[HTCondor-devel] Thoughts on CUDA version and ClassAds

["Bockelman, Brian" <BBockelman@xxxxxxxxxxxxx>]

Hi all,

After digging in to the various CUDA attributes and a good chat with TJ, we've arrived at the following proposal for evolving attributes:

â Deprecate the CUDARuntimeVersion attribute.  This has not been working for over a year and was a misleading attribute when working.  (We should document it was removed, however.)
    â For the 8.8.x series, we can revive it if someone depends on it.
â In 8.9.x: Alias CUDADriverVersion to CUDAMaxSupportedRuntime.  That is, advertise both for the time being.
    â Open to arguments on aliased name (CUDAMaxRuntimeVersion?)!  We want to indicate this is the maximum supported version
â In 8.9.x: Enumerate all the available libcudart.so variants (equivalent of `ldconfig -p | grep 'libcudart.so\.' | awk '{print $1;}'`; on Windows, this might be a crawl through the registry).  Invoke the library to determine the correct version to advertise.  Advertise this as CUDARuntimeLibraries.  The attribute will be of the form CUDARuntimeLibraries=â8.0, 9.1, 10.1â.
â On request (8.9.x only): If a user asks for it, we can also advertise the kernel driver version as CUDAKernelDriverVersion; the attribute is of the form CUDAKernelDriverVersion=â418.87.00â.  If there's no strong argument for this, then we can skip it.

Thoughts?  Opposition?  TJ, did I forget anything from yesterday?

Thanks,

Brian