Continuous Deep Analytics
0 likes•244 views
The document discusses the integration of diverse workloads in modern data pipelines, emphasizing critical decision-making and the importance of efficiency in processing. It highlights concepts such as cross-platform computation, data dependencies, and the implementation of intermediate representations for optimized execution. The content includes various techniques, algorithms, and the current research team involved in advancing deep analytics and machine learning capabilities.
Continuous Deep Analytics
- 1. RISE Research Institutes of Sweden knowledge Computer Systems Laboratory Paris Carbone - [email protected] PhD Candidate @ KTH Committer @ Apache Foundation CONTINUOUS DEEP ANALYTICS RISE Open House 2018
- 2. !2 knowledge Data it is raining it is cold those who listen to A also like B someone is crossing the street credit fraud detected users hate the website redesign water/light sensor events temperature sensor events app-user likes web user clicks vehicle camera feed user data and transactions a storm is approachingclimate data and simulations PROCESSING
- 3. !3 knowledge Decision Making it is raining bring an umbrella it is cold turn heating on those who listen to A also like B recommend B to fans of A someone is crossing the street stop the vehicle credit fraud detected cancel transactions users hate the website redesign switch to the old version a storm is approaching alert boats! REASONING
- 4. !4 complexity > data dependencies < response time time criticality
- 5. !5 > data dependencies it is raining it is cold those who listen to A will love B someone is crossing the street credit fraud people hate the new menu a storm approaches < response time complexity time criticality
- 6. !6 CEP bulk synchronous iterative processing deep neural networks graph analysis derivative approximation simulations > data dependencies < response time complexity time criticality relational algebra online ML
- 7. CEP relational algebra online ML !7 bulk synchronous iterative processing deep neural networks graph analysis derivative approximation simulations > data dependencies < response time complexity time criticality Tensor Programming Platforms (GPUs/TPUs) Batch Processing Platforms (mem/CPU) High Performance Computing Platforms (petaflops) Data Stream/DBMS Platforms (mem/CPU)
- 8. CEP relational algebra online ML !7 bulk synchronous iterative processing deep neural networks graph analysis derivative approximation simulations > data dependencies < response time complexity time criticality Tensor Programming Platforms (GPUs/TPUs) Batch Processing Platforms (mem/CPU) High Performance Computing Platforms (petaflops) Data Stream/DBMS Platforms (mem/CPU) Continuous Deep Analytics
- 9. ▪ Modern Data Pipelines need to combine diverse workloads! (ML Training & Serving, Relational Algebra, Streams, Tensors, Graphs) !8 ⋈ ⋈ ⋈ σθ σθ σθ σθ π π Relational Data Streams Feature Learning Tensor Programming Dynamic Graphs
- 10. !9 Hardware Acceleration: Important but Not Enough Pipeline (CPU) Pipeline (GPU/TPU) time until decision critical decision making
- 11. !10 Cross-Platform Computation is Inefficient Stream Tasks Tensor Tasks Graph Tasks computationcomputation computation
- 12. !10 Cross-Platform Computation is Inefficient Stream Tasks Tensor Tasks Graph Tasks computationcomputation computation - No computation sharing optimisations
- 13. !10 Cross-Platform Computation is Inefficient Stream Tasks Tensor Tasks Graph Tasks computationcomputation computation - expensive data exchange through disk - No computation sharing optimisations
- 14. !11 Computation Sharing f f’…. ….Data knowledge
- 15. !11 Computation Sharing f f’…. ….Data knowledge
- 16. !11 Computation Sharing f f’ f + f’ loop fusion vectorization …. +less data passes +no additional overhead …. ….Data knowledge
- 17. !12 Critical Decision Making demands Efficiency time until decision Pipeline (CPU) Pipeline (GPU/TPU) critical decision making
- 18. !12 Critical Decision Making demands Efficiency Pipeline (CPU) - Optimised Pipeline (GPU/TPU) - Optimised time until decision Pipeline (CPU) Pipeline (GPU/TPU) critical decision making
- 19. !13 The Problem f f’ ? …. ….Data knowledge
- 20. !14
- 21. !15 Using an Intermediate Representation (IR) f f’…. ….Data knowledge
- 22. !15 Using an Intermediate Representation (IR) f f’…. ….Data knowledge f+f’ IR IR IR f f’
- 23. !16 1) Weld IR (Stanford DAWN Project) Related IR Projects Matei Zaharia (Spark architect) et. al. !16
- 24. !16 1) Weld IR (Stanford DAWN Project) Related IR Projects + supports large number of existing libraries - currently limited to short-lived local task execution Matei Zaharia (Spark architect) et. al. !16
- 25. !17 2) Tensor Comprehensions (Facebook) Related IR Projects
- 26. !17 2) Tensor Comprehensions (Facebook) Related IR Projects + JIT compiler, multi-GPU autotuner - overspecialized to tensor operations in CUDA
- 27. !18 ⋈ σθ σθ π Dynamic Graphs Tensors Relational Streams Data Pipeline Our Goals: Model and Language Independence
- 28. !18 ⋈ σθ σθ π Dynamic Graphs Tensors Relational Streams Data Pipeline Our Goals: Model and Language Independence
- 29. !18 ⋈ σθ σθ π Dynamic Graphs Tensors Relational Streams Data Pipeline Our Goals: Model and Language Independence …
- 30. !19 A Distributed Runtime for Heterogeneous Hardware CDA Platform Worker Worker Worker LLVM code Rust-Based
- 32. !21 Intermediate Representation (IR) Logically Optimised IR Physically Optimised IR Compilation Steps
- 33. !22 Physically Optimised IR Distributed JIT Compilation and Reconfiguration Rust AST LLVM
- 34. !23 The Current CDA Team (RISE SICS + KTH) Computer Systems Machine Learning Lars Kroll Paris Carbone Christian Schulte Seif Haridi Theodore Vasiloudis Daniel Gillblad MSc Students • Klas Segeljakt • Oscar Bjuhr • Johan Mickos
- 35. !24 Thank You! Don’t miss our poster :)