Presentation is loading. Please wait.

Presentation is loading. Please wait.

SU YUXIN JAN 20, 2014 Petuum: An Iterative-Convergent Distributed Machine Learning Framework.

Similar presentations


Presentation on theme: "SU YUXIN JAN 20, 2014 Petuum: An Iterative-Convergent Distributed Machine Learning Framework."— Presentation transcript:

1 SU YUXIN JAN 20, 2014 Petuum: An Iterative-Convergent Distributed Machine Learning Framework

2 Outline Introduction Implementation Questions Demo

3 Introduction to Petuum

4 Bulk Synchronous Parallel

5 Asynchronous Parameters read / update at any time

6 Stale Synchronous Parallel

7 Convergence

8 Programming read(table, row, col) inc(table, row, col, value) iteration()

9 Implementation

10 Overview in Logic

11 Overview in the Real

12 Main Components

13 Table

14 ConsistencyController::DoGet()

15 ConsistencyController::iterate()

16 Server::GetRow()

17 Least-Recently-Used(LRU) Strategy

18 Questions

19 Is Lock-Free Possible ? Data exchange in real-time ? next …

20 Is Auto-Rescheduling Possible ? sub-centralized server reduce communication cost

21 Is Auto-Partition Possible ? Run ML algorithms like that in a single thread A Solution for all ML algorithms

22 In-Memory or In-Storage ? Data capacity is greater than memory size. Memory should be a cache for disk storage. Solution for disk storage:  Hadoop  Spark  ….

23 New Schema to Reduce the Upper Bound?

24 STRADS Scheduler Variable Correlations  Auto-Parallelization Dynamic Prioritization  Monitor the contribution of variables to objective function Load-Balancing in Task

25 Demo Switch to my laptop …


Download ppt "SU YUXIN JAN 20, 2014 Petuum: An Iterative-Convergent Distributed Machine Learning Framework."

Similar presentations


Ads by Google