This page summarizes which UFAL servers have some GPU card, and suggests basic diagnostic commands, paths to installed tools, etc., simply everything necessary at the very beginning of using GPUs for experiments.

GPU cluster gpu-ms.q at Malá Strana:

GPU cluster gpu-troja.q at Troja:

Desktop machines:

Multiple versions of CUDA library are accessible on each machine together with cudnn. Theano and TensorFlow is supported.

GPU usage rolling graphs

• First, read Linux network and Grid.
• All the rules from Grid apply, even more strictly than for CPU because there are too many GPU users and not as many GPUs available. So as a reminder: always use GPUs via qsub (or qrsh), never via ssh. You can ssh to any machine e.g. to run nvidia-smi or htop, but not to start computing on GPU. Don't forget to specify you RAM requirements with e.g. -l mem_free=8G,act_mem_free=8G,h_data=12G.
  • Note that you need to use h_data instead of h_vmem for GPU jobs. CUDA driver allocates a lot of “unused” virtual memory (tens of GB per card), which is counted in h_vmem, but not in h_data. All usual allocations (malloc, new, Python allocations) seem to be included in h_data.
• Always specify the number of GPU cards (e.g. gpu=1), the minimal Cuda capability you need (e.g. gpu_cc_min3.5=1) and your GPU memory requirements (e.g. gpu_ram=2G). Thus e.g.

  qsub -q 'gpu*' -l gpu=1,gpu_cc_min3.5=1,gpu_ram=2G

• If you need more than one GPU card (on a single machine), always require as many CPU cores (-pe smp X) as many GPU cards you need. E.g.

  qsub -q 'gpu*' -l gpu=4,gpu_cc_min3.5=1,gpu_ram=7G -pe smp 4

• For interactive jobs, you can use qrsh, but make sure to end your job as soon as you don't need the GPU (so don't use qrsh for long training). Warning: -pty yes bash -l is necessary, otherwise the variable $CUDA_VISIBLE_DEVICES will not be set correctly. E.g.

  qrsh -q 'gpu*' -l gpu=1,gpu_ram=2G -pty yes bash -l

  In general: don't reserve a GPU (as described above) without actually using it for longer time. (E.g. try separating steps which need GPU and steps which do not and execute those separately on our GPU resp. CPU cluster.) Ondřej Bojar has a script /home/bojar/tools/servers/watch_gpus for watching reserved but unused GPU on most machines which will e-mail you, but don't rely on it only.

• Note that the dll machines have typically 10 cards, but “just” 250 GB RAM (DLL7 has only 128 GB). So the expected (maximal) mem_free requirement for jobs with 1 GPU is 25GB. If your 1-GPU job takes e.g. 80 GB and you submit three such jobs on the same machine, you have effectively blocked the whole machine and seven GPUs remain unused. If you really need to submit more high-memory jobs, send each on a different machine.
• If you know an approximate runtime of your job, please specify it with -l s_rt=hh:mm:ss - this is a soft constraint so your job won't be killed if it runs longer than specified. It will help SGE to better schedule the jobs, especially multi-gpu reservations (see qconf -ssconf).

Multiple versions of cuda can be accessed in /opt/cuda.

You need to set library path from your ~/.bashrc:

CUDA_version=10.1
CUDNN_version=7.6
CUDA_DIR_OPT=/opt/cuda/$CUDA_version
if [ -d "$CUDA_DIR_OPT" ] ; then
  CUDA_DIR=$CUDA_DIR_OPT
  export CUDA_HOME=$CUDA_DIR
  export THEANO_FLAGS="cuda.root=$CUDA_HOME,device=gpu,floatX=float32"
  export PATH=$PATH:$CUDA_DIR/bin
  export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$CUDA_DIR/cudnn/$CUDNN_version/lib64:$CUDA_DIR/lib64
  export CPATH=$CUDA_DIR/cudnn/$CUDNN_version/include:$CPATH
fi

• When not using Theano, just Tensorflow this can be simplified to export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/opt/cuda/10.1/lib64:/opt/cuda/10.1/cudnn/7.6/lib64.
• Note that the cudnn library is compiled for specific version of cuda. If you need specific version of cudnn, you can look in /opt/cuda/$CUDA_version/cudnn/ whether it is available for a given $CUDA_version.

Many people at UFAL use TensorFlow. To start using it it is recommended to create a Python virtual environment (or use Anaconda for it). Into the environment you must place TensorFlow. The TF is either in CPU or GPU version.

pip install tensorflow
pip install tensorflow-gpu

You can use prepared environment by adding into your ~/.bashrc

export PATH=/a/merkur3/kocmanek/ANACONDA/bin:$PATH

And then you can activate your environment:

source activate tf18
source activate tf18cpu

This environment have TensorFlow 1.8.0 and all necessary requirements for NeuralMonkey.

If you want to use pytorch, there is a ready-made environment in

/home/hajicj/anaconda3/envs/pytorch/bin

It does rely on the CUDA and CuDNN setup above.

As an alternative to qsub, you can use /home/bojar/tools/shell/qsubmit

qsubmit --gpumem=2G --queue="gpu-ms.q" WHAT_SHOULD_BE_RUN

It is recommended to use priority lower than the default -100 if you are not rushing for the results and don't need to leap over your colleagues jobs. Please, do not use priority between -99 to 0 for jobs taking longer than a few hours, unless it is absolutely necessary for your work.

lspci
  # is any such hardware there?
nvidia-smi
  # more details, incl. running processes on the GPU
  # nvidia-* are typically located in /usr/bin
watch nvidia-smi
  # For monitoring GPU activity in a separate terminal (thanks to Jindrich Libovicky for this!)
nvcc --version
  # this should tell CUDA version
  # nvcc is typically installed in /usr/local/cuda/bin/
theano-test
  # dela to vubec neco uzitecneho? :-)
  # theano-* are typically located in /usr/local/bin/
/usr/local/cuda/samples/1_Utilities/deviceQuery/deviceQuery
  # shows CUDA capability etc.
ssh dll1; ~popel/bin/gpu_allocations
  # who occupies which card on a given machine
  

The variable CUDA_VISIBLE_DEVICES constrains tensorflow and other toolkits to compute only on the selected GPUs. Do not set this variable yourself (unless debugging SGE), it is set for you automatically by SGE if you ask for some GPUs (see above).

To list available devices, use:

/opt/cuda/samples/1_Utilities/deviceQuery/deviceQuery | grep ^Device

* 980 vs 1080 vs Titan X (not the Titan Z we have)

In the following table is the experiment conducted by Tom Kocmi. You can replicate experiment: /a/merkur3/kocmanek/Projects/GPUBenchmark (you will need to prepare environment of TensorFlow11 or use my ANACONDA). The benchmark uses 2GB model of seq2seq machine translation in Neural Monkey (De > EN). If not specified, the benchmark had an access only to one GPU.

The previous benchmark only compares the speed of processing units within the GPUs and do not take into account the size of memory. Therefore I have conducted another benchmark, this time for each graphic card I have increased the batch size as much as possible so the model still could fit into the GPU (the previous benchmark model had batch size 20). This way the results should be more representative of the power for each GPU.

• CUDA Supported GPUs (Wikipedia table of Compute Capability)

GPU specs for those GPUs we have:

• Tesla K40c

There is an easy way to get one high-end GPU: ask NVIDIA for an Academic Hardware Grant. All it takes is writing a short grant application (at most ~2 hrs of work from scratch; if you have a GAUK, ~15 minutes of copy-pasting). Due to the GPU housing issues (mainly rack space and cooling), Milan F. said we should request the Tesla-line cards (2017; check with Milan about this issue). If you want to have a look at an application, feel free to ask at hajicj@ufal.mff.cuni.cz :)

Take care, however, to coordinate the grant applications a little, so that not too many arrive from UFAL within a short time: these grants are explicitly not intended to build GPU clusters, they are “seeding” grants meant for researchers to try out GPUs (and fall in love with them, and buy a cluster later). If you are planning to submit the hardware grant, have submitted one, or have already been awarded one, please add yourself here.

Known NVIDIA Academic Hardware Grants:

• Ondřej Plátek - granted (2015)
• Jan Hajič jr. - granted (early 2016)