Presented October 18, 2017 While rapid innovation is occurring across the GPU software ecosystem, the platforms themselves still remain isolated from each other - until now. Aaron Williams, the VP of Global Community at MapD, will demo the GPU Open Analytics Initiative's (GOAI) first project on stage - the GPU Data Frame (GDF) - and explain how this approach will enable efficient intra-GPU communication between different processes running on the GPUs.

1. Accelerate Analytics
with a GPU Data Frame
Aaron Williams
October 18, 2017

2. MapD: Extreme Analytics
2
100x Faster Queries
MapD Core
The world’s fastest
columnar database, powered
by GPUs
+
Visualization at the Speed of Thought
MapD Immerse
A visualization front end that
leverages the speed &
rendering superiority of GPUs

3. MapD System Architecture
Accelerating the existing data infrastructure
3

4. 4
MAPD DEMO

5. MapD Benchmarks
Blogger Mark Litwintschik benchmarked MapD on a billion-row taxi data set and
found it to be up to orders-of-magnitude faster than the fastest CPU databases
5
MapD Core: Comparative Query Acceleration*
System Q 1 Q 2 Q 3 Q 4
BrytlytDB & 2-node p2.16xlarge cluster 36x 47x 25x 12x
ClickHouse, Intel Core i5 4670K 49x 58x 32x 25x
Redshift, 6-node ds2.8xlarge cluster 74x 24x 14x 6x
BigQuery 95x 38x 6x 6x
Presto, 50-node n1-standard-4 cluster 190x 75x 61x 41x
Amazon Athena 305x 117x 37x 13x
Elasticsearch (heavily tuned) 386x 343x n/a n/a
Spark 2.1, 11 x m3.xlarge cluster w/ HDFS 485x 153x 119x 169x
Presto, 10-node n1-standard-4 cluster 524x 189x 127x 61x
Vertica, Intel Core i5 4670K 685x 607x 203x 132x
Elasticsearch (lightly tuned) 1,642x 1,194x n/a n/a
Presto, 5-node m3.xlarge cluster w/ HDFS 1,667x 735x 388x 159x
Presto, 50-node m3.xlarge cluster w/ S3 2,048x 849x 164x 86x
PostgreSQL 9.5 & cstore_fdw 7,238x 3,302x 1,424x 722x
Spark 1.6, 5-node m3.xlarge cluster w/ S3 12,571x 5,906x 3,758x 1,884x
*All speed comparisons are to the “MapD & 1 Nvidia Pascal DGX-1” benchmark
Source: http://tech.marksblogg.com/benchmarks.html

6. Query Compilation with LLVM
6
Traditional DBs can be highly inefficient
• each operator in SQL treated as a separate function
• incurs tremendous overhead and prevents vectorization
MapD compiles queries w/LLVM to create one custom function
• Queries run at speeds approaching hand-written functions
• LLVM enables generic targeting of different architectures (GPUs, X86, ARM, etc).
• Code can be generated to run query on CPU and GPU simultaneously
10111010101001010110101101010101
00110101101101010101010101011101
LLVM

7. Keeping Data Close to Compute
MapD maximizes performance by optimizing memory use
7
SSD or NVRAM STORAGE (L3)
250GB to 20TB
1-2 GB/sec
CPU RAM (L2)
32GB to 3TB
70-120 GB/sec
GPU RAM (L1)
24GB to 256GB
1000-6000 GB/sec
Hot Data
Speedup = 1500x to 5000x
Over Cold Data
Warm Data
Speedup = 35x to 120x
Over Cold Data
Cold Data
COMPUTE
LAYER
STORAGE
LAYER
Data Lake/Data Warehouse/System Of Record
SpeedIncreases
SpaceIncreases

8. The Status Quo: Memory Bottlenecks
8
PCIe
4-16GB/s

9. The GPU Open Analytics Initiative Model
Standard in-memory format; zero-copy interchange
9
GPU

10. The GPU Open Analytics Initiative Model
Standard in-memory format; zero-copy interchange
10

11. Interactive Machine Learning
Empowering the People in the Pipeline
11
Personas in
Analytics Lifecycle
(Illustrative)
Business Analyst
Data Scientist
Data Engineer
IT Systems Admin
Data Scientist / Business Analyst
Data
Preparation
Data
Discovery
& Feature
Engineering
Model &
Validate
Predict
Operationalize
Monitoring &
Refinement
Evaluate
& Decide
GPUsMapD H20.ai MapD

12. 12
GOAI DEMO

13. Try MapD
It’s free and it’s easy (and @ortelius sez “it’s the new h0t sh1t”)
13
Play with the live demos:
https://www.mapd.com/demos/
Download the Community Edition:
https://www.mapd.com/platform/download-community/
Join our forums:
https://community.mapd.com/
Review these slides:
https://www.slideshare.net/aaronrogerwilliams

14. Aaron Williams
VP of Global Community
@_arw_
aaron@mapd.com
/in/aaronwilliams/
/williamsaaron