Budapest Big Data Meetup: GPU Analytics

The goal of this talk is to deliver an introduction to Graphical Processing Units (GPUs) and their application to general purpose analytics. There is a great deal of excitement and hype around GPU computing, with good reason. The implications of GPU technology for machine and deep learning have already been enormous. But GPUs can do more than train neural networks - and there is a blossoming open source community to prove it. Aaron will discuss an application of GPU computing: the use of GPUs for performing super fast database operations on big data.

Transcript

1. GPU Analytics Budapest Big Data Meetup | Hungary | October 8, 2018 slides: https://speakerdeck.com/mapd
2. © OmniSci 2018 © OmniSci 2018
3. © OmniSci 2018 Aaron Williams VP, Global Community at OmniSci @_arw_ aaron@omnisci.com /in/aaronwilliams/ /williamsaaron
4. The Fastest Software Designed for the Fastest Hardware HARNESS GPUs
5. GPU Processing CPU Processing 40,000 Cores 20 Cores *fictitious example Latency Throughput CPU 1 ns per task (1 task/ns) x (20 cores) = 20 tasks/ns GPU 10 ns per task (0.1 task per ns) x (40,000 cores) = 4,000 task per ns Latency: Time to do a task. | Throughput: Number of tasks per unit time.
6. © OmniSci 2018 GPU Processing Keeps Moore’s Law Alive
7. © OmniSci 2018 7 * open source for single node github.com/omnisci/mapd-core
8. D E M O https://www.omnisci.com/demos/
9. © OmniSci 2018 9 GPU Parallelism Drives Fast Analytics at Scale High Memory Bandwidth Native Rendering Pipeline Supercomputer Processing
10. © OmniSci 2018 10 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 Advanced Memory Management
11. © OmniSci 2018 11 MapD Core: Query Compilation with LLVM 10111010101001010110101101010101 00110101101101010101010101011101 Traditional DBs can be highly inefficient • Each operator in SQL treated as a separate function • Incurs tremendous overhead and prevents vectorization OmniSci 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
12. © OmniSci 2018 12 GPU Data Frames /mapd/pymapd /gpuopenanalytics/pygdf OMNISCI CORE
13. © OmniSci 2018 TOP-TIER VENTURE BACKING USED BY 100+ GLOBAL ORGS $37 MILLION IN FUNDING OPEN-SOURCE COMMUNITY About OmniSci 13
14. © OmniSci 2018 © OmniSci 2018 • omnisci.com/demos Play with our demos - everything demo you saw in this talk was live! • omnisci.cloud Get an OmniSci instance in 60 seconds • omnisci.com/platform/downloads/ Download the Community Edition • community.omnisci.com Ask questions and share your experiences Next Steps
15. © OmniSci 2018 Aaron Williams VP, Global Community at OmniSci @_arw_ aaron@omnisci.com /in/aaronwilliams/ /williamsaaron Thank you! Any Questions?