![© 2020 IBM Corporation
IBM Cognitive Systems
Enabling tensorflow Large Model Support
29
Keras API
from tensorflow.contrib.lms import LMSKerasCallback
lms_callback = LMSKerasCallback()
…
model.fit_generator(generator=training_gen,
callbacks=[lms_callback])
Estimator API
with tf.name_scope(‘gradientscope'):
optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.001)
train_op = optimizer.minimize(loss=loss,
global_step=tf.train.get_global_step())
from tensorflow.contrib.lms import LMSHook
lms_hook = LMSHook({‘gradientscope '})
mnist_classifier.train(input_fn=train_input_fn, steps=20000,
hooks=[logging_hook, lms_hook])](https://image.slidesharecdn.com/deeplearningusingcloudpakfordata-210713231201/85/Deeplearningusingcloudpakfordata-29-320.jpg)
![© 2020 IBM Corporation
IBM Cognitive Systems
Demo with Slurm
33
ssh bsc18651@plogin2.bsc.es #Slurm login node
# You should have a ~/data dir with the dataset downloaded or internet conection to download it
#Edit and include the following line in ~/.bashrc
export TMPDIR=/tmp/
# To pass all the variables, like activate ...., you may need to write a simple submission script: Run as “sbatch script.sh”
[bsc18651@p9login2 ~]$ cat script.sh
#!/bin/bash
#SBATCH -J test
#SBATCH -D .
#SBATCH -o test_%j.out
#SBATCH -e test_%j.err
#SBATCH -N 2
#SBATCH --ntasks-per-node=4
#SBATCH --gres="gpu:4"
#SBATCH --time=01:00:00
module purge
module load anaconda2 powerAI
source /opt/DL/ddl-tensorflow/bin/ddl-tensorflow-activate
export TMPDIR="/tmp/"
export DDL_OPTIONS="-mode b:4x2"
NODE_LIST=$(scontrol show hostname $SLURM_JOB_NODELIST | tr 'n' ',')
NODE_LIST=${NODE_LIST%?}
cd examples/mnist
ddlrun -n 8 -H $NODE_LIST python mnist-init.py --ddl_options="-mode b:4x2" --data_dir
/home/bsc18/bsc18651/examples/mnist/data
[bsc18651@p9login2 ~]$ sbatch script.sh
https://developer.ibm.com/linuxonpower/2018/05/01/improved-ease-use-ddl-powerai/](https://image.slidesharecdn.com/deeplearningusingcloudpakfordata-210713231201/85/Deeplearningusingcloudpakfordata-33-320.jpg)
Deeplearningusingcloudpakfordata
- 1. © 2020 IBM Corporation IBM Cognitive Systems Ander Ochoa – [email protected] Cognitive Systems Architect for SPGI OpenPOWER Foundation member https://es.linkedin.com/in/anderotxoa Deep learning using Cloud Pak for Data CINECA OpenPOWER Workshop
- 2. © 2020 IBM Corporation IBM Cognitive Systems Agenda §Introduction to Open POWER servers §Open POWER based Super Computers §Software Solutions for Open POWER accelerated servers §State of the Art SW Solutions leverage AI on POWER §Demo - Cloud Pak for Data
- 3. © 2020 IBM Corporation IBM Cognitive Systems Open POWER architecture Based Servers
- 4. © 2020 IBM Corporation IBM Cognitive Systems Current Mayor Processor Architectures PROPIETARY ARCHITECTURES X86 - Intel - AMD LICENCEABLE ARCHITECTURES ARM - Samsung - Apple - … OPEN ARCHITECTURES - RISC-V - OpenPOWER
- 5. © 2020 IBM Corporation IBM Cognitive Systems 5
- 6. © 2020 IBM Corporation IBM Cognitive Systems 6 OPENNESS: POWER9 Ecosystem
- 7. © 2020 IBM Corporation IBM Cognitive Systems POWER9 The CPU 7 POWER9 designed for data 2.6x2 Performance per core Memory per socket (8TB / socket) POWER9 vs x86 Xeon SP (1) 2X performance per core is based on IBM Internal measurements as of 2/28/18 on various system configuration and workload environments including (1) Enterprise Database (2.22X per core): 20c L922 (2x10-core/2.9 GHz/256 GB memory): 1,039,365 Ops/sec versus 2-socket Intel Xeon Skylake Gold 6148 (2x20-core/2.4 GHz/256 GB memory): 932,273 Ops/sec. (2) DB2 Warehouse (2.43X per core): 20c S922 (2x10-core/2.9 GHz/512 GB memory): 3242 QpH versus 2-socket Intel Xeon Skylake Platinum 8168 (2x24-core/2.7 GHz/512 GB memory): 3203 QpH. (3) DayTrader 7 (3.19X per core): 24c S924 (2x12-core/3.4 GHz/512 GB memory): 32221.4 tps versus 2-socket Intel Xeon Skylake Platinum 8180 (2x28-core/2.5 GHz/512 GB memory): 23497.4 tps. (2) 2.6X memory capacity is based on 4TB per socket for POWER9 and 1.5TB per socket for x86 Scalable Platform Intel product brief: https://www.intel.com/content/dam/www/public/us/en/documents/product-briefs/xeon-scalable-platform-brief.pdf?asset=14606 (3) 1.8X bandwidth is based on 230 GB/sec per socket for POWER9 and 128GB/sec per socket for x86 Scalable Platform Intel product brief: https://www.intel.com/content/dam/www/public/us/en/documents/product-briefs/xeon-scalable-platform-brief.pdf?asset=14606 1.8x3 Memory bandwidth per socket (230GB / sec) 2x1 4x Up to 8 Threads per core
- 8. © 2020 IBM Corporation IBM Cognitive Systems Server virtualization security is critical for DB workloads since many are run in virtual environments • The PowerVM hypervisor has only had one hypotetical reported security vulnerability and provides the bullet-proof security that customers demand for mission-critical workloads • The VIOS, which is part of the overall virtualization has had 0 reported security vulnerabilities • Dare to compare – search any security tracking DB and compare Power against x86 1 reported hypothetical security breeches on the PowerVM hypervisor (in Dec 2020) Power VM security March 2021
- 9. © 2020 IBM Corporation IBM Cognitive Systems A Portfolio for the Data & AI Era From Mission-Critical workloads to AI and Cloud Computing leadership AC922 • Industry first and only in advanced IO with 2nd Generation CPU - GPU NVLink delivering ~5.6x higher data throughput •Up to 4 integrated NVIDIA “Volta” GPUs air cooled (GTH) and up to 6 GPUs with water cooled (GTX) version •OpenCAPI support •Memory coherence IC922 •Storage dense, high bandwidth server – up to 24 NVMe or SAS/SATA in 2U1 •Advanced IO with PCIe Gen4 •Optimized inferencing server with up to 6 Nvidia T4 GPUs at GA and additional accelerators in roadmap1 •OpenCAPI support1 •Price/performance server Accelerated Compute Data, Inferencing, and Cloud Enterprise Private Cloud servers GPUs! • With up to 192 POWER9 cores, up to 64 TB memory and the fastest POWER9 processors in the Power Systems portfolio, the Power E980 delivers extraordinary performance and availability for data centers with demanding AIX, IBM i and Linux applications. • The Power E950 is ideal for cloud deployments with built-in virtualization and flexible capacity. It allows you to deliver faster business results by increasing throughput and reducing response time with POWER9™ processors and increased memory and I/O bandwidth. E950 / E980 S914 S922 S924 •Three different form factors: Tower (S914), 2U (S922) and 4U (S924) •Industry leading reliability and computing capability •PowerVM ecosystem focus for outstanding utilization •Focus on memory capacity with up to 4TB of RAM PowerVM and high RAS GPUs!
- 10. © 2020 IBM Corporation IBM Cognitive Systems 10 https://www.nextplatform.com/2018/08/28/ibm-power-chips-blur-the-lines-to-memory-and-accelerators/ Longevity brings maturity and stability to ecosystem
- 11. © 2020 IBM Corporation IBM Cognitive Systems Open POWER based Super Computers
- 12. © 2020 IBM Corporation IBM Cognitive Systems Three of the 10 most POWERFUL HPC systems: Summit, Sierra & Marconi The United States Department of Energy together with Oak Ridge National Laboratory and Lawrence Livermore National Laboratory have contracted IBM and Nvidia to build two supercomputers, the Summit and the Sierra, that are based on POWER9 processors coupled with Nvidia's Volta GPUs. These systems went online in 2018. http://www.teratec.eu/actu/calcul/Nvidia_Coral_White_Paper_Final_3_1.pdf IBM Summit #2 !! IBM Sierra #3 !! IBM Marconi100 #9 !! Online in 2020 #1 and #2 from 2018- 2020
- 13. © 2020 IBM Corporation IBM Cognitive Systems 13
- 14. © 2020 IBM Corporation IBM Cognitive Systems CINECA MARCONI100 14 © Copyright IBM Corporation 2020 https://www.ibm.com/case-studies/cineca-systems-power-hpc-exascale IBM POWER9 + NVIDIA A new accelerated HPC system will be installed at Cineca in February 2020. This system, acquired by Cineca within the PPI4HPC European initiative, opens the way to the pre-exascale Leonardo supercomputer expected to be installed in 2021. MARCONI100, is based on the IBM Power9 architecture with NVIDIA Volta GPUs. Specifically, each node will host 2x16 cores IBM POWER9 AC922 at 3.1 GHz with 256 GB/node of RAM memory and 4 x NVIDIA Volta V100 GPUs per node, Nvlink 2.0, 16GB. The number of nodes will be 980, totallying 31360 cores. Internal Network: Mellanox Infiniband EDR DragonFly+ Model: IBM Power AC922 (Whiterspoon) Racks: 55 total (49 compute) Nodes: 980 Cores: 31360 Processors: 2 x 16 cores IBM POWER9 AC922 at 3.1 GHz Accelerators: 4 x NVIDIA Volta V100 GPUs, Nvlink 2.0, 16GB Cores: 32 cores/node RAM: 256 GB/node Peak Performance: about 32 Pflop/s Internal Network: Mellanox Infiniband EDR DragonFly+ Disk Space: 8PB Gpfs storage
- 15. © 2020 IBM Corporation IBM Cognitive Systems Part of Mare Nostrum 4 • 3 Racks • 54 Power9 Systems • 54 AC922 servers • 4x Nvidia V100 • 512 GB RAM • 6.4 TB NVMe storage • 1.48 Pflops ! BSC Mare Nostrum 4
- 16. © 2020 IBM Corporation IBM Cognitive Systems IBM's Pangea III is the world's most powerful commercial supercomputer 16 Total’s Supercomputer Ranked First in Industry Worldwide The new IBM POWER9-based supercomputer (25 PFLOPS & 50 Pbytes) will help Total more accurately locate new resources and better assess the potential of new opportunities. According to Total Pangea III requires 1.5 Megawatts, compared to 4.5 MW for its predecessor system. Combined with the increased performance of Pangea III, Total has reported that they have observed that the new system uses less than 10% the energy consumption per petaflop as its predecessor. • Higher Resolution Seismic Imaging in exploration and development phase • Reliable Development and Production Models • Asset Valuation and Selectivity
- 17. © 2020 IBM Corporation IBM Cognitive Systems Satori PowerAI Cluster at MIT satori.mit.edu is the name of a new scalable AI oriented hardware resource for research computing at MIT. It is made possible by a donation through IBM Global Universities Program. Provided as a gift from IBM it will help further the aims of the new MIT Stephen A. Schwarzman College of Computing and other campus initiatives that are combining supercomputing power and AI algorithmic innovation. Announced by Dr John Kelly at the Schwarzman School Kick Off Feb 28. : ‘A slice of Summit’ Satori Specs: • 2560 POWER9 Cores with NVLINK 2.0 • 256 NVIDIA V100 GPUs with NVLINK 2.0 • 64TB RAM DDR4 (8 channels) • PCIe4.0 InfiniBand EDR Interconnect • 2.9 PB of shared storage • 1.46 PFLop @ 94 kW (#4 #Green500) Satori lives at Mass Green HPC Center (MGHPCC) Operational since MGHPCC Sept 30, 2019 MIT - IBM Watson AI Lab / version 1.0 https://researchcomputing.mit.edu/satori/home/
- 18. © 2020 IBM Corporation IBM Cognitive Systems 18 The Best Server for Enterprise AI IBM® Power System™ Accelerated Compute Server (AC922)
- 19. © 2020 IBM Corporation IBM Cognitive Systems Mechanical Overview Operator Interface • 1 USB 3.0 • Power Button • Service LED’s 4X - Cooling Fans • Counter- Rotating • Hot swap • 80mm Memory DIMM’s (16x) • 8 DDR4 IS DIMMs per socket Power 9 Processor (2x) • 190W & 250W BMC (Service Processor Card) • IPMI • 2x 1 Gb Ethernet • 1 VGA • 1 USB 3.0 PCIe slot (4x) • Gen4 PCIe • 2, x16 HHHL Adapter • 1, x8,x8 Shared HHHL Adapter • 1 x4 HHHL Adapter NVidia Volta GPU • 2 per socket • SXM2 form factor • 300W • NVLink 2.0 • Air Cooled Power Supplies (2x) • 2200W • Configuration limits for redundancy • Hot Swap • 200VAC, 277VAC, 400VDC input Storage • Optional 2x SFF SATA Disk • Optional 2x SFF SATA SSD • Disk are tray based for hot swap Note: Front Bezel removed 19 § 4 PCIe Gen4 Slots § 2x SFF (HDD/SSD), SATA, Up to 7.7 TB storage § Supports 1.6TB and 3.2TB NVMe Adapters § Redundant Hot Swap Power Supplies and Fans § Default 3 year 9x5 warranty, 100% CRU § 4 PCIe Gen4 Slots § 2x SFF (HDD/SSD), SATA, Up to 7.7 TB storage § Supports 1.6TB and 3.2TB NVMe Adapters § Redundant Hot Swap Power Supplies and Fans § Default 3 year 9x5 warranty, 100% CRU
- 20. © 2020 IBM Corporation IBM Cognitive Systems Front & Rear Details Front Rear 80mm CR Cooling Fans (4x) Note: Front bezel is removed in this illustration USB 3.0 SFF-4 Carrier (2X) • SFF SATA HDD or SSD Service Indicators USB 3.0 1Gb Eth (2x) IPMI VGA PCIe Slot 2 • Gen4 Shared x8,x8 • HHHL Slot • CAPI Enabled PCIe Slot 1 • Gen4 x4 (x8 Connector) • HHHL Slot Power Supplies (2X) Water lines (Option) Service Indicators Power Button PCIe Slot 3 & 4 • Gen4 x16 • HHHL Slot • CAPI Enabled 20
- 21. © 2020 IBM Corporation IBM Cognitive Systems https://www.nvidia.com/en-us/data-center/tesla-v100/ Volta architecture fits in the same power envelope and form factor as the previous Pascal generation GPU with 1.5x the memory performance, 2x the NVLink performance, and 7.8 teraflops of FP64 processing (15.7 teraflops at FP32) on the GPU (CUDA) cores and a total of 125 teraflops of processing performance with the tensor cores.
- 22. © 2020 IBM Corporation IBM Cognitive Systems 22 AC922 System buses and components diagram 32 -140+GB/s 64GB/s Fast link to exchange memory contents between servers Fast link to share memory contents with the GPUs
- 23. © 2020 IBM Corporation IBM Cognitive Systems Large Memory Support (LMS) Distributed Deep Learning (DDL) Other Frameworks (Snap ML) Software Solutions for Open POWER accelerated servers
- 24. © 2020 IBM Corporation IBM Cognitive Systems 24 Large Memory Support (LMS) Objective: Overcome GPU Memory Limitations in DL Training. Increase the Batch Size and/or increase the resolution of the features. LMS enables processing of high definition images, large models, and higher batch sizes that doesn’t fit in GPU memory today (Maximum GPU memory available in Nvidia P100 and V100 GPUs is 16/32GB). Available for - Caffe - TensorFlow - Chainer https://www.sysml.cc/doc/127.pdf GPU RAM System RAM NVLink v2.0 2 TB 16/32 GB Accelerated by NVLink Dataset
- 25. © 2020 IBM Corporation IBM Cognitive Systems https://www.linkedin.com/pulse/deep-learning-high-resolution-images-large-models-sumit-gupta/ 25
- 26. © 2020 IBM Corporation IBM Cognitive Systems Performance results with TensorFlow Large Model Support v2 26 ResNet50 3D U-Net TensorFlow Large Model Support in PowerAI 1.6 allows training models with much higher resolution data. Combining the large model support with the IBM Power Systems AC922 server allows the training of these high resolution models with low data rate overhead. https://developer.ibm.com/linuxonpower/2019/05/17/performance-results-with-tensorflow-large-model-support-v2/ TF LMS v2 DeepLabV3+
- 27. © 2020 IBM Corporation IBM Cognitive Systems 27 LMS in Caffe $caffe train -solver solver.prototxt -bvlc -gpu 0,1,2,3 -lms 10000 -lms_frac 0.5 • -lms 10000. Any memory chunk allocation larger than 10000KB will be done in CPU memory, and fetched to GPU memory only when needed for computation. • -lms_frac 0.5. LMS doesn’t kick in until more than at least 50% of GPU memory is expected to be utilized. Note that configuring the “lms” and “lms_frac” values depends on the below factors: • Batch size used • Model used • Number of GPUs used • System memory available Arriving at an optimal configuration requires understanding of the above and experimentation based on that. A general guideline is that the optimal configuration should utilize GPU memory close to fullest.
- 28. © 2020 IBM Corporation IBM Cognitive Systems 28 Demo https://developer.ibm.com/linuxonpower/2017/09/22/realizing-value-large-model-support-lms-powerai-ibm-caffe/ #Set the cpu to performance mode lscpu; ppc64_cpu –smt; ppc64_cpu --smt=2 cpupower -c all frequency-set -g performance #Check gpu status nvidia-smi ; nvidia-smi -i 0 –q; nvidia-smi -ac 877,1530 #Activate caffe cd lms source /opt/DL/caffe/bin/caffe-activate #Show solver to check the paths vi solver.prototxt #Check in model batchsize = 1 and paths vi models/googlenet_big.prototxt #Run the training caffe train -solver solver.prototxt -bvlc -gpu 0,1 #Change model batchsize = 5, it will show OOM error (out of memory) once we run the training. vi models/googlenet_big.prototxt caffe train -solver solver.prototxt -bvlc -gpu 0,1 #Try running the training again with lms support (tip, it will work now) caffe train -solver solver.prototxt -bvlc -gpu 0,1 -lms 10000 28
- 29. © 2020 IBM Corporation IBM Cognitive Systems Enabling tensorflow Large Model Support 29 Keras API from tensorflow.contrib.lms import LMSKerasCallback lms_callback = LMSKerasCallback() … model.fit_generator(generator=training_gen, callbacks=[lms_callback]) Estimator API with tf.name_scope(‘gradientscope'): optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.001) train_op = optimizer.minimize(loss=loss, global_step=tf.train.get_global_step()) from tensorflow.contrib.lms import LMSHook lms_hook = LMSHook({‘gradientscope '}) mnist_classifier.train(input_fn=train_input_fn, steps=20000, hooks=[logging_hook, lms_hook])
- 30. © 2020 IBM Corporation IBM Cognitive Systems More information 30 ü TensorFlow Large Model Support Code / Pull Request: ü https://github.com/tensorflow/tensorflow/pull/19845/ ü TensorFlow Large Model Support Research Paper: ü https://arxiv.org/pdf/1807.02037.pdf ü TensorFlow Large Model Support Case Study: ü https://developer.ibm.com/linuxonpower/2018/07/27/tensorflow-large-model-support-case-study-3d-image-segmentation/ ü IBM AC922 with NVLink 2.0 connections between CPU and GPU: ü https://www.ibm.com/us-en/marketplace/power-systems-ac922
- 31. © 2020 IBM Corporation IBM Cognitive Systems 31 Distributed Deep Learning Objective: Overcome the server boundaries of some DL frameworks. How: Scaling. Using “ddlrun” applied to Topology aware distributed frameworks. Our software does deep learning training fully synchronously with very low communication overhead. The overall goal of ddlrun is to improve the user experience DDL users. To this end the primary features of ddlrun are: • Error Checking/Configuration Verification • Automatic Rankfile generation • Automatic mpirun option handling Available for: • Tensorflow • IBM Caffe • Torch https://www.sysml.cc/doc/127.pdf Good for: • Speed • Accuracy
- 32. © 2020 IBM Corporation IBM Cognitive Systems Distributed Deep Learning (DDL) for Training phase Using the Power of 100s of Servers August 8, 2017 16 Days Down to 7 Hours: Near Ideal Scaling to 256 GPUs and Beyond 1 System 64 Systems 16 Days 7 Hours ResNet-101, ImageNet-22K, Caffe with PowerAI DDL, Running on Minsky (S822Lc) Power System 58x Faster https://www.ibm.com/blogs/research/2017/08/distributed-deep-learning/
- 33. © 2020 IBM Corporation IBM Cognitive Systems Demo with Slurm 33 ssh [email protected] #Slurm login node # You should have a ~/data dir with the dataset downloaded or internet conection to download it #Edit and include the following line in ~/.bashrc export TMPDIR=/tmp/ # To pass all the variables, like activate ...., you may need to write a simple submission script: Run as “sbatch script.sh” [bsc18651@p9login2 ~]$ cat script.sh #!/bin/bash #SBATCH -J test #SBATCH -D . #SBATCH -o test_%j.out #SBATCH -e test_%j.err #SBATCH -N 2 #SBATCH --ntasks-per-node=4 #SBATCH --gres="gpu:4" #SBATCH --time=01:00:00 module purge module load anaconda2 powerAI source /opt/DL/ddl-tensorflow/bin/ddl-tensorflow-activate export TMPDIR="/tmp/" export DDL_OPTIONS="-mode b:4x2" NODE_LIST=$(scontrol show hostname $SLURM_JOB_NODELIST | tr 'n' ',') NODE_LIST=${NODE_LIST%?} cd examples/mnist ddlrun -n 8 -H $NODE_LIST python mnist-init.py --ddl_options="-mode b:4x2" --data_dir /home/bsc18/bsc18651/examples/mnist/data [bsc18651@p9login2 ~]$ sbatch script.sh https://developer.ibm.com/linuxonpower/2018/05/01/improved-ease-use-ddl-powerai/
- 34. © 2020 IBM Corporation IBM Cognitive Systems 34 Example without slurm 1. Install PowerAI 1.6.0 in all nodes participant 2. Locate tf_cnn_benchmarks/tf_cnn_benchmarks.py under your pai conda env. 3. verify you have passwordless ssh access across two AC922/Minsky nodes 4. RUN: $ ddlrun -H host1,host2 python tf_cnn_benchmarks.py --batch_size=256 -- num_batches=128 --data_format=NCHW --optimizer=sgd --variable_update=ddl - -num_gpus=8 --model=alexnet --tcp ddlrun -H minsky1,minskyx --tcp python tf_cnn_benchmarks.py --batch_size=256 -- num_batches=128 --data_format=NCHW --optimizer=sgd --variable_update=ddl -- num_gpus=1 --model=alexnet 5. Have FUN!
- 35. © 2020 IBM Corporation IBM Cognitive Systems Most used data science methods in 2017 (according to Kaggle) 35 https://www.kaggle.com/surveys/2017
- 36. © 2020 IBM Corporation IBM Cognitive Systems SnapML v2 36 Our latest version of Snap ML adds high performance implementations of Decision Trees and Random Forests. Our implementation of these algorithms takes advantage of multiple CPU cores and threads (but so far do not take advantage of GPUs). Click here for documentation on Snap ML. https://medium.com/@sumitg_16893/snap-ml-2x-faster-machine-learning-than-scikit-learn-c3529a1a6172 Snap ML, a python-based machine learning framework that is designed to be a high-performance machine learning software framework. Snap ML is bundled as part of the WML Community Edition or WML CE (aka PowerAI) software distribution that is available for free on Power systems. The first release of Snap ML enabled GPU-acceleration of generalized linear models (GLMs) and also enabled scaling these models to multiple GPUs and multiple servers. GLMs are popular machine learning algorithms, which include logistic regression, linear regression, ridge and lasso regression, and support vector machines (SVMs). http://www.cirrascale.com/ibmpower_snapml.php Snap ML (PowerAI 1.6.0) supports: •Generalized Linear models • Logistic Regression • Linear Regression • Ridge Regression • Lasso Regression • Support Vector Machines •Tree-based models • Decision Trees • Random Forest •2H19 release will add support for •Gradient Boosting Machines (GBMs)
- 37. © 2020 IBM Corporation IBM Cognitive Systems The change to 100% Community Maintained SW! 37 POWERAI / WATSON MACHINE LEARNING COMMUNITY EDITION https://mit-satori.github.io/satori-ai-frameworks.html
- 38. © 2020 IBM Corporation IBM Cognitive Systems 38 https://github.com/open-ce
- 39. © 2020 IBM Corporation IBM Cognitive Systems State of the Art SW Solutions leverage AI on POWER
- 40. © 2020 IBM Corporation IBM Cognitive Systems 41 Red Hat OpenShift Cloud Pak for Applications Cloud Pak for Data Red Hat OpenShift Cloud Pak for Integration Red Hat OpenShift Red Hat OpenShift Cloud Pak for Automation Red Hat OpenShift Cloud Pak for Multicloud Management Node.js React Kitura.js Jenkins Nginx Spring Istio Swift Open Liberty WAS WAS ND App Connect Enterprise API Connect DataPower Event Streams MQ Aspera Watson Voice Gateway * RabbitMQ Integration Explorer Business Automation Workflow Operational Decision Management Business Automation Insight Business Automation Content Analyzer File Connect Manager + Add-on * Multicloud Manager Cloud Application Management Cloud Event Manager Cloud Automation Manager Cloudant DB2 Data Virtualization Cognos Streams MongoDB MariaDB* Redis* etcd + Add-on * (including Watson AI, …) WAS Liberty Transformation Advisor JBoss Kabanero Enterprise RHOAR - OpenShift Application Runtimes *UrbanCode Deploy *Developer Team Orch *Dev Team Governance + Add-on * + Add-on * Robotic Process Automation * * PostgreSQL * NetApp Persistent Storage * Wand Taxonomies and Ontologies * Knowis for Banking * Lightbend Reactive Microservices * Prolifics Prospecting Accelerator IBM Mobile Foundations Runs on choice of IBM Power Systems Infrastructure-as- a-Service (IaaS) Bare-metal Available today IBM Cloud Paks
- 41. © 2020 IBM Corporation IBM Cognitive Systems IBM Cloud Pak For Data: Base Platform vs Cartridges ü Db2 Warehouse ü Data Virtualization ü Db2 Eventstore ü IBM Streams ü Watson Knowledge Catalog (including IGC) ü IBM Regulatory Accelerator ( included in WKC) ü Information Analyzer ü Watson Studio (includes Data Refinery) ü Watson Machine learning (includes AutoAI ) ü Watson OpenScale ü Cognos Dashboards Embedded ü Analytics Engine for Apache Spark ü Open source governance ü IBM Performance Server (only on CPD System) ü Db2 ü DataStage ü Cognos Analytics ü Information Server ü Watson Studio Premium ü Watson Assistant ü Watson Discovery ü Watson API Kit ü Watson Financial Crimes Insights Base Platform Extensions 42 Services supported on Power Systems in YELLOW
- 42. © 2020 IBM Corporation IBM Cognitive Systems Data Sources 43
- 43. © 2020 IBM Corporation IBM Cognitive Systems “Cloud Pak for Data” is an integrated Data & AI Platform Common Services 44
- 44. © 2020 IBM Corporation IBM Cognitive Systems IBM Cloud Pak for Data DEMO
- 45. © 2020 IBM Corporation IBM Cognitive Systems Power IC922 server details IBM Cognitive Systems / January 2020 / © 2020 IBM Corporation 8x SAS/SATA 8x SAS/SATA 8x SAS/SATA 1. Statements regarding IBM’s future direction and intent are subject to change or withdrawal without notice, and represent goals and objectives only. Full Statement of Direction in the speaker notes • 19” Rack, 2 Socket 2U Form Factor • OpenPOWER partner design • Two POWER9 chips with SMT4 capability – 12, 16, or 20 cores = 40 cores in our case • Maximum memory bandwidth – 2TB maximum system memory – 8 DDR4 ports at 2666Mhz enables 170 GB/s peak memory bandwidth per chip – 32 DDR4 RDIMM slots • 10 PCIe Slots – Support for up to 6 Nvidia T4 GPUs (future support up to 8 accelerated devices)1 • 24 SFF (2.5”) Storage bays - SAS/SATA (future NVMe support)1 • Linux-only system – RHEL 7 & 8
- 46. © 2020 IBM Corporation IBM Cognitive Systems Red Hat OpenShift 4.6.35 OCP architecture: - Nodes running as KVM VMs - 3 Master nodes - 6 Worker nodes - 128GB RAM & 24 VCPUs & 200GB Storage https://developer.ibm.com/tutorials/installation-of-cloud-pak-on-ocp-on-powervs/
- 47. © 2020 IBM Corporation IBM Cognitive Systems Login to the Cloud Pak for Data running in a POWER9 server 48
- 48. © 2020 IBM Corporation IBM Cognitive Systems Dataset used for the Auto AI experiment 49
- 49. © 2020 IBM Corporation IBM Cognitive Systems Select winner model and create Notebook for accountability. 50
- 50. © 2020 IBM Corporation IBM Cognitive Systems Put the model in production sharing access code and online quick test. 51
- 51. © 2020 IBM Corporation IBM Cognitive Systems 52 by Your Innovation Platform!
- 52. © 2020 IBM Corporation IBM Cognitive Systems Notice and disclaimers ü Copyright © 2017 by International Business Machines Corporation (IBM). No part of this document may be reproduced or transmitted in any form without written permission from IBM. ü U.S. Government Users Restricted Rights — use, duplication or disclosure restricted by GSA ADP Schedule Contract with IBM. ü Information in these presentations (including information relating to products that have not yet been announced by IBM) has been reviewed for accuracy as of the date of initial publication and could include unintentional technical or typographical errors. IBM shall have no responsibility to update this information. This document is distributed “as is” without any warranty, either express or implied. In no event shall IBM be liable for any damage arising from the use of this information, including but not limited to, loss of data, business interruption, loss of profit or loss of opportunity. IBM products and services are warranted according to the terms and conditions of the agreements under which they are provided. ü IBM products are manufactured from new parts or new and used parts. In some cases, a product may not be new and may have been previously installed. Regardless, our warranty terms apply.” ü Any statements regarding IBM's future direction, intent or product plans are subject to change or withdrawal without notice. ü Performance data contained herein was generally obtained in a controlled, isolated environments. Customer examples are presented as illustrations of how those customers have used IBM products and the results they may have achieved. Actual performance, cost, savings or other results in other operating environments may vary. ü References in this document to IBM products, programs, or services does not imply that IBM intends to make such products, programs or services available in all countries in which IBM operates or does business. ü Workshops, sessions and associated materials may have been prepared by independent session speakers, and do not necessarily reflect the views of IBM. All materials and discussions are provided for informational purposes only, and are neither intended to, nor shall constitute legal or other guidance or advice to any individual participant or their specific situation. ü It is the customer’s responsibility to insure its own compliance with legal requirements and to obtain advice of competent legal counsel as to the identification and interpretation of any relevant laws and regulatory requirements that may affect the customer’s business and any actions the customer may need to take to comply with such laws. IBM does not provide legal advice or represent or warrant that its services or products will ensure that the customer is in compliance with any law.
- 53. © 2020 IBM Corporation IBM Cognitive Systems Notice and disclaimers continued Information concerning non-IBM products was obtained from the suppliers of those products, their published announcements or other publicly available sources. IBM has not tested those products in connection with this publication and cannot confirm the accuracy of performance, compatibility or any other claims related to non-IBM products. Questions on the capabilities of non-IBM products should be addressed to the suppliers of those products. IBM does not warrant the quality of any third-party products, or the ability of any such third-party products to interoperate with IBM’s products. IBM expressly disclaims all warranties, expressed or implied, including but not limited to, the implied warranties of merchantability and fitness for a particular, purpose. The provision of the information contained herein is not intended to, and does not, grant any right or license under any IBM patents, copyrights, trademarks or other intellectual property right. IBM, the IBM logo, ibm.com, AIX, BigInsights, Bluemix, CICS, Easy Tier, FlashCopy, FlashSystem, GDPS, GPFS, Guardium, HyperSwap, IBM Cloud Managed Services, IBM Elastic Storage, IBM FlashCore, IBM FlashSystem, IBM MobileFirst, IBM Power Systems, IBM PureSystems, IBM Spectrum, IBM Spectrum Accelerate, IBM Spectrum Archive, IBM Spectrum Control, IBM Spectrum Protect, IBM Spectrum Scale, IBM Spectrum Storage, IBM Spectrum Virtualize, IBM Watson, IBM z Systems, IBM z13, IMS, InfoSphere, Linear Tape File System, OMEGAMON, OpenPower, Parallel Sysplex, Power, POWER, POWER4, POWER7, POWER8, Power Series, Power Systems, Power Systems Software, PowerHA, PowerLinux, PowerVM, PureApplica- tion, RACF, Real-time Compression, Redbooks, RMF, SPSS, Storwize, Symphony, SystemMirror, System Storage, Tivoli, WebSphere, XIV, z Systems, z/OS, z/VM, z/VSE, zEnterprise and zSecure are trademarks of International Business Machines Corporation, registered in many jurisdictions worldwide. Other product and service names might be trademarks of IBM or other companies. A current list of IBM trademarks is available on the Web at "Copyright and trademark information" at: www.ibm.com/legal/copytrade.shtml. Linux is a registered trademark of Linus Torvalds in the United States, other countries, or both. Java and all Java-based trademarks and logos are trademarks or registered trademarks of Oracle and/or its affiliates.