RAPIDS: PLATFORM INSIDE AND OUT Joshua Patterson 3-19-2019 RAPIDS - - PowerPoint PPT Presentation

▶
rapids platform inside and out
SMART_READER_LITE
LIVE PREVIEW

RAPIDS: PLATFORM INSIDE AND OUT Joshua Patterson 3-19-2019 RAPIDS - - PowerPoint PPT Presentation

Sep 17, 2022 209 likes •713 views

RAPIDS: PLATFORM INSIDE AND OUT Joshua Patterson 3-19-2019 RAPIDS End to End Accelerate GPU Data Science Data Preparation Model Training Visualization cuDF cuIO cuML cuGraph PyTorch Chainer MxNet cuXfilter <> Kepler.gl Analytics

slide-1
SLIDE 1

Joshua Patterson 3-19-2019

RAPIDS: PLATFORM INSIDE AND OUT

slide-2
SLIDE 2

2

cuDF cuIO Analytics GPU Memory Data Preparation Visualization Model Training cuML Machine Learning cuGraph Graph Analytics PyTorch Chainer MxNet Deep Learning cuXfilter <> Kepler.gl Visualization

RAPIDS

End to End Accelerate GPU Data Science

slide-3
SLIDE 3

3

cuDF cuIO Analytics GPU Memory Data Preparation Visualization Model Training cuML Machine Learning cuGraph Graph Analytics PyTorch Chainer MxNet Deep Learning cuXfilter <> Kepler.gl Visualization

RAPIDS

End to End Accelerate GPU Data Science

slide-4
SLIDE 4

4

APP A

DATA MOVEMENT AND TRANSFORMATION

The bane of productivity and performance

CPU GPU

APP B Read Data Copy & Convert Copy & Convert Copy & Convert Load Data APP A

GPU Data

APP B

GPU Data

APP A APP B

slide-5
SLIDE 5

5

APP A

DATA MOVEMENT AND TRANSFORMATION

What if we could keep data on the GPU?

APP B Copy & Convert Copy & Convert Copy & Convert APP A

GPU Data

APP B

GPU Data

Read Data Load Data APP B

CPU GPU

APP A

slide-6
SLIDE 6

6

LEARNING FROM APACHE ARROW

From Apache Arrow Home Page - https://arrow.apache.org/

slide-7
SLIDE 7

7

cuDF cuIO Analytics GPU Memory Data Preparation Visualization Model Training cuML Machine Learning cuGraph Graph Analytics PyTorch Chainer MxNet Deep Learning cuXfilter <> Kepler.gl Visualization

RAPIDS

End to End Accelerate GPU Data Science

slide-8
SLIDE 8

8

cuDF cuIO Analytics GPU Memory Data Preparation Visualization Model Training cuML Machine Learning cuGraph Graph Analytics PyTorch Chainer MxNet Deep Learning cuXfilter <> Kepler.gl Visualization

RAPIDS

Core of RAPIDS

slide-9
SLIDE 9

9

ROAD TO 1.0

RAPIDS Is Fast…

0.00 100.00 200.00 300.00 400.00 500.00 600.00 700.00 800.00 lower() find(#) slice(1,15)

milliseconds

Pandas cudastrings

  • CPU to GPU 10-25x improvement on average
  • Simple Python Interface
  • … could be even faster!
  • JIT Compression/Decompression
  • Improved caching
  • More static compiled kernels
slide-10
SLIDE 10

10

ROAD TO 1.0

Focused on Robust Functionality, Deployment, and User Experience

https://9-volt.github.io/bug-life/?repo=rapidsai/cudf Feature Request

  • Window and Rolling Window Functions
  • Improved Apply function
  • Improved String Support
  • Words to Vec
  • Geospatial Functions
  • Improved Integration with Numba
  • Statistical functions (ANOVA, Covariance, etc…)
slide-11
SLIDE 11

11

ROAD TO 1.0

Focused on Robust Functionality, Deployment, and User Experience

https://9-volt.github.io/bug-life/?repo=rapidsai/cudf Bugs cuDF Python cuDF C++

  • Better error handling
  • Replace CFFI with Cython for more descriptive

errors and exceptions

  • Cover more edge cases of functionality
  • Push more functionality down from cuDF Python into

cuDF C++ for performance and future languages

  • Support a proper C++ API
slide-12
SLIDE 12

12

ROAD TO 1.0

GTC Europe – Launch - RAPIDS 0.1

cuGraph SG MG MGMN Jaccard Weighted Jaccard PageRank Louvain SSSP BFS SSWP Triangle Counting Subgraph Extraction

cuML SG MG MGMN Gradient Boosted Decision Trees (GBDT) GLM Logistic Regression Random Forest (regression) K-Means K-NN DBSCAN UMAP ARIMA Kalman Filter Holts-Winters Principal Components Singular Value Decomposition

slide-13
SLIDE 13

13

ROAD TO 1.0

GTC San Jose – Today - RAPIDS 0.6

cuGraph SG MG MGMN Jaccard Weighted Jaccard PageRank Louvain SSSP BFS SSWP Triangle Counting Subgraph Extraction

cuML SG MG MGMN Gradient Boosted Decision Trees (GBDT) GLM Logistic Regression Random Forest (regression) K-Means K-NN DBSCAN UMAP ARIMA Kalman Filter Holts-Winters Principal Components Singular Value Decomposition

slide-14
SLIDE 14

14

ROAD TO 1.0

Q4 – 2019 - RAPIDS 0.12?

cuML SG MG MGMN Gradient Boosted Decision Trees (GBDT) GLM Logistic Regression Random Forest (regression) K-Means K-NN DBSCAN UMAP ARIMA Kalman Filter Holts-Winters Principal Components Singular Value Decomposition

cuGraph SG MG MGMN Jaccard Weighted Jaccard PageRank Louvain SSSP BFS SSWP Triangle Counting Subgraph Extraction

slide-15
SLIDE 15

15

ROAD TO 1.0

Focused on Robust Functionality, Deployment, and User Experience

Integration with every major cloud provider Both containers and cloud specific machine instances Support for Enterprise and HPC Orchestration Layers

slide-16
SLIDE 16

16

ROAD TO 1.0

Focused on Robust Functionality, Deployment, and User Experience

3/19/19 S9788 Building a GPU-Focused CI Solution Michael Wendt

  • CI/CD essential to RAPIDS
  • Airspeed Velocity (ASV) for regression
  • Nightlies!
  • https://hub.docker.com/r/rapid

sai/rapidsai-nightly

  • https://anaconda.org/rapidsai-

nightly

slide-17
SLIDE 17

17

cuDF cuIO Analytics GPU Memory Data Preparation Visualization Model Training cuML Machine Learning cuGraph Graph Analytics PyTorch Chainer MxNet Deep Learning cuXfilter <> Kepler.gl Visualization

RAPIDS

Core of RAPIDS

slide-18
SLIDE 18

18

ETL - THE BACKBONE OF DATA SCIENCE

cuDF is…

CUDA With Python Bindings

  • Low level library containing function implementations and

C/C++ API

  • Importing/exporting Apache Arrow using the CUDA IPC

mechanism

  • CUDA kernels to perform element-wise math operations on GPU

DataFrame columns

  • CUDA sort, join, groupby, and reduction operations on GPU

DataFrames

  • A Python library for manipulating GPU DataFrames
  • Python interface to CUDA C++ with additional functionality
  • Creating Apache Arrow from Numpy arrays, Pandas

DataFrames, and PyArrow Tables

  • JIT compilation of User-Defined Functions (UDFs) using Numba
  • Apache Arrow data format
  • Pandas-like API
  • Unary and Binary Operations
  • Joins / Merges
  • GroupBys
  • Filters
  • User-Defined Functions (UDFs)
  • Accelerated file readers
  • Etc.
slide-19
SLIDE 19

19

cuDF cuIO Analytics GPU Memory Data Preparation Visualization Model Training cuML Machine Learning cuGraph Graph Analytics PyTorch Chainer MxNet Deep Learning cuXfilter <> Kepler.gl Visualization Dask

ETL - THE BACKBONE OF DATA SCIENCE

cuDF is not the end of the story

3/19/19 S9793 cuDF: RAPIDS GPU-Accelerated Data Frame Library Keith Kraus & Dante Gama Dessavre 3/20/19 RAPIDS CUDA DataFrame Internals for C++ Developers Jake Hemstad

slide-20
SLIDE 20

20

ETL - THE BACKBONE OF DATA SCIENCE

Why Dask

  • PyData Native
  • Built on top of NumPy, Pandas Scikit-Learn,

… (easy to migrate)

  • With the same APIs (easy to train)
  • With the same developer community (well

trusted)

  • Scales
  • Easy to install and use on a laptop
  • Scales out to thousand-node clusters
  • Popular
  • Most common parallelism framework today

at PyData and SciPy conferences

  • Deployable
  • HPC: SLURM, PBS, LSF

, SGE

  • Cloud: Kubernetes
  • Hadoop/Spark: Yarn
slide-21
SLIDE 21

21

ETL - THE BACKBONE OF DATA SCIENCE

Why Dask

  • PyData Native
  • Built on top of NumPy, Pandas Scikit-Learn,

… (easy to migrate)

  • With the same APIs (easy to train)
  • With the same developer community (well

trusted)

  • Scales
  • Easy to install and use on a laptop
  • Scales out to thousand-node clusters
  • Popular
  • Most common parallelism framework today

at PyData and SciPy conferences

  • Deployable
  • HPC: SLURM, PBS, LSF

, SGE

  • Cloud: Kubernetes
  • Hadoop/Spark: Yarn
slide-22
SLIDE 22

22

ETL - THE BACKBONE OF DATA SCIENCE

Dask-cuDF improvements in 0.7 & 0.8

  • Make cuDF more Pandas like
  • The more cuDF follows the Pandas API, the fewer changes to Dask

DataFrame

  • Replace Dask communications with Open UCX
  • Pickling CUDA IPC was a clever hack, but would not scale past a

single node

  • Focus on Dask-cuDF errors
  • Dask will prevent most out of memory errors users currently

experience with cuDF alone

  • Improvements to Dask-cuDF still improve cuDF
  • Better memory monitoring in Dask
  • Improve String Support…
slide-23
SLIDE 23

23

ETL - THE BACKBONE OF DATA SCIENCE

String Support in Dask-cuDF

Now 0.6 String Support:

  • Element-wise operations
  • Split, Find, Extract, Cat, Typecasting, etc…
  • String GroupBys
  • String Joins
  • Power Support now possible

Future 0.7 & 0.8 String Support:

  • GPU accelerated to_csv
  • More Pandas String API compatibility
  • Element-wise String Comparisons
  • Improved Categorical column support
slide-24
SLIDE 24

24

EXTRACTION IS THE CORNERSTONE OF ETL

cuIO Is Born

  • CSV Reader
  • Follows API of pandas.read_csv
  • Current implementation is >10x speed

improvement over pandas

  • Parquet Reader – v0.7
  • Work in progress: Will follow API of

pandas.read_parquet

  • ORC Reader – v0.7
  • Work in progress: Will have similar API of

Parquet reader

  • Additionally looking towards GPU-accelerating

decompression for common compression schemes

Source: Apache Crail blog: SQL Performance: Part 1 - Input File Formats

slide-25
SLIDE 25

25

ETL IS NOT JUST DATAFRAMES!

slide-26
SLIDE 26

26

cuDF cuIO Analytics GPU Memory Data Preparation Visualization Model Training cuML Machine Learning cuGraph Graph Analytics PyTorch Chainer MxNet Deep Learning cuXfilter <> Kepler.gl Visualization

RAPIDS

Core of RAPIDS

Dask

slide-27
SLIDE 27

27

cuDF cuIO CuPy Numba Analytics GPU Memory Data Preparation Visualization Model Training cuML Machine Learning cuGraph Graph Analytics PyTorch Chainer MxNet Deep Learning cuXfilter <> Kepler.gl Visualization

RAPIDS

Core of RAPIDS

Dask

slide-28
SLIDE 28

28

INTEROPERABILITY FOR THE WIN

DLPack and __cuda_array_interface__

slide-29
SLIDE 29

29

INTEROPERABILITY FOR THE WIN

DLPack and __cuda_array_interface__

slide-30
SLIDE 30

30

ETL – ARRAYS & DATAFRAMES

Dask and CUDA Python Arrays

  • Scales NumPy to distributed clusters
  • Used in climate science, imaging, HPC analysis

up to 100TB size

  • Now seamlessly accelerated with GPUs
slide-31
SLIDE 31

31

ETL – ARRAYS & DATAFRAMES

Dask-CuPy

https://blog.dask.org/2019/01/03/dask-array-gpus-first- steps

slide-32
SLIDE 32

32

ETL – ARRAYS & DATAFRAMES

Dask-CuPy

https://blog.dask.org/2019/01/03/dask-array-gpus-first- steps https://cloud.google.com/blog/products/ai-machine-learning/nvidias-rapids-joins-our- set-of-deep-learning-vm-images-for-faster-data-science

slide-33
SLIDE 33

33

ETL – ARRAYS & DATAFRAMES

Dask-CuPy

Cluster configuration: 20x GCP instances, each instance has: CPU: 1 VM socket (Intel Xeon CPU @ 2.30GHz), 2- core, 2 threads/core, 132GB mem, GbE ethernet, 950 GB disk GPU: 4x NVIDIA Tesla P100-16GB-PCIe (total GPU DRAM across nodes 1.22 TB) Software: Ubuntu 18.04, RAPIDS 0.5.1, Dask=1.1.1, Dask-Distributed=1.1.1, CuPY=5.2.0, CUDA 10.0.130

https://blog.dask.org/2019/01/03/dask-array-gpus-first- steps

slide-34
SLIDE 34

34

ETL – ARRAYS & DATAFRAMES

Dask-CuPy

slide-35
SLIDE 35

35

ETL – ARRAYS & DATAFRAMES

Dask-CuPy is Easy to Implement

slide-36
SLIDE 36

36

ETL – ARRAYS & DATAFRAMES

More Dask Awesomeness from RAPIDS

https://youtu.be/gV0cykgsTPM https://youtu.be/R5CiXti_MWo 3/19/19 S9797 Dask Extensions and New Developments with RAPIDS Matthew Rockling

slide-37
SLIDE 37

37

cuDF cuIO Analytics GPU Memory Data Preparation Visualization Model Training cuML Machine Learning cuGraph Graph Analytics PyTorch Chainer MxNet Deep Learning cuXfilter <> Kepler.gl Visualization Dask

NEW FEATURES WITH GRAPH

Connected Data the Next Frontier

slide-38
SLIDE 38

38

PAGERANK

1 RTX8000 or Many CPU Nodes

cuGraph SG MG MGMN Jaccard Weighted Jaccard PageRank Louvain SSSP BFS SSWP Triangle Counting Subgraph Extraction

slide-39
SLIDE 39

39

PAGERANK

Future of Graph

cuGraph SG MG MGMN Jaccard Weighted Jaccard PageRank Louvain SSSP BFS SSWP Triangle Counting Subgraph Extraction

  • For PageRank Less than 3 seconds spent in algorithm!
  • Multi-GPU with Dask
  • Better Graph Partitioning on GPU

cuGraph SG MG MGMN Jaccard Weighted Jaccard PageRank Louvain SSSP BFS SSWP Triangle Counting Subgraph Extraction

slide-40
SLIDE 40

40

PAGERANK

1 RTX8000 or Many CPU Nodes

cuGraph SG MG MGMN Jaccard Weighted Jaccard PageRank Louvain SSSP BFS SSWP Triangle Counting Subgraph Extraction

More Talks on Graph! 3/21/19 S9802 Context-Aware Network Mapping and Asset Classification Bartley Richardson 3/21/19 S9783 Accelerating Graph Algorithms with RAPIDS Joe Eaton & Brad Rees

cuGraph SG MG MGMN Jaccard Weighted Jaccard PageRank Louvain SSSP BFS SSWP Triangle Counting Subgraph Extraction

slide-41
SLIDE 41

41

cuDF cuIO Analytics GPU Memory Data Preparation Visualization Model Training cuML Machine Learning cuGraph Graph Analytics PyTorch Chainer MxNet Deep Learning cuXfilter <> Kepler.gl Visualization Dask

MACHINE LEARNING

More Models More Problem

slide-42
SLIDE 42

42

UMAP

New Dimensionality Reduction on GPU

https://ai.googleblog.com/2019/03/exploring-neural-networks.html

Uniform Manifold Approximation and Projection (UMAP) is a dimension reduction technique that can be used for visualisation similarly to t-SNE, but also for general non-linear dimension reduction.

  • First of all UMAP is fast… scaling beyond

what most t-SNE packages can manage…

  • Second, UMAP scales well in embedding

dimension -- it isn't just for visualisation! You can use UMAP as a general purpose dimension reduction technique…

  • Third, UMAP often performs better at

preserving aspects of global structure of the data than t-SNE…

  • Fourth, UMAP supports a wide variety of

distance functions…

  • Fifth, UMAP supports adding new points to

an existing embedding via the standard sklearn transform method…

  • Sixth, UMAP supports supervised and semi-

supervised dimension reduction…

  • Finally UMAP has solid theoretical

foundations in manifold learning…

https://arxiv.org/pdf/1802.03426.pdf

Thank You Leland McInnes!

slide-43
SLIDE 43

43

CUML

Single GPU and XGBoost

cuML SG MG MGMN Gradient Boosted Decision Trees (GBDT) GLM Logistic Regression Random Forest (regression) K-Means K-NN DBSCAN UMAP ARIMA Kalman Filter Holts-Winters Principal Components Singular Value Decomposition

slide-44
SLIDE 44

44

DASK-CUML

OLS, tSVD, and KNN in RAPIDS 0.6

cuML SG MG MGMN Gradient Boosted Decision Trees (GBDT) GLM Logistic Regression Random Forest (regression) K-Means K-NN DBSCAN UMAP ARIMA Kalman Filter Holts-Winters Principal Components Singular Value Decomposition

slide-45
SLIDE 45

45

DASK-CUML

K-Means*, DBSCAN & PCA in RAPIDS 0.7/0.8

cuML SG MG MGMN Gradient Boosted Decision Trees (GBDT) GLM Logistic Regression Random Forest (regression) K-Means K-NN DBSCAN UMAP ARIMA Kalman Filter Holts-Winters Principal Components Singular Value Decomposition

  • Deprecating the current K-means in 0.6 for new K-means built on MLPrims
slide-46
SLIDE 46

46

DASK-CUML

cuML and cuMLPrims

cuML SG MG MGMN Gradient Boosted Decision Trees (GBDT) GLM Logistic Regression Random Forest (regression) K-Means K-NN DBSCAN UMAP ARIMA Kalman Filter Holts-Winters Principal Components Singular Value Decomposition

cuML cuMLPrims CUDA cuBLAS CUTLASS

3/20/19 S9817 RAPIDS cuML: GPU Accelerated Machine Learning Onur Yilmaz & Corey Nolet

slide-47
SLIDE 47

47

  • https://ngc.nvidia.com/registry/nvidia-

rapidsai-rapidsai

  • https://hub.docker.com/r/rapidsai/rapidsai/
  • https://github.com/rapidsai
  • https://anaconda.org/rapidsai/
  • https://pypi.org/project/cudf
  • https://pypi.org/project/cuml

RAPIDS

How do I get the software?

slide-48
SLIDE 48

48

JOIN THE MOVEMENT

Everyone Can Help!

Integrations, feedback, documentation support, pull requests, new issues, or code donations welcomed!

APACHE ARROW GPU Open Analytics Initiative

https://arrow.apache.org/ @ApacheArrow http://gpuopenanalytics.com/ @GPUOAI

RAPIDS

https://rapids.ai @RAPIDSAI

slide-49
SLIDE 49

THANK YOU

Joshua Patterson @datametrician