cool_gpu

#!/bin/bash
#
# Based on 2011 script by Axel Kohlmeyer <akohlmey@gmail.com>,
# https://sites.google.com/site/akohlmey/random-hacks/nvidia-gpu-coolness
#
# Adapted for Nvidia Titan X Pascal Linux Driver 367
# by Boris Dimitrov <dimiroll@gmail.com> in 2016.
# Use at your own risk!
#
# Suggested installation:
#
#     Unpack somewhere, say in directory
#
#         ${HOME}/set-gpu-fans
#
#     and create a link there from /opt, like so
#
#         ln -sf ${HOME}/set-gpu-fans /opt/set-gpu-fans
#
# Already running an X server?
#
#     Make sure you are not running an X server.  This tool is meant
#     only for headless systems (no display attached).
#
# Suggested usage:
#
#    cd /opt/set-gpu-fans
#    sudo tcsh
#    ./cool_gpu >& controller.log &
#    tail -f controller.log
#
# Seems to work fine on headless ubuntu systems with one or two compute GPUs.
# Not intended to run automatically at system startup, sorry!  However,
# once kicked off, should be good until reboot.
#
# See 'nvscmd' script for the rudimentary fan control policy (one designed
# to eek out a few percent more perf with very noticeable but still
# tolerable fan noise).

dir=/home/liuk/src/set_gpu_fans_public

/opt/bin/nvidia-smi -pm 1

# we need to launch the X server once for each display
# making each of the PCI IDs the primary device in turn.

# get PCI bus ids of Nvidia cards and convert from hexadecimal to decimal. watch out for the falling toothpicks.
pciid=`lspci  | sed -n -e '/VGA compatib.*NVIDIA/s/^\(..\):\(..\).\(.\).*/printf "PCI:%d:%d:%d\\\\\\\\n" 0x\1 0x\2 0x\3;/p'`
disp=0
sar=(`eval ${pciid}`)
gpuid=`lspci  | sed -n -e '/VGA compatib.*NVIDIA/s/^\(..\):\(..\).\(.\).*/\1:\2.\3/p'`
sgar=(${gpuid})
for i in "${!sar[@]}"; do
    cfg=`mktemp /tmp/xorg-XXXXXXXX.conf`
    sed -e s,@GPU_BUS_ID@,${sar[$i]},    \
        -e s,@SET_GPU_DIR@,${dir}, \
        ${dir}/xorg.conf >> ${cfg}
    # we keep one X server per GPU running indefinitely
    # (if we let the X server die, the fan config does not persist, no matter what else)
    xinit ${dir}/nvscmd ${sgar[$i]} run_forever -- :${disp} -once -config ${cfg} &
    echo "xinit ${dir}/nvscmd ${sgar[$i]} run_forever -- :${disp} -once -config ${cfg}"
    disp=$((disp+1))
    sleep 10
    rm -f ${cfg}
done