Differences

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

--- gpu [2017/05/16 10:48]
kocmanek [Performance tests]
+++ gpu [2024/10/02 15:21] (current)
popel
@@ Line 3: / Line 3: @@
 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.
-===== Servers with GPU units =====
+**TODO: IN 2022 MOVING FROM SGE TO SLURM** (see [[slurm|slurm guidelines]]) -- **commands like ''qsub'' and ''qstat'' will stop working!**
-| machine    | GPU; [[https://en.wikipedia.org/wiki/CUDA#Supported_GPUs|Capability]] [cc]  | cores | GPU RAM | Comment |
+**IN 2024: Newly, all the documentation is at a dedicated wiki https://ufal.mff.cuni.cz/lrc (you need to use ufal and [[internal:welcome-at-ufal#small-linguistic-password|small-linguistic password]] to access the wiki from outside of the UFAL network).***
-| titan      | GeForce GTX 1080 Ti; cc6.1 | 1 | 12 GB |        |
-| titan-gpu  | GeForce GTX Titan Z; cc3.5 | 2 | 6 GB each core |        |
-| twister1   | Tesla K40c; cc3.5          | 1 | 12 GB          |        |
-| twister2   | Tesla K40c; cc3.5          | 1 | 12 GB          |        |
-| kronos-dev | Tesla K40c; cc3.5          | 1 | 12 GB          |        |
-| iridium    | Quadro K2000; cc3.0        | 1 | 2 GB           |        |
-| victoria   | GeForce GT 630; cc3.0      | 1 | 2 GB           | Ondrej Bojar's desktop machine |
-| arc        | GeForce GT 630; cc3.0      | 1 | 2 GB           | Lucka's desktop machine |
-| athena     | GeForce GTX 1080; cc6.1    | 1 | 8 GB           | Tom's desktop machine |
-| dll1       | GeForce GTX 1080; cc6.1    | 8 | 8 GB each core |  |
-| dll2       | GeForce GTX 1080; cc6.1    | 8 | 8 GB each core |  |
-not used at the moment: GeForce GTX 570 (from twister2)
+===== Servers with GPU units =====
-All machines have CUDA8.0 and should support both Theano and TensorFlow.
+GPU cluster ''gpu-ms.q'' at Malá Strana:
-Summary of future plans:
+| machine | GPU type | GPU driver version | [[https://en.wikipedia.org/wiki/CUDA#GPUs_supported|cc]] | GPU cnt | GPU RAM (GB) | machine RAM (GB)|
-  * Current Troja servers won't get any GPUs (the only option would be [[http://www.czc.cz/hp-quadro-k1200-4gb/171662/produkt?ppcbee-adtext-variant=Produkt%3B+kategorie+%2B+cena%3B+Pobo%C4%8Dky&gclid=CKbKkbrWrswCFQUq0wodHDELCw|Quadro K1200 4GB]], horribly cost-inefficient)
+| dll1 |  Quadro RTX 5000 |   455.23.05 |  7.5 |  8 |  16 |  366.2 |
-  * The old Quadro K2000 we have is a much more low end piece, so we can't test is in Troja.
+| dll3 |  NVIDIA RTX A4000 |   510.73.08 |  8.6 |  10 |  16 |  248.8 |
-  * There is MetaCentrum which also has GPUs, so testing can be done there.
+| dll4 |  GeForce GTX 1080 Ti |   455.23.05 |  6.1 |  10 |  11 |  248.8 |
-  * It is impossible (wasteful in terms of space and forbidden by a dean regulation) to put non-rack machines to our servers rooms. So we won't be buying GeForce GTX 1080 (~20000CZK, out of stock now), for a non-rack machine since we most likely don't have any available.
+| dll5 |  GeForce GTX 1080 Ti |   455.23.05 |  6.1 |  10 |  11 |  248.8 |
-  * Yes, there are grant applications under review which include rack machines with GPUs, e.g. 5x2 or something like that; more will be known in 2017.
+| dll6 |  NVIDIA RTX A4000 |   510.73.08 |  8.6 |  8 |  16 |  248.8 |
+| dll7 |  NVIDIA RTX A4000 |   510.73.08 |  8.6 |  8 |  16 |  248.8 |
+| dll8 |  Quadro RTX 5000 |   455.23.05 |  7.5 |  8 |  16 |  366.2 |
+| dll9 |  GeForce RTX 3090 |   455.23.05 |  8.6 |  4 |  25 |  183.0 |
+| dll10 |  GeForce RTX 3090 |   455.23.05 |  8.6 |  4 |  25 |  183.0 |
+GPU cluster ''gpu-troja.q'' at Troja:
-=== Individual acquisitions: NVIDIA Academic Hardware Grants ==
+| 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 |  6.1 |  8 |  16 |  245.0 |
+| tdll2 |  Quadro P5000 |   455.23.05 |  6.1 |  8 |  16 |  245.0 |
+| tdll3 |  Quadro P5000 |   455.23.05 |  6.1 |  8 |  16 |  245.0 |
+| tdll4 |  Quadro P5000 |   455.23.05 |  6.1 |  8 |  16 |  245.0 |
+| tdll5 |  Quadro P5000 |   455.23.05 |  6.1 |  8 |  16 |  245.0 |
-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 :)
+Desktop machines:
+| machine                    | GPU type | [[https://en.wikipedia.org/wiki/CUDA#GPUs_supported|cc]] | GPUs | GPU RAM | Comment |
+| victoria; arc              | GeForce GT 630   | cc3.0 |  1 |  2 GB | desktop machine |
+| athena                     | GeForce GTX 1080 | cc6.1 |  1 |  8 GB | Tom's desktop machine |
-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.
+Multiple versions of CUDA library are accessible on each machine together with cudnn. Theano and TensorFlow is supported.
-Known NVIDIA Academic Hardware Grants:
+[[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]]
-  * Ondřej Plátek - granted (2015)
-  * Jan Hajič jr. - granted (early 2016)
-  * Jindra Helcl - planning to apply (fall 2016)
+===== Rules =====
+  * 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_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'') and your GPU memory requirements (e.g. ''gpu_ram=2G''). Thus e.g. <code>qsub -q 'gpu*' -l gpu=1,gpu_ram=2G</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 ''$CUDA_VISIBLE_DEVICES'' will not be set correctly. E.g. <code>qrsh -q 'gpu*' -l gpu=1,gpu_ram=2G -pty yes bash -l</code>In general: don't reserve a GPU (as described above) without actually using it for longer time. (E.g. try separating steps which need GPU and steps which do not and execute those separately on our GPU resp. CPU cluster.) Ondřej Bojar has a script /home/bojar/tools/servers/watch_gpus for watching reserved but unused GPU on most machines which will e-mail you, but don't rely on it only.
+  * Note that the dll machines have typically 10 cards, but "just" 250 GB RAM (DLL7 has only 128 GB). So the expected (maximal) ''mem_free'' requirement for jobs with 1 GPU is 25GB. If your 1-GPU job takes e.g. 80 GB and you submit three such jobs on the same machine, you have effectively blocked the whole machine and seven GPUs remain unused. If you really need to submit more high-memory jobs, send each on a different machine.
+  * If you know an approximate runtime of your job, please specify it with ''-l s_rt=hh:mm:ss'' - this is a soft constraint so your job won't be killed if it runs longer than specified. It will help SGE to better schedule the jobs, especially multi-gpu reservations (see ''qconf -ssconf'').
 ===== How to use cluster =====
-In this section will be explained how to use cluster properly.
+==== Set-up CUDA and CUDNN ====
+Multiple versions of ''cuda'' can be accessed in ''/opt/cuda''.
+You need to set library path from your ''~/.bashrc'':
+  CUDA_version=11.1
+  CUDNN_version=8.0
+  CUDA_DIR_OPT=/opt/cuda/$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/11.1/cudnn/8.0/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 ====
-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
@@ Line 60: / Line 88: @@
 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.
-==== Using cluster ====
+==== PyTorch Environment ====
-Rule number one, always use the GPU queue (never log in machine by ssh). Always use qsub or qsubmit with proper arguments.
+Install PyTorch following the instructions on https://pytorch.org.
-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.
+At the time of writing, the recommended setup for CUDA 11.1 (supported by all GPU nodes) is:
-  qrsh -q gpu.q -l gpu=1 -pty yes bash
+   pip3 install torch==1.10.0+cu111 torchvision==0.11.1+cu111 torchaudio==0.10.0+cu111 -f https://download.pytorch.org/whl/cu111/torch_stable.html
+==== Using cluster ====
-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
+As an alternative to ''qsub'', you can use /home/bojar/tools/shell/qsubmit
-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 ====
@@ Line 91: / Line 116: @@
   watch nvidia-smi
     # For monitoring GPU activity in a separate terminal (thanks to Jindrich Libovicky for this!)
+    # You can also use nvidia-smi -l TIME
   nvcc --version
     # this should tell CUDA version
@@ Line 99: / Line 125: @@
   /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
 === Select GPU device ===
-Use variable CUDA_VISIBLE_DEVICES to constrain tensorflow to compute only on the selected one. For the use of first GPU use (GPU queue do this for you):
+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).
-  export CUDA_VISIBLE_DEVICES=0
 To list available devices, use:
@@ Line 112: / Line 139: @@
 * [[http://www.trustedreviews.com/nvidia-geforce-gtx-1080-review-performance-benchmarks-and-conclusion-page-2| 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.
+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.
 | machine | Setup; CPU/GPU; [[https://en.wikipedia.org/wiki/CUDA#Supported_GPUs|Capability]] [cc] | Walltime | Note |
 | athena     | GeForce GTX 1080; cc6.1            |    9:55:58 | Tom's desktop  |
-| dll2       | (2 GPU) GeForce GTX 1080; cc6.1    |   10:19:40 | with CUDA_VISIBLE_DEVICES=0 |
+| dll2       | GeForce GTX 1080; cc6.1            |   10:19:40 |  |
-| titan      | GeForce GTX 1080 Ti                |   11:41:08 |  |
+| 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 |
-| dll2       | (2 GPU) GeForce GTX 1080; cc6.1    |   13:01:05 | 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 |
 | titan-gpu  | (2 GPU) GeForce GTX Titan Z; cc3.5 |   16:05:24 | Probably only one GPU was used |
 | kronos-dev | Tesla K40c; cc3.5                  |   22:41:01 |  |
@@ Line 130: / Line 158: @@
 | arc        | GeForce GT 630; cc3.0              |  103:42:30 | (approximated after 66 hours) |
 | lucifer4   | 8x cores CPU                       |  134:41:22 |  |
-| victoria   | GeForce GT 630; cc3.0              |        --- | never run, same GPU as Arc |
+=== Second Benchmark ===
+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   |
+| 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      |
+| Quadro P5000               |   16 GB |  01:17:00 |       3400 | twister2  |
+| 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   |
 ===== Links =====
@@ Line 141: / Line 178: @@
 GPU specs for those GPUs we have:
   * [[http://www.nvidia.com/content/PDF/kepler/Tesla-K40-Active-Board-Spec-BD-06949-001_v03.pdf|Tesla K40c]]
+==== 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 :)
+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)

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

Institute of Formal and Applied Linguistics Wiki

Differences