[ Skip to the content ]

Institute of Formal and Applied Linguistics Wiki


[ Back to the navigation ]

Differences

This shows you the differences between two versions of the page.

Link to this comparison view

Both sides previous revision Previous revision
Next revision
Previous revision
Next revision Both sides next revision
gpu [2017/10/17 16:38]
popel [Rules]
gpu [2021/07/02 16:50]
ptacek
Line 4: Line 4:
  
 ===== Servers with GPU units ===== ===== Servers with GPU units =====
-GPU cluster ''gpu.q'' at Malá Strana:+GPU cluster ''gpu-ms.q'' at Malá Strana:
  
-| machine                    | GPU type | [[https://en.wikipedia.org/wiki/CUDA#GPUs_supported|cc]] | GPUs | GPU RAM | Comment +| machine | GPU type | GPU driver version | [[https://en.wikipedia.org/wiki/CUDA#GPUs_supported|cc]] | GPU cnt | GPU RAM (GB) machine RAM (GB)
-iridium                    | Quadro K2000        |  cc3.0 |   1|   2 GB |  | +dll3 |  GeForce GTX 1080 Ti |   455.23.05 |  6.1 |  10 |  11 |  248.8 
-titan-gpu                  | GeForce GTX Titan Z |  cc3.  2  6 GB |  | +dll4  GeForce GTX 1080 Ti |   455.23.05 |  6. 10  11 |  248.8 
-twister1; twister2; kronos | Tesla K40c          |  cc3.5 |   1|  12 GB |  | +dll5 |  GeForce GTX 1080 Ti |   455.23.05 |  6.1 |  10 |  11 |  248.8 
-dll1; dll2                 | GeForce GTX 1080    |  cc6.1 |   8|   GB |  | +dll6  GeForce GTX 1080 Ti |   455.23.05 |  6.1 |  8 |  11 |  248.8 | 
-titan                      | GeForce GTX 1080    |  cc6.|   1|   GB |  | +| dll7 |  GeForce RTX 2080 Ti |   455.23.05 |  7.5 |  |  11 |  248.8 | 
-dll3; dll4; dll5           GeForce GTX 1080 Ti |  cc6.1 |  10|  11 GB dll3 has only 9 GPUs since 2017/07 +| dll9 |  GeForce RTX 3090 |   455.23.05 |  8.6 |  4 |  25 |  183.0 
-dll6                       GeForce GTX 1080 Ti |  cc6.1 |   3|  11 GB |  |+dll10  GeForce RTX 3090 |   455.23.05 |  8.6 |  4 |  25 |  183.0 | 
 + 
 +GPU cluster ''gpu-troja.q'' at Troja: 
 + 
 +| machine | GPU type | GPU driver version | [[https://en.wikipedia.org/wiki/CUDA#GPUs_supported|cc]] | GPU cnt | GPU RAM (GB) | machine RAM (GB)| 
 +| tdll1 |  Quadro P5000 |   455.23.05 |  6.|  8 |  17 |  245.0 | 
 +| tdll2 |  Quadro P5000 |   455.23.05 |  6.1 |  8 |  17 |  245.0 
 +tdll3  Quadro P5000 |   455.23.05 |  6.1 |  |  17  245.0 
 +tdll4  Quadro P5000 |   455.23.05 |  6.1 |  8 |  17 |  245.0 | 
 +| tdll5 |  Quadro P5000 |   455.23.05 |  6.1 |  8 |  17 |  245.0 |
  
 Desktop machines: Desktop machines:
Line 20: Line 29:
 | athena                     | GeForce GTX 1080 | cc6.1 |  1 |  8 GB | Tom's desktop machine | | athena                     | GeForce GTX 1080 | cc6.1 |  1 |  8 GB | Tom's desktop machine |
  
-Not used at the moment: GeForce GTX 570 (from twister2) +Multiple versions of CUDA library are accessible on each machine together with cudnn. Theano and TensorFlow is supported. 
-All machines have CUDA8.0 and should support both Theano and TensorFlow.+ 
 +[[http://ufaladm2.ufal.hide.ms.mff.cuni.cz/munin/ufal.hide.ms.mff.cuni.cz/lrc-master.ufal.hide.ms.mff.cuni.cz/index.html#dll|GPU usage rolling graphs]] 
  
 ===== Rules ===== ===== Rules =====
   * First, read [[internal:Linux network]] and [[:Grid]].   * First, read [[internal: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_vmem=12G''+  * 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''
-  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 you GPU memory requirements (e.g. ''gpu_ram=2G''). Thus e.g. <code>qsub -q gpu.q -l gpu=1,gpu_cc_min3.5=1,gpu_ram=2G</code> +    **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''
-  * If you need more than one GPU card, always require as many CPU cores as many GPU cards you need. E.g. <code>qsub -q gpu.q -l gpu=4,gpu_cc_min3.5=1,gpu_ram=7G -pe smp 4</code> +  * Always specify the number of GPU cards (e.g. ''gpu=1'') and your GPU memory requirements (e.g. ''gpu_ram=2G''). Thus e.g. <code>qsub -q 'gpu*' -l gpu=1,gpu_ram=2G</code> 
-  * 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). E.g. <code>qrsh -q gpu.q -l gpu=1,gpu_ram=2G -pty yes bash</code>+  * If you need more than one GPU card (on a single machine), always require at least as many CPU cores (''-pe smp X''as many GPU cards you need. E.g. <code>qsub -q 'gpu*' -l gpu=4,gpu_ram=7G -pe smp 4</code> 
 +  * 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. <code>qrsh -q 'gpu*' -l gpu=1,gpu_ram=2G -pty yes bash -l</code>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''). 
  
 ===== How to use cluster ===== ===== How to use cluster =====
Line 34: Line 49:
 ==== Set-up CUDA and CUDNN ==== ==== Set-up CUDA and CUDNN ====
  
-You can add following command into your ~/.bashrc+Multiple versions of ''cuda'' can be accessed in ''/opt/cuda''
  
-  CUDNN_version=6.0 +You need to set library path from your ''~/.bashrc'': 
-  CUDA_version=8.0 + 
-  CUDA_DIR_OPT=/opt/cuda-$CUDA_version+  CUDA_version=10.
 +  CUDNN_version=7.6 
 +  CUDA_DIR_OPT=/opt/cuda/$CUDA_version
   if [ -d "$CUDA_DIR_OPT" ] ; then   if [ -d "$CUDA_DIR_OPT" ] ; then
     CUDA_DIR=$CUDA_DIR_OPT     CUDA_DIR=$CUDA_DIR_OPT
Line 47: Line 64:
     export CPATH=$CUDA_DIR/cudnn/$CUDNN_version/include:$CPATH     export CPATH=$CUDA_DIR/cudnn/$CUDNN_version/include:$CPATH
   fi   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''.
 +
  
 ==== TensorFlow Environment ==== ==== TensorFlow Environment ====
  
-Majority people at UFAL use TensorFlow. To start using it you need to create python virtual environment (virtualenv or use Anaconda for it). Into the environment you must place TensorFlow. The TF is either in CPU or GPU version.+Many people at UFAL use TensorFlow. To start using it it is recommended to create a [[python|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
Line 61: Line 82:
 And then you can activate your environment: And then you can activate your environment:
  
-  source activate tf1 +  source activate tf18 
-  source activate tf1cpu+  source activate tf18cpu
  
-This environment have TensorFlow 1.0 and all necessary requirements for NeuralMonkey.+This environment have TensorFlow 1.8.0 and all necessary requirements for NeuralMonkey.
  
 ==== Pytorch Environment ==== ==== Pytorch Environment ====
Line 76: Line 97:
 ==== Using cluster ==== ==== Using cluster ====
  
-Rule number onealways use the GPU queue (never log in machine by ssh). Always use qsub or qsubmit with proper arguments. +As an alternative to ''qsub'', you can use /home/bojar/tools/shell/qsubmit
- +
-For testing and using the cluster interactively you can use qrsh (this should not be used for long running experiments since the console is not closed on the end of the experiment). Following command will assign you a GPU and creates interactive console. +
- +
-  qrsh -q gpu.q -l gpu=1,gpu_ram=2G -pty yes bash +
-   +
-For running experiments you must use qsub command: +
- +
-  qsub -q gpu.q -l gpu=1,gpu_cc_min3.5=1,gpu_ram=2G WHAT_SHOULD_BE_RUN +
-   +
-Cleaner way to use cluster is with /home/bojar/tools/shell/qsubmit+
  
-  qsubmit --gpumem=2G --queue="gpu.q" WHAT_SHOULD_BE_RUN+  qsubmit --gpumem=2G --queue="gpu-ms.q" WHAT_SHOULD_BE_RUN
      
-It is recommended to use priority -100 if you are not rushing for the results and don't need to leap over your colleagues jobs.+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.
 ==== Basic commands ==== ==== Basic commands ====
  
Line 100: Line 111:
   watch nvidia-smi   watch nvidia-smi
     # For monitoring GPU activity in a separate terminal (thanks to Jindrich Libovicky for this!)     # For monitoring GPU activity in a separate terminal (thanks to Jindrich Libovicky for this!)
 +    # You can also use nvidia-smi -l TIME
   nvcc --version   nvcc --version
     # this should tell CUDA version     # this should tell CUDA version
Line 129: Line 141:
 | titan      | GeForce GTX 1080 Ti                |   10:45:11 | (new result with correct CUDA version) | | titan      | GeForce GTX 1080 Ti                |   10:45:11 | (new result with correct CUDA version) |
 | dll1       | (2 GPU) GeForce GTX 1080; cc6.1    |   12:34:34 | Probably only one GPU was used | | dll1       | (2 GPU) GeForce GTX 1080; cc6.1    |   12:34:34 | Probably only one GPU was used |
 +| twister2   | Quadro P5000                         13:19:00 |  |
 | dll2       | GeForce GTX 1080; cc6.1            |   13:01:05 | Only one GPU was used | | dll2       | GeForce GTX 1080; cc6.1            |   13:01:05 | Only one GPU was used |
 | titan-gpu  | (2 GPU) GeForce GTX Titan Z; cc3.5 |   16:05:24 | Probably only one GPU was used | | titan-gpu  | (2 GPU) GeForce GTX Titan Z; cc3.5 |   16:05:24 | Probably only one GPU was used |
Line 146: Line 159:
 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. 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.
  
-| GPU; Cuda capability       | GPU RAM |  Walltime | Batch size | Machine +| GPU; Cuda capability       | GPU RAM |  Walltime | Batch size | Machine   | 
-| Tesla K40c; cc3.5          |   12 GB |                  |  +| GeForce GTX 1080 Ti; cc6.1 |   11 GB |  00:55:56 |       2300 | dll5      
-| GeForce GTX 1080 Ti; cc6.1 |   11 GB |  00:55:56 |       2300 | dll5 | +| GeForce GTX 1080; cc6.1    |    8 GB |  01:10:57 |       1700 | dll1      | 
-| GeForce GTX 1080; cc6.1    |    8 GB |  01:10:57 |       1700 | dll1 |+| Quadro P5000                 16 GB |  01:17:00 |       3400 | twister2  |
 | GeForce GTX Titan Z; cc3.5 |    6 GB |  02:20:47 |       1100 | titan-gpu | | GeForce GTX Titan Z; cc3.5 |    6 GB |  02:20:47 |       1100 | titan-gpu |
-| Quadro K2000; cc3.0        |    2 GB |  28:15:26 |         50 | iridium |+| Quadro K2000; cc3.0        |    2 GB |  28:15:26 |         50 | iridium   |
  
 ===== Links ===== ===== Links =====
Line 163: Line 176:
 ==== Individual acquisitions: NVIDIA Academic Hardware Grants ==== ==== Individual acquisitions: NVIDIA Academic Hardware Grants ====
  
-There is an easy way to get one high-end GPU: [[https://developer.nvidia.com/academic_gpu_seeding|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 :)+There is an easy way to get one high-end GPU: [[https://developer.nvidia.com/academic_gpu_seeding|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 (2017check 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. 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.

[ Back to the navigation ] [ Back to the content ]