I’m guessing Bernd tried to release a non-static build (smaller file size), but it got built with some dependencies from the CUDA toolkit that not everyone has. I think installing the CUDA toolkit would probably resolve the errors, but most end users won’t know to do this extra step.
the statically linked version 1.08 continues to work fine though. I never stopped using the app_info since I have some other customizations that need it.
<core_client_version>7.16.5</core_client_version>
<![CDATA[
<message>
process exited with code 127 (0x7f, -129)</message>
<stderr_txt>
../../projects/einstein.phys.uwm.edu/einstein_O3AS_1.09_x86_64-pc-linux-gnu__GW-cuda: error while loading shared libraries: libcufft.so.10.0: cannot open shared object file: No such file or directory
Sorry, the build went wrong for 1.09. I'm issuing 1.10 now. What the new app should do better than 1.08 is that the "Recalc" step which previous versions did on the CPU, should now also run on the GPU, keeping it more utilized. How much faster that is depends on the CPU of course as well. And validation is also an issue left to check.
Note: 1.08 is "official" now, the new app versions (1.09 and 1.10) are "Beta test", so you can avoid these when having problems by disabling "Beta test work".
I think it's going to be pretty hard to see the difference, if any, given the variability of the workunits. Like you, I've eyeballed some running through and can't really detect much on the older hardware I run (Dual Xeon X5675s with 3-GTX980 Superclocks) considering any change could just be related to a single wu. Perhaps averaging the runtimes of 100 wus might yield some clarity, maybe not. Someone with a newer hotter GPU might benefit.
Either way, the Cuda development and support is much appreciated here! I do see quite a bit of difference versus the 1.07 OpenCL app, perhaps as much as 5 minutes per wu on a 40 minute runtime.
I’m guessing Bernd tried to
)
I’m guessing Bernd tried to release a non-static build (smaller file size), but it got built with some dependencies from the CUDA toolkit that not everyone has. I think installing the CUDA toolkit would probably resolve the errors, but most end users won’t know to do this extra step.
the statically linked version 1.08 continues to work fine though. I never stopped using the app_info since I have some other customizations that need it.
_________________________________________________________________________
Have disabled the beta
)
Have disabled the beta setting as the opencl app works
As a bit of extra
)
As a bit of extra confirmation
https://einsteinathome.org/host/12791541
Ubuntu 22.04.4
all 1.09 tasks failing:
e.g https://einsteinathome.org/task/1587398635
</stderr_txt>
]]>
Just for grins, I installed
)
Just for grins, I installed the CUDA Toolkit.
No change, but whether that's cause there's another issue or the app just didn't know how to find the toolkit I don't know.
Probably needs to be fixed in the build.
Now I'm getting 1.08s, so perhaps 1.09 got pulled back.
Sorry, the build went wrong
)
Sorry, the build went wrong for 1.09. I'm issuing 1.10 now. What the new app should do better than 1.08 is that the "Recalc" step which previous versions did on the CPU, should now also run on the GPU, keeping it more utilized. How much faster that is depends on the CPU of course as well. And validation is also an issue left to check.
BM
Note: 1.08 is "official" now,
)
Note: 1.08 is "official" now, the new app versions (1.09 and 1.10) are "Beta test", so you can avoid these when having problems by disabling "Beta test work".
BM
FYI - The first few 1.10s
)
FYI - The first few 1.10s have passed through here. No problems seen so far.
With the new version 1.10 i
)
With the new version 1.10 i don't see any significant increase in gpu utilization in the part that runs on the cpu.
Almost identical to 1.08!?
I think it's going to be
)
I think it's going to be pretty hard to see the difference, if any, given the variability of the workunits. Like you, I've eyeballed some running through and can't really detect much on the older hardware I run (Dual Xeon X5675s with 3-GTX980 Superclocks) considering any change could just be related to a single wu. Perhaps averaging the runtimes of 100 wus might yield some clarity, maybe not. Someone with a newer hotter GPU might benefit.
Either way, the Cuda development and support is much appreciated here! I do see quite a bit of difference versus the 1.07 OpenCL app, perhaps as much as 5 minutes per wu on a 40 minute runtime.
Judging from the results,
)
Judging from the results, enabling this feature didn't work as planned for 1.10. Hm. I'll keep digging.
BM