su yuxin jan 20, 2014 petuum: an iterative-convergent distributed machine learning framework

25
SU YUXIN JAN 20, 2014 Petuum: An Iterative- Convergent Distributed Machine Learning Framework

Upload: evan-maxwell

Post on 29-Dec-2015

216 views

Category:

Documents


2 download

TRANSCRIPT

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

SU YUXINJAN 20, 2014

Petuum: An Iterative-Convergent Distributed

Machine Learning Framework

Page 2: SU YUXIN JAN 20, 2014 Petuum: An Iterative-Convergent Distributed Machine Learning Framework

Outline

Introduction

Implementation

Questions

Demo

Page 3: SU YUXIN JAN 20, 2014 Petuum: An Iterative-Convergent Distributed Machine Learning Framework

Introduction to Petuum

Page 4: SU YUXIN JAN 20, 2014 Petuum: An Iterative-Convergent Distributed Machine Learning Framework

Bulk Synchronous Parallel

Page 5: SU YUXIN JAN 20, 2014 Petuum: An Iterative-Convergent Distributed Machine Learning Framework

Asynchronous

Parameters read / update at any time

Page 6: SU YUXIN JAN 20, 2014 Petuum: An Iterative-Convergent Distributed Machine Learning Framework

Stale Synchronous Parallel

Page 7: SU YUXIN JAN 20, 2014 Petuum: An Iterative-Convergent Distributed Machine Learning Framework

Convergence

Page 8: SU YUXIN JAN 20, 2014 Petuum: An Iterative-Convergent Distributed Machine Learning Framework

Programming

read(table, row, col)

inc(table, row, col, value)

iteration()

Page 9: SU YUXIN JAN 20, 2014 Petuum: An Iterative-Convergent Distributed Machine Learning Framework

Implementation

Page 10: SU YUXIN JAN 20, 2014 Petuum: An Iterative-Convergent Distributed Machine Learning Framework

Overview in Logic

Page 11: SU YUXIN JAN 20, 2014 Petuum: An Iterative-Convergent Distributed Machine Learning Framework

Overview in the Real

Page 12: SU YUXIN JAN 20, 2014 Petuum: An Iterative-Convergent Distributed Machine Learning Framework

Main Components

Page 13: SU YUXIN JAN 20, 2014 Petuum: An Iterative-Convergent Distributed Machine Learning Framework

Table

Page 14: SU YUXIN JAN 20, 2014 Petuum: An Iterative-Convergent Distributed Machine Learning Framework

ConsistencyController::DoGet()

Page 15: SU YUXIN JAN 20, 2014 Petuum: An Iterative-Convergent Distributed Machine Learning Framework

ConsistencyController::iterate()

Page 16: SU YUXIN JAN 20, 2014 Petuum: An Iterative-Convergent Distributed Machine Learning Framework

Server::GetRow()

Page 17: SU YUXIN JAN 20, 2014 Petuum: An Iterative-Convergent Distributed Machine Learning Framework

Least-Recently-Used(LRU) Strategy

Page 18: SU YUXIN JAN 20, 2014 Petuum: An Iterative-Convergent Distributed Machine Learning Framework

Questions

Page 19: SU YUXIN JAN 20, 2014 Petuum: An Iterative-Convergent Distributed Machine Learning Framework

Is Lock-Free Possible ?

Data exchange in real-time ? next …

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

Is Auto-Rescheduling Possible ?

sub-centralized server

reduce communication cost

Page 21: SU YUXIN JAN 20, 2014 Petuum: An Iterative-Convergent Distributed Machine Learning Framework

Is Auto-Partition Possible ?

Run ML algorithms like that in a single thread

A Solution for all ML algorithms

Page 22: SU YUXIN JAN 20, 2014 Petuum: An Iterative-Convergent Distributed Machine Learning Framework

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 ….

Page 23: SU YUXIN JAN 20, 2014 Petuum: An Iterative-Convergent Distributed Machine Learning Framework

New Schema to Reduce the Upper Bound?

Page 24: SU YUXIN JAN 20, 2014 Petuum: An Iterative-Convergent Distributed Machine Learning Framework

STRADS Scheduler

Variable Correlations Auto-Parallelization

Dynamic Prioritization Monitor the contribution of variables to objective

function

Load-Balancing in Task

Page 25: SU YUXIN JAN 20, 2014 Petuum: An Iterative-Convergent Distributed Machine Learning Framework

Demo

Switch to my laptop …