Differences
This shows you the differences between two versions of the page.
Both sides previous revision Previous revision Next revision | Previous revision Next revision Both sides next revision | ||
gpu [2016/12/20 12:17] kocmanek |
gpu [2018/11/12 10:36] vodrazka [Servers with GPU units] |
||
---|---|---|---|
Line 4: | Line 4: | ||
===== Servers with GPU units ===== | ===== Servers with GPU units ===== | ||
+ | GPU cluster '' | ||
+ | | machine | GPU type | GPU driver version | [[https:// | ||
+ | | dll1 | GeForce GTX 1080 | 396.24 | 6.1 | 8 | 8 | 249 | yes | | ||
+ | | dll2 (out of order) | GeForce GTX 1080 | 396.24 | 6.1 | 8 | 8 | 249 | yes | | ||
+ | | dll3 | GeForce GTX 1080 Ti | 396.24 | 6.1 | 10 | 11 | 249 | yes | | ||
+ | | dll4 | GeForce GTX 1080 Ti | 396.24 | 6.1 | 10 | 11 | 249 | yes | | ||
+ | | dll5 | GeForce GTX 1080 Ti | 396.24 | 6.1 | 10 | 11 | 249 | yes | | ||
+ | | dll6 | GeForce GTX 1080 Ti | 396.24 | 6.1 | 10 | 11 | 123 | yes | | ||
+ | | kronos | GeForce GTX 1080 Ti | 396.24 | 6.1 | 1 | 11 | 123 | yes | | ||
+ | | titan1 | GeForce GTX 1080 | 396.24 | 6.1 | 1 | 8 | 30 | yes | | ||
+ | | titan2 | Tesla K40c | 396.24 | 3.5 | 1 | 11 | 30 | yes | | ||
+ | | twister1 | Tesla K40c | 396.24 | 3.5 | 1 | 11 | 45 | no | | ||
+ | | twister2 | Tesla K40c | 396.24 | 3.5 | 1 | 11 | 45 | no | | ||
- | | machine | GPU; [[https:// | + | Desktop machines: |
- | | titan-gpu | + | | machine |
- | | twister1 | + | | victoria; arc | GeForce GT 630 | cc3.0 | 1 | 2 GB | desktop machine | |
- | | twister2 | + | | athena |
- | | kronos-dev | Tesla K40c; cc3.5 | 1 | 12 GB | | | + | |
- | | iridium | + | |
- | | victoria | + | |
- | | arc | GeForce GT 630; cc3.0 | 1 | 2 GB | + | |
- | | athena | + | |
- | not used at the moment: GeForce GTX 570 (from twister2) | + | Not used at the moment: GeForce GTX 570 (from twister2) |
- | All machines have CUDA8.0 and should support both Theano and TensorFlow. | + | Multiple versions of CUDA library are accessible on each machine together with cudnn. Theano and TensorFlow |
- | Summary of future plans: | + | [[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# |
- | * Current Troja servers won't get any GPUs (the only option would be [[http://www.czc.cz/hp-quadro-k1200-4gb/171662/produkt? | + | |
- | * The old Quadro K2000 we have is a much more low end piece, so we can't test is in Troja. | + | |
- | * There is MetaCentrum which also has GPUs, so testing can be done there. | + | |
- | * 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. | + | |
- | * 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. | + | |
- | === Individual acquisitions: NVIDIA Academic Hardware Grants | + | ===== Rules ===== |
+ | * First, read [[internal:Linux network]] and [[: | ||
+ | * 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 '' | ||
+ | * Always specify the number of GPU cards (e.g. '' | ||
+ | * If you need more than one GPU card (on a single machine), always require as many CPU cores ('' | ||
+ | * For interactive jobs, you can use '' | ||
+ | * Note that the dll machines have typically 10 cards, but " | ||
- | There is an easy way to get one high-end GPU: [[https:// | + | ===== How to use cluster ===== |
- | 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 " | + | ==== Set-up CUDA and CUDNN ==== |
- | Known NVIDIA Academic Hardware Grants: | + | Multiple versions of '' |
- | * Ondřej Plátek - granted (2015) | + | You need to set library path from your '' |
- | * Jan Hajič jr. - granted (early 2016) | + | |
- | * Jindra Helcl - planning | + | |
+ | CUDNN_version=7.0 | ||
+ | CUDA_version=9.0 | ||
+ | CUDA_DIR_OPT=/ | ||
+ | if [ -d " | ||
+ | CUDA_DIR=$CUDA_DIR_OPT | ||
+ | export CUDA_HOME=$CUDA_DIR | ||
+ | export THEANO_FLAGS=" | ||
+ | export PATH=$PATH: | ||
+ | export LD_LIBRARY_PATH=$LD_LIBRARY_PATH: | ||
+ | export CPATH=$CUDA_DIR/ | ||
+ | fi | ||
- | | + | |
+ | * Note that the '' | ||
- | ===== Performance tests ===== | ||
- | * [[http:// | + | ==== TensorFlow Environment ==== |
- | In the following table is the experiment conducted by Tom Kocmi. You can replicate experiment: / | + | Many people at UFAL use TensorFlow. To start using it it is recommended |
- | I am preparing department-wide benchmark, but meanwhile the results with same setting: | + | pip install tensorflow |
- | * Athena (GTX 1080) - 2 hodiny 32 minut | + | pip install tensorflow-gpu |
- | * Twister (Tesla K40c) - 6 hodin 46 minut | + | |
+ | You can use prepared environment by adding into your ~/.bashrc | ||
- | | machine | GPU; [[https://en.wikipedia.org/wiki/CUDA# | + | export PATH=/a/merkur3/kocmanek/ANACONDA/ |
- | | athena | + | |
- | | titan-gpu | + | |
- | | twister1 | + | |
- | | twister2 | + | |
- | | kronos-dev | Tesla K40c; cc3.5 | CUDA8.0 | | | + | |
- | | iridium | + | |
- | | victoria | + | |
- | | arc | GeForce GT 630; cc3.0 | CUDA8.0 | | | + | |
- | | lucifer4 | + | |
- | | ? | - | 16x CPU | | | + | |
+ | And then you can activate your environment: | ||
+ | source activate tf18 | ||
+ | source activate tf18cpu | ||
- | ===== Installed toolkits ===== | + | This environment have TensorFlow 1.8.0 and all necessary requirements for NeuralMonkey. |
- | //This should mention where each interesting toolkit lives (on a particular machine).// | + | ==== Pytorch Environment ==== |
- | ==== TensorFlow ==== | + | If you want to use pytorch, there is a ready-made environment in |
- | [[https://redmine.ms.mff.cuni.cz/projects/mmmt/repository/revisions/ | + | |
+ | |||
+ | It does rely on the CUDA and CuDNN setup above. | ||
- | OP: I created [[https:// | + | ==== Using cluster ==== |
- | === Select GPU device === | + | As an alternative to '' |
- | Use variable CUDA_VISIBLE_DEVICES to constrain tensorflow to compute only on the selected one. For the use of first GPU use: | + | qsubmit --gpumem=2G --queue=" |
- | < | + | |
- | + | 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. | |
- | To list available devices, | + | ==== Basic commands ==== |
- | < | + | |
- | + | ||
- | ===== Basic commands | + | |
lspci | lspci | ||
Line 100: | Line 111: | ||
/ | / | ||
# shows CUDA capability etc. | # shows CUDA capability etc. | ||
+ | ssh dll1; ~popel/ | ||
+ | # who occupies which card on a given machine | ||
+ | | ||
+ | === Select GPU device === | ||
+ | |||
+ | 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: | ||
+ | / | ||
+ | |||
+ | ===== Performance tests ===== | ||
+ | |||
+ | * [[http:// | ||
+ | |||
+ | In the following table is the experiment conducted by Tom Kocmi. You can replicate experiment: / | ||
+ | |||
+ | | machine | Setup; CPU/GPU; [[https:// | ||
+ | | athena | ||
+ | | dll2 | GeForce GTX 1080; cc6.1 | | ||
+ | | titan | GeForce GTX 1080 Ti | | ||
+ | | dll1 | (2 GPU) GeForce GTX 1080; cc6.1 | | ||
+ | | twister2 | ||
+ | | dll2 | GeForce GTX 1080; cc6.1 | | ||
+ | | titan-gpu | ||
+ | | kronos-dev | Tesla K40c; cc3.5 | | ||
+ | | twister2 | ||
+ | | twister1 | ||
+ | | helena1 | ||
+ | | belzebub | ||
+ | | iridium | ||
+ | | helena7 | ||
+ | | arc | GeForce GT 630; cc3.0 | 103:42:30 | (approximated after 66 hours) | | ||
+ | | lucifer4 | ||
+ | |||
+ | |||
+ | === 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 | ||
+ | | 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 | ||
+ | | 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 ===== | ===== Links ===== | ||
Line 108: | Line 164: | ||
GPU specs for those GPUs we have: | GPU specs for those GPUs we have: | ||
* [[http:// | * [[http:// | ||
+ | |||
+ | ==== Individual acquisitions: | ||
+ | |||
+ | There is an easy way to get one high-end GPU: [[https:// | ||
+ | |||
+ | 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 " | ||
+ | |||
+ | Known NVIDIA Academic Hardware Grants: | ||
+ | |||
+ | * Ondřej Plátek - granted (2015) | ||
+ | * Jan Hajič jr. - granted (early 2016) | ||
+ | |||
+ | |||
+ |