Parallelizing Video Transcoding With Load Balancing On Cloud Computing Song Lin, Xinfeng Zhang, Qin...

Post on 31-Mar-2015

214 views 1 download

Tags:

transcript

Parallelizing Video Transcoding With Load Balancing On Cloud Computing

Song Lin, Xinfeng Zhang, Qin Y, Siwei Ma

Circuits and Systems, 2013 IEEE

Outline

Introduction

Related work

Problem formulation and system architecture

Proposed method

Experiment Results

Conclusion

Introduction #1

Parallel programmingShare memory

Pthread – data dependency

Message passing MPI – time delay

Introduction #2

IssuesData dependencyCost of data passingLoad balance

Introduction #3

Cloud computationData segmentationComputing capacity

Introduction #4

GOP-based encodingIndependence between GOPs

...........

Introduction #5

Paralleling in GOP-based

Thread1

Thread2

Thread3

Related work #1

FCFS - First come first server [2]Easy to implementLoad balancing problem is still exist

Related work #3

MCT – Minimal complete time [6]Map-Reduce-based

Problem formulation and system architecture #1 Load balance problem on cloud computing

Executing timeDelay time

Data passing

C is complexity and P is computing capacity

Problem formulation and system architecture #2 The overall completion time of set Sk is

.

Goal .

Problem formulation and system architecture #3 Optimal solution

.

n means n task and m means m cores

Problem formulation and system architecture #4 Flow chart of the proposed method

Problem formulation and system architecture #5For video coding, if the input sequence has

instantaneous decoder refresh (IDR) frame, this video coding task can be divided into sub-tasks.[7]

Problem formulation and system architecture #6For complexity estimation of video transcoding

tasks, the existing algorithms [8] [9] can be utilized.

Proposed method #1

The framework includes three modulesTask pre-analysisAdaptive threshold segmentationMinimal finish time

Proposed method #2

The threshold of segmentation

Proposed method #3

Proposed method #4

The optical finish time

The finish time

Proposed method #5

Assign all the tasks sequentially in descending complexity order

For each unassigned task j, the cores are judged in their descending computing capacity order according to the following criterion: assuming the task j is assigned to core k, if Τκ ≤ Tthr, the assignment is verified. Otherwise, we will judge the next core.

Proposed method #6

If all the cores are traversed and all the computing time are beyond Tthr, the task j will be assigned by MCT algorithm. and Tthr is updated to be the new finish time of the received core Tk

Proposed method #7

Experiment results #1

Experiment results #2

Experiment results #3

Conclusion

Load balancing problem is a NP-hard problem.The proposed algorithm has strong robustness to

the task launching delay.