CUDA fails in v3.4.0 amd64 DISKLESS mode

[skylar]

NVRM: Your system is not currently configured to drive a VGA console NVRM: on the primary VGA device. The NVIDIA Linux graphics driver NVRM: requires the use of a text-mode VGA console. Use of other console NVRM: drivers including, but not limited to, vesafb, may result in NVRM: corruption and stability problems, and is not supported.

bccd@node011:~/CUDA$ ./device-query cudaDriverGetVersion() FAILED, status = 30 (unknown error)

[skylar]

In 5211//cluster/svnroot:

v3.4.0 needs explicit compat32 libdir re #896

[skylar]

In 5215//cluster/svnroot:

fix nvidia_build for v3.4.0 re #896

[skylar]

In 5219//cluster/svnroot:

remove kernel dependency for nvidia build re #896

[skylar]

In 5240//cluster/svnroot:

merge addresses #896, #898, #915, #946, #949, #950

[skylar]

In 5255//cluster/svnroot:

disable nouveau by default re #896

[skylar]

In 5260//cluster/svnroot:

rolling back gcc 5 changes, no openacc for now re #910, #896

[skylar]

In 5261//cluster/svnroot:

cuda 5 comes with both i386 and amd64 re #896

[skylar]

In 5264//cluster/svnroot:

roll back cuda module version re #896

[skylar]

In 5265//cluster/svnroot:

need gcc-4.8 for cuda re #896

[skylar]

In 5266//cluster/svnroot:

remove gcc-4.9 re #896

[skylar]

In 5268//cluster/svnroot:

nvidia build needs ruby-ffi re #896

[skylar]

sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-4.8 60 --slave /usr/bin/g++ g++ /usr/bin/g++-4.8 --slave /usr/bin/cpp cpp /usr/bin/cpp-4.8 --slave /usr/bin/cc cc /usr/bin/gcc-4.8

[skylar]

​https://devtalk.nvidia.com/default/topic/813458/linux/linux-4-0-rc1-346-35-build-error-_cr4-functions-fix/

[skylar]

​https://devtalk.nvidia.com/default/topic/813458/linux-4-0-rc1-346-47-build-error-_cr4-functions-fix/

[skylar]

Apply patch:

1. Extract NVIDIA package: ./NVIDIA-Linux-x86_64-340.76.run -x
2. pushd NVIDIA-Linux-x86_64-340.76/kernel
3. patch < /path/to/nv-pat.patch

[skylar]

In 5269//cluster/svnroot:

create gcc 4.8 symlinks re #896

[skylar]

In 5270//cluster/svnroot:

paren re #896

[skylar]

In 5271//cluster/svnroot:

remove cpp alternatives re #896

[skylar]

In 5272//cluster/svnroot:

add cc alternative re #896

[skylar]

In 5273//cluster/svnroot:

need to get uvm module too re #896

[skylar]

In 5274//cluster/svnroot:

properly manage /usr/bin/cc links re #896[

[skylar]

In 5275//cluster/svnroot:

link to /usr/bin/cpp re #896

[skylar]

In 5276//cluster/svnroot:

versioned cpp re #896

[skylar]

In 5280//cluster/svnroot:

merge addresses #896, #910 #951, #953, #954

[skylar]

rebuild kernel and .deb's w/ gcc 4.8

[skylar]

In 5282//cluster/svnroot:

new kernel revision re #896

[skylar]

Confirmed that NVIDIA-Linux-x86_64-340.76 and cuda 6.5.14 play nicely, but cuda 6.5.14 is 1.5GB.

Next up is trying cuda 6.0.37

[skylar]

cuda 6.0.37 nvidia driver (331) does not build against linux 4.0.0, does not like nvidia 340.76

size is 1.8GB, could be reduced closer to 1GB though

[skylar]

seems that loading nvidia_uvm and then running as root will solve the problem with 5.5.22

[skylar]

nvidia device nodes:

crw-rw-rw- 1 root root 249, 0 Aug 19 00:21 /dev/nvidia-uvm crw-rw-rw- 1 root root 195, 0 Aug 18 23:29 /dev/nvidia0 crw-rw-rw- 1 root root 195, 255 Aug 18 23:29 /dev/nvidiactl

[skylar]

In 5285//cluster/svnroot:

adding nvidia_uvm to load list re #896

[skylar]

In 5288//cluster/svnroot:

create nvidia-uvm node re #896

[skylar]

In 5291//cluster/svnroot:

merge addresses #896, #933, #939, #953