Reduce snap size by splitting CUDA support into different channels #2
Description
This snap is very large due to the size of the included dcgmi deb (874MB compressed).
Looking at its contents, most of the used space comes from libdcgm_cublas_proxy1{0,1,2}.so libraries, which are needed by the dcgmproftester binaries. For example:
# CUDA 12
root@jammy:~# ldd /usr/bin/dcgmproftester12
linux-vdso.so.1 (0x00007ffe2b5ba000)
libdl.so.2 => /usr/bin/../lib/x86_64-linux-gnu/libdl.so.2 (0x00007ca83fb88000)
librt.so.1 => /usr/bin/../lib/x86_64-linux-gnu/librt.so.1 (0x00007ca83fb83000)
libcuda.so.1 => /usr/bin/../lib/x86_64-linux-gnu/libcuda.so.1 (0x00007ca83df16000)
libpthread.so.0 => /usr/bin/../lib/x86_64-linux-gnu/libpthread.so.0 (0x00007ca83df11000)
libdcgm_cublas_proxy12.so => /usr/bin/../lib/x86_64-linux-gnu/libdcgm_cublas_proxy12.so (0x00007ca81cefe000)
libdcgm.so.3 => /usr/bin/../lib/x86_64-linux-gnu/libdcgm.so.3 (0x00007ca81cb57000)
libm.so.6 => /usr/bin/../lib/x86_64-linux-gnu/libm.so.6 (0x00007ca81ca70000)
libgcc_s.so.1 => /usr/bin/../lib/x86_64-linux-gnu/libgcc_s.so.1 (0x00007ca81ca50000)
libc.so.6 => /usr/bin/../lib/x86_64-linux-gnu/libc.so.6 (0x00007ca81c827000)
/lib64/ld-linux-x86-64.so.2 (0x00007ca83fb8f000)
root@jammy:~# ls -lah /usr/bin/../lib/x86_64-linux-gnu/libdcgm_cublas_proxy12.so
-rwxr-xr-x 1 root root 489M Jul 9 19:58 /usr/bin/../lib/x86_64-linux-gnu/libdcgm_cublas_proxy12.so
root@jammy:~# du -sh /usr/share/nvidia-validation-suite/plugins/cuda12
127M /usr/share/nvidia-validation-suite/plugins/cuda12
# CUDA 11
root@jammy:~# ldd /usr/bin/dcgmproftester11
linux-vdso.so.1 (0x00007ffc362df000)
libdl.so.2 => /usr/bin/../lib/x86_64-linux-gnu/libdl.so.2 (0x000071f029863000)
librt.so.1 => /usr/bin/../lib/x86_64-linux-gnu/librt.so.1 (0x000071f02985e000)
libcuda.so.1 => /usr/bin/../lib/x86_64-linux-gnu/libcuda.so.1 (0x000071f027bf1000)
libpthread.so.0 => /usr/bin/../lib/x86_64-linux-gnu/libpthread.so.0 (0x000071f027bec000)
libdcgm_cublas_proxy11.so => /usr/bin/../lib/x86_64-linux-gnu/libdcgm_cublas_proxy11.so (0x000071f003afe000)
libdcgm.so.3 => /usr/bin/../lib/x86_64-linux-gnu/libdcgm.so.3 (0x000071f003757000)
libm.so.6 => /usr/bin/../lib/x86_64-linux-gnu/libm.so.6 (0x000071f003670000)
libgcc_s.so.1 => /usr/bin/../lib/x86_64-linux-gnu/libgcc_s.so.1 (0x000071f003650000)
libc.so.6 => /usr/bin/../lib/x86_64-linux-gnu/libc.so.6 (0x000071f003427000)
/lib64/ld-linux-x86-64.so.2 (0x000071f02986a000)
root@jammy:~# ls -lah /usr/bin/../lib/x86_64-linux-gnu/libdcgm_cublas_proxy11.so
-rwxr-xr-x 1 root root 548M Jul 9 19:58 /usr/bin/../lib/x86_64-linux-gnu/libdcgm_cublas_proxy11.so
root@jammy:~# du -sh /usr/share/nvidia-validation-suite/plugins/cuda11
132M /usr/share/nvidia-validation-suite/plugins/cuda11
iiuc you can't really have multiple CUDA versions installed at the same time, so if we had separate snap channels for CUDA 11 and 12 we'd split the almost 1.3GB shown above into two separate snaps.