OMFS An Object-Oriented Multimedia File System for Cluster Streaming Server CHENG Bin, JIN Hai...

Preview:

Citation preview

OMFSAn Object-Oriented Multimedia File System

for Cluster Streaming Server

CHENG Bin, JIN HaiCluster & Grid Computing Lab

Huazhong University of Science & TechnologyHPCAsia, BeiJing, China, Dec. 2005

2

OutlineBackground

Our Object-Oriented Method

Improvement Schemes

Performance Evaluation

Conclusion

3

Background

From Centralized Streaming Server to Cluster Streaming Server.

Single Server Cluster Server

4

BackgroundExisting cluster streaming servers

I/O BottleneckPoor Scalability

Cluster File SystemBlock-based File OperationsHave to care about the media file format

5

BackgroundExisting multimedia file systemsTiger Shark, PVFS A general cluster file system

HERMES, Symphony focus on a QoS-aware disk scheduling algorithm, data placement, and

cache policies.

EXT3NS ( MM’05) a local multimedia file system, not special for cluster streaming server

7

BackgroundThe features of media filesVarious of file formats, such mp4, mov, aviDifferent files have different objection organizationsBe consist of logical object units

The features of cluster streaming serverOnly perform reading operationsSeek objections according to timestamp.Retrieve objects, such as meta object, data objects

8

Our goalsAn object-oriented multimedia file system specially designed for cluster streaming server To provide an object-oriented method to fetch the media data To present a single system image To mask differences of multimedia file format To separate the processing of multimedia files from cluster

streaming servers To enhance the performance of cluster streaming server.

9

Object Oriented MethodTraditional processing method

avi

mov

rm

Cluster Streaming Server

Parser_avi

Parser_mov

Parser_rm

BlockParsing

Object

RTP Server

RTSP Server

consume a lot of CPU and memory resource, not efficient

10

Object Oriented Method

avi

mov

rm

Cluster Streaming Server

Parser_avi

Parser_mov

Parser_rm

RTP Server

RTSP Server

DB

Meta Objects

RTP Ojbects

Uniform Interface

Pre-processing

Reading

Apply the rule: Make the common case faster

11

Objects

Original File

MPEG 1

MPEG 4

MOV

AVI

PARSER A

PARSER B

PARSER C

PARSER D

MetaData Objects

Data Objects

Metadata Servers

Data Servers

Control Servers

Data Servers

Writing Interface

Reading Interface

Object-Oriented Uniform Interface

(OMFS)

SDP, Poster Picture, Track Infor, Time, Size, Location,

Directory, and so on

RTP Packet

Parsing Phrase Writing Phrase Reading Phrase

Web ServersHTTP

RTP

RTSP

Object-Oriented Framework

12

Optimization SchemesBypassing kernel buffer

Data Server

NIC DISK

1

23

4

Data Server

NIC DISK

12

DataSend Command

Kernel Module

Bypassing

13

Optimization SchemesCreating time-based block indexing map

Write Read

Data Object Data Object Data Object Data Object

Key Frame

Map

14

Performance EvaluationSimulation Setup

Our cluster streaming server has one Web Server, one Control Server, eight Data Servers, and one OMFS admin node.

---Two 1.4 GB AMD Opteron CPU---2 GB Memory---1000Mb/s Network Card---avi, mp4, wmv Format

A multithreads simulator to generate requests, which accord with Poisson Distribution.

15

Performance EvaluationMetrics Max Concurrent Stream Number Max Throughput CPU Overhead

Average VCR Waiting Time

16

Simulation ResultsMax Concurrent Stream Number

0 1 2 3 4 5 6 7 8 90

500

1000

1500

2000

2500

3000

3500

Co

ncu

rre

nt

Str

ea

m N

um

be

r

Data Server Number

without OMFS OMFS without Bypassing OMFS with Bypassing

about 12% improvement

17

Simulation ResultsMax Throughput

0 1 2 3 4 5 6 7 8 90

500

1000

1500

2000

2500

3000

3500

4000

Thr

ough

put (

Mb/

s)

Data Server Number

without OMFS OMFS without Bypassing OMFS with Bypassing

18

Simulation ResultsCPU Overhead

0 2 4 6 8 10 12 14 160

20

40

60

80

100C

PU

Util

iza

tion

(%

)

Time (minute)

without OMFS OMFS without Bypassing OMFS with Bypassing

19

Simulation ResultsDelay of VCR Operation

Streaming Server TypeTotalTime

(Sec.)AverageTime

(Sec.)

Darwin Streaming Server 1050 10.5

Our Cluster streaming Server with Time-based Indexing Map

647 6.4

20

ConclusionOMFS has the following features: Simplify the design of cluster streaming server Improve its performance optimization schemes have great positive effect to the

performance improvement of cluster streaming server.

21

That is all, Thank you!

22

Overview of OMFSNetwork

Client

SchedulerWeb

Server

RTP Stream

Control ServerMetadata Server

Data Server

MetaData

Data Slice

OMFS Admin

Movie Files

23

Inherited Tree of Objects

Object

MetaDataObject

DataObject

S IDT

S: SDP

T: Track Info

P: Preview Picture

ID: Index Data

RH: RTP HEAD

RD: RTP Packet DataP RD RH

24

Uniform Interface

SP T OMFS_WriteRHRD OMFS_Encoder

OMFS_Decoder

Block Write

Block Read

SP TRHRD OMFS_Read

Input Sequence

Output Sequence

Local File System

Transfer

OMFS_Write(int object_type, void * object_content)

OMFS_Read(int object_type, void * object_content)

Recommended