1. 1 Dryad Distributed Data-Parallel Programs from Sequential Building Blocks Michael Isard, Mihai Budiu, Yuan Yu, Andrew Birrell, Dennis Fetterly of Microsoft Research, Silicon Valley Presented by: Thomas Hummel

2. 2  Introduction  System Overview  Dryad Graph  Program Development  Program Execution  Experimental Results  Future Work Agenda

3. Introduction  Problem  How to write efficient distributed programs easily?  Environment  Parallel Processors  High Speed Links  Administered Domain  Ignore Low Level Issues 3

4. Introduction  Parallel Execution  Faster Execution  Automatic Specification  Manual Specification  GPU Shader  Distributed Databases  MapReduce 4

5. Introduction 5  Graph Model  Verticies Are Programs  Edges Are Communication Links  Forced Parallelism Mindset  Necessary Abstraction

6. Introduction 6  GPU Shader  Low Level  Hardware Specific  MapReduce  Simplicity Paramount  Performance Sacrificed  Database  Implicit Communication  Algebra Optimized

7. Introduction 7  Dryad  Fine Communication Control  Multiple Input/Output Sets  Must Consider Resources  Execution Engine  Executes DAG Of Programs  Outputs Directed To Inputs  No Recursion

8. System Overview 8  Dryad Job  DAG  Data Passed On Edges  Vertex is a Program  Message Structure  User Defined  Shared Memory  TCP  Files

9. System Overview 9  Dryad Job  DAG  Data Passed On Edges  Vertex is a Program  Message Structure  User Defined  Shared Memory  TCP  Files

10. System Overview 10  System Organization  Job Manager  Name Server  Dameon (Work Nodes)

11. Dryad Graph 11  Graph Description Language  “Embedded” in C++  Combine Sub-Graphs  C++ Class  Inherited By Vertex Program  Program Name  Program Factory

12. Dryad Graph 12  Vertex Creation  C++ Class  Inherited By Vertex Program  Program Name  Program Factory  One Vertex Is a Graph  Factory Called  Program Specific Arguments Applied

13. Dryad Graph 13  Edge Creation  Composition (Combine) Operation  Two Graphs  Varying Assignment Methods

14. Dryad Graph 14

15. Dryad Graph 15  Communication Channel  File I/O By Default  TCP  Shared Memory  Pitfall: Connected Vertices Must Be On Same Process  Deadlock Avoidance  DAG Architecture

16. Program Development 16  Vertex Program Development  C++ Base Classes  Status And Errors Reported to Job Manager  Standard “Main” Method  Channel Readers/Writers  Supplied Via Argument List  Legacy Programs  C++ Wrapper

17. Program Development 17  Pipelined Execution  Assuming Sequential Code  Event Based Programming  Channels Are Asynchronous  Thread Pool  Optimized For Verticies

18. Program Execution 18  Job Manager  Job Ends If JM Machine Fails  Different Schemes Possible To Avoid This  Versioning System For Execution Instances  Vertex Execution  Starts When All Input Channels Ready  User Can Specify Execution Machine  Can Be Re-Run On Failures  Job Ends After All Verticies Have Run

19. Program Execution 19  Fault Tolerance  Re-Run Vertex If Failed  Channel Re-Creation (File Recreation)  TCP/Shared Memory Failures Cause Failures On All Connected Vertices  Staged Execution Allows Intermediate Error Checking

20. Experimental Results 20  SQL Operation  10 Computer Cluster  Gigabit Connections  Data Mining Operation  1800 Computer Cluster  10 TB Data Set  11 Minute Execution Time

21. Future Work 21  Scripting Language  Nebula  Additional Abstraction  SISS Integration  SQL Server Integration  Distributed SQL Queries  Query Optimizer