over-optimistic compiler optimisation had invoked a faster, but less precise, fused multiply+add opcode.
Well, in similar thread I got
)
Well, in similar thread I got the answer that issues with Linux vs Windows cross-validation are known.
So, project's validator is too rigid (if Windows vs Windows is OK, Linux vs Linux is OK and Windows vs Linux results in validation error one can't talk about scientifically correct or incorrect result, just about little too rigid validator cause DB will be filled randomly with both "Windows" and "Linux" results anyway).
In that view GPU vs CPU validation errors shouldn't amuse.
The question is what overall overhead of re-sends because of that and why it tolerated?
If it's impossible to make different builds agree more and still any of builds deems good one then validator could be tuned to allow little more divergence than currently tolerated.
I have five individual iGPUs
)
I have five individual iGPUs working on this project. The first three are all identical, Intel i5-4690 CPUs with HD 4600 GPUs.
TASKS FOR COMPUTER 1001562
All (929) | In progress (7) | Pending (290) | Valid (563) | Invalid (22) | Error (0)
TASKS FOR COMPUTER 1001564
All (908) | In progress (8) | Pending (279) | Valid (550) | Invalid (25) | Error (0)
TASKS FOR COMPUTER 1226365
All (875) | In progress (7) | Pending (260) | Valid (545) | Invalid (23) | Error (0)
The fourth is an Intel i5-4570 CPU, also with HD 4600 GPU
TASKS FOR COMPUTER 8864187
All (932) | In progress (9) | Pending (261) | Valid (589) | Invalid (28) | Error (0)
See how all four HD 4600 GPUs have a comparable ratio of Valid/Invalid results. Also note that 'All' is inaccurate, as status "Completed, validation inconclusive" is excluded.
The machine I linked yesterday was the exact machine we used for that SETI@Home bug-hunt. It has an Intel i5-6500 CPU with HD 530 GPU.
TASKS FOR COMPUTER 12496320
All (1121) | In progress (10) | Pending (297) | Valid (666) | Invalid (65) | Error (0)
A little faster, but significantly more invalids. I don't put that down to a greater affinity to Linux wingmates (though somebody else is welcome to check through them ...)
For the record, all five machines are running Windows 7 / x64
So, from ~ 5% to 10% error
)
So, from ~ 5% to 10% error rate. Seems not too insignificant....
And NV/ATi cards don't show such big validation error rate?
Maybe worth to rebuild iGPU binary with similar replacements as we did for SETI.
It would possibly help the
)
It would possibly help the devs if you could post exactly which compiler you used, and which optimisation flag you had to disable.
Richard Haselgrove wrote: It
)
Need to look through SVN history for that. Forgot already :/ But will do on weekend perhaps.
Raistmer* wrote: Richard
)
To save you time searching, our testing was completed on Sunday, 16 October 2016, and the final product was uploaded as r3541.
Oh, thanks!So I tried
)
Oh, thanks!
So I tried just now... but where is SVN now?? I still have my super-old (few changed from that time) netbook where most of ocl AP was developed left just in "working" state... but it can't connect SVN.
So I tried web-interface...
https://setisvn.ssl.berkeley.edu/trac/browser/seti_boinc (beware, certificate is outdated so Chrome will yell) but ... "unsupported module SVN"...
Seems they finally moved away from SVN but how to browse code now??
EDIT: I wrote Eric about that. Hope log history alive somewhere.