PowerAI Deep Dive ( key points )
Download as PPTX, PDF2 likes101 views
The document provides an overview of PowerAI, IBM's set of libraries for developing machine learning and deep learning applications. It discusses what PowerAI is, its hardware requirements, the differences between CPUs and GPUs for machine learning, how to use PowerAI components like TensorFlow and Theano, and tuning recommendations for PowerAI performance.
PowerAI Deep Dive ( key points )
- 1. 1 IBM Systems Technical Events | ibm.com/training/events © Copyright IBM Corporation 2017. Technical University/Symposia materials may not be reproduced in whole or in part without the prior written permission of IBM. Paulo Sergio Lemes Queiroz Systems Consultant IBM PowerAI Deep Dive
- 2. 2 IBM Systems Technical Events | ibm.com/training/events © Copyright IBM Corporation 2017. Technical University/Symposia materials may not be reproduced in whole or in part without the prior written permission of IBM. • What is PowerAI • PowerAI Components Hardware requirements CPU VS GPU Volta / Tensors • Using PowerAI Components • Extending PowerAI Session objectives
- 3. 3 IBM Systems Technical Events | ibm.com/training/events © Copyright IBM Corporation 2017. Technical University/Symposia materials may not be reproduced in whole or in part without the prior written permission of IBM. What is PowerAI • Set of support Libraries to develop machine learning and Deep learning applications
- 4. 4 IBM Systems Technical Events | ibm.com/training/events © Copyright IBM Corporation 2017. Technical University/Symposia materials may not be reproduced in whole or in part without the prior written permission of IBM. Hardware Requirements Minimal requirements for accelerated Machine Learning is: - Any Power Server with a Nvidia GPU IBM Power System S824L (with NVIDIA technology) IBM Power System S822LC for High Performance Computing IBM Power System S822LC for Big Data IBM Power System S821LC Key start point is have the GPU for acceleration Nvlink enable even further acceleration Nvlink usage is transparent
- 5. 5 IBM Systems Technical Events | ibm.com/training/events © Copyright IBM Corporation 2017. Technical University/Symposia materials may not be reproduced in whole or in part without the prior written permission of IBM. CPU vs GPU CPU is a generic core, wich has access to the system main memmory in order to do all kind of tasks GPU is composed by thousands of specialized cores that handles mathematical operations GPUs are specialized on Tensor / Matrix / Vector / Float point parallel operations Sequential operations are slower on GPUS
- 6. 6 IBM Systems Technical Events | ibm.com/training/events © Copyright IBM Corporation 2017. Technical University/Symposia materials may not be reproduced in whole or in part without the prior written permission of IBM. GPU limitations GPU doesn't has access to the system main memmory, all data must be copied to the GPU and copied back when processing is done ( batch like operations ) Non SIMD ( Same Instruction Multiple Datapoints ) operations are not THAT fast
- 7. 7 IBM Systems Technical Events | ibm.com/training/events © Copyright IBM Corporation 2017. Technical University/Symposia materials may not be reproduced in whole or in part without the prior written permission of IBM. Volta / Tensors Volta GPU will have dedicated Tensor function units, which will allow processing of Neural Networks to run even faster However this will require even bigger bandwidth ( and lower lattency ) between the main memory and the GPU
- 8. 8 IBM Systems Technical Events | ibm.com/training/events © Copyright IBM Corporation 2017. Technical University/Symposia materials may not be reproduced in whole or in part without the prior written permission of IBM. Using PowerAI Components PowerAI, by default, require Ubunto 16.04 and it's install files can be located at: https://public.dhe.ibm.com/software/server/POWER/Linux/mldl/ubuntu/README.html The following components come along at the free package: caffe-bvlc - Berkeley Vision and Learning Center (BVLC) upstream Caffe, v1.0.0 caffe-ibm - IBM Optimized version of BVLC Caffe, v1.0.0 caffe-nv - NVIDIA fork of Caffe, v0.15.14 chainer - Chainer, v1.23.0 digits - DIGITS, v5.0.0 tensorflow - Google TensorFlow, v1.1.0 ddl-tensorflow - Distributed Deep Learning custom operator for TensorFlow theano - Theano, v0.9.0 torch - Torch, v7 It's important to highlight that PowerAI isn't limited by these pre-compiled libraries or versions As on any Ubuntu machine, the user can compile any software
- 9. 9 IBM Systems Technical Events | ibm.com/training/events © Copyright IBM Corporation 2017. Technical University/Symposia materials may not be reproduced in whole or in part without the prior written permission of IBM. Using PowerAI Components All software components of PowerAI are installed at /opt/DL root@pq-s824l-kvm:/opt/DL# ls -l total 0 drwxr-xr-x 4 root root 29 Jun 23 16:08 bazel lrwxrwxrwx 1 root root 23 Jun 23 16:39 caffe -> /etc/alternatives/caffe drwxr-xr-x 13 root root 151 Jun 23 16:38 caffe-bvlc drwxr-xr-x 13 root root 151 Jun 23 16:38 caffe-ibm drwxr-xr-x 13 root root 151 Jun 23 16:38 caffe-nv drwxr-xr-x 7 root root 68 Jun 23 16:38 chainer drwxr-xr-x 4 root root 28 Aug 8 14:21 ddl drwxr-xr-x 6 root root 55 Aug 8 14:21 ddl-tensorflow drwxr-xr-x 8 root root 210 Aug 6 16:47 digits drwxr-xr-x 7 root root 67 Jun 23 16:38 nccl drwxr-xr-x 6 root root 54 Jun 23 16:38 openblas drwxr-xr-x 3 root root 44 Aug 8 14:19 repo drwxr-xr-x 5 root root 54 Jun 23 16:10 tensorflow drwxr-xr-x 6 root root 54 Jun 23 16:38 theano drwxr-xr-x 9 root root 94 Jun 23 16:38 torch root@pq-s824l-kvm:/opt/DL# pwd /opt/DL root@pq-s824l-kvm:/opt/DL#
- 10. 10 IBM Systems Technical Events | ibm.com/training/events © Copyright IBM Corporation 2017. Technical University/Symposia materials may not be reproduced in whole or in part without the prior written permission of IBM. Using PowerAI Components By default these packages are not in the system PATH, however to enable those to exist at the PATH a few helper scripts can be used: . /opt/DL/tensorflow/bin/tensorflow-activate . /opt/DL/theano/bin/theano-activate export PATH="${PATH}:/opt/DL/bazel/bin"
- 11. 11 IBM Systems Technical Events | ibm.com/training/events © Copyright IBM Corporation 2017. Technical University/Symposia materials may not be reproduced in whole or in part without the prior written permission of IBM. Tuning recommendation Change the CPU govenor to performance, in order to avoid performance fluctuations for i in ./devices/system/cpu/cpufreq/policy* ; do echo performance > $i/scaling_governor ; done cpupower -c all frequency-set -g performance Enable Persistent memory mode for the GPU: nvidia-smi -pm ENABLED
- 12. 22 IBM Systems Technical Events | ibm.com/training/events © Copyright IBM Corporation 2017. Technical University/Symposia materials may not be reproduced in whole or in part without the prior written permission of IBM. Your feedback about this session is very important to us. Submit a survey at: ibmtechu.com
- 13. 23 IBM Systems Technical Events | ibm.com/training/events © Copyright IBM Corporation 2017. Technical University/Symposia materials may not be reproduced in whole or in part without the prior written permission of IBM. Continue the conversation view event highlights talk to tech experts connect with attendees read training articles IBM Systems Technical Events LinkedIn community Join today bit.ly/IBMTechUconnect
- 14. 24 IBM Systems Technical Events | ibm.com/training/events © Copyright IBM Corporation 2017. Technical University/Symposia materials may not be reproduced in whole or in part without the prior written permission of IBM. ibm.com/training provides a comprehensive portfolio of skills and career accelerators that are designed to meet all your training needs. If you can’t find the training that is right for you with our Global Training Providers, we can help. Contact IBM Training at dpmc@us.ibm.com Continue growing your IBM skills