GPUs

Add your tips for effective work with GPUs:

Verify that

you are a member of dsi-gpu-users, and read their threads

Show GPU usage on your machine

source /cortex/code/scripts/gpustat.py

Run a command directly on a dsigpu

ssh dsihead ssh dsigpu01 ls

To make your life easier, add the following to your .bashrc:

alias dgx01="ssh -A -t dsihead.lnx.biu.ac.il ssh -A -t dgx01"

alias dgx02="ssh -A -t dsihead.lnx.biu.ac.il ssh -A -t dgx02"

alias dgx03="ssh -A -t dsihead.lnx.biu.ac.il ssh -A -t dgx03"

Snapshot of GPU usage (repeatedly SSHs into machines)

source /cortex/code/scripts/scan_gpu.sh

Run on a specific Server/GPU

source /cortex/code/scripts/run_gpu.sh 43 ls

Does the following:

ssh into dsigpu04,

export the CUDA visible GPU to 3

run ls

How to profile GPU utilization:

Instructions here

Legacy instructions for local compgpu1 (2016)

For the local lab GPU (updated 2016)

GPU computer IP address: 132.71.84.233

GPU computer hostname: gpucomp1

Checking GPU usage (from the command line): nvidia-smi

To setup the GPU (and CAFFE) libraries paths, add the following to your ~/.bashrc file:

GPU1=gpucomp1

host=$(hostname)

if [[ $host == $GPU1 ]]

then

# Setting GPU envs

export LD_LIBRARY_PATH="$LD_LIBRARY_PATH:/usr/local/cuda/lib64:/usr/local/cuda-7.5/extras/CUPTI/lib64"

export CUDA_HOME=/usr/local/cuda-7.5

export CUDA=/usr/local/cuda-7.5

export PATH="$PATH:/usr/local/cuda-7.5/bin"

# CAFFE

export PYTHONPATH=/local_stuff/caffe/python:$PYTHONPATH

fi