cloud streaming

Cloud StreamingJingwen Wang

Video content distribution

Nearly 90% of all the consumer IP traffic is expected to consist of video content distribution- Web video like YouTube, P2P video like BitTorrent

Content distribution requirements:- Scalable and secure media storage, processing and distribution- Anytime, anywhere, any device consumption- Low latency, global distribution

Cloud Provides a Better way

Massive Scale

Rapid File Transfer

Low IT Costs

High Reliability

Accredited Security

CloudStream

Motivation:- Current solution for deliver videos: progressive download via CDN

Non-adaptive codec Video freeezes

- WANT: a SVC based video proxy that delivers high-quality Internet streaming adapting to variable conditions Video transcoding from original formats to SVC Video streaming to different users under Internet dynamics

CloudStream

Implement on one processor:- Video transcoding to SVC is highly complex and transcoding speed

is relatively slow a long duration before a user can access the transcoded video video freezes because of unavailability of transcoded video data

To enable real-time transcoding and allow scalable support for multiple concurrent videos:- Use Cloud: CloudStream

Partition a video into clips and maps them to different compute nodes in order to achieve encoding parallelization

5

Concerns

Encoding parallelization:- Multiple video clips can be mapped to compute nodes at different time- First-task first-server scheme can introduce unbalanced computation

load transcoding jitter- The transcoding component should not speed up video encoding at the

expense of degrading the encoded video quality

Streaming jitter:- Video clips arrive at the streaming component in batches- Demand surge of network resources leads to some data not arrive at

the user at the expected arrival time

6

Metrics affecting Streaming Quality

Streaming Quality:- Access time

Transcoding and streaming latencies- Video freezes

Transcoding and streaming jitters

Video Content:- The temporal motion metric TM- The spatial detail metric SD

Encoding Parallelization

SVC coding structure:- A video non-overlapping coding-independent GOPs- A picture layers- A layer coding-independent slices- A slice macro-blocks

Parallelism- Across different compute nodes: inter-node parallelism- Shared-memory address parallelism inside on compute node:

intra-node parallelism

Multi-level parallelization Scheme

Multi-level encoding parallelization:- GOPs: have the largest work granularity

Inter-node parallelism !

- Slices: independence, relative larger amount of work Intra-node parallelism! Each slice on a different CPU

Intra-node Parallelism

Intra-node Parallelism- Limit the average computation time spend over the GOP to an

upper bound Tth Shorten the access time !

- The minimum number of slices encoded in parallel: Mmin

Notations Definitions

M Number of encoded parallel slices in a picture

NMB, i, Nslice, i Number of MBs or slices in the i-th layer of a picture

TMB, i(M), Tslice, i(M) Average encoding time of one NM or slice in the i-th layer with M parallel slices

Tpic, i(M) Average encoding time of the i-th layer of a picture

Tpic(M) Average encoding time of a picture

TGOP(M) Average encoding time of a GOP

Inter-node Parallelism

Inter-node Parallelism- Achieve real-time transcoding- Transcoding jitters introduced by variation of GOP encoding time- Goal:

Minimize transcoding jitters Minimize the number of compute nodes

Estimation of GOP’s Encoding Time

A multi-variable regression model- At a given encoding configuration- Train videos with different video content characteristics TM and

SD to build the regression model- 90% of predicted values of the testing data are fallen within the

10% of error

Problem Formulation

Problem Formulation- Based on the approximation of each GOP’s encoding time- Given Q jobs- Each job i has a deadline di and a processing time pi

- Multiple nodes in parallel, each job is processed with out preemption on each machine until its completion

- Lateness li can be computed as ci (actual completion time) – di

- Upper bound of lateness: τ- WANT: bound the lateness of these jobs find the

minimal number of machines N and minimize τ

Complexity: NP-hardSolution:

- Hallsh-based Mapping- Lateness-first Mapping

Hallsh-based Mapping

Hallsh-based Mapping(HM): - Set an upper bound of τ and find the minimal number of N satisfies

it- Use Hallsh machine scheduling algorithm as a blackbox

minMS2approx algorithm

1. Pick ε = mini{(di - pi)/τ}

2. Run HallSh by increasing the number of machines until the maximum lateness among all jobs satisfies <(1 + ε) *τ, and set the machine number at this point to be K

3. HallSh will returns the scheduling results of all jobs. For a job with lateness over the upper bound on a particular machine j, move it along with all future jobs on machine K to a new machine K + j. Then compute the new completion time for all jobs on this new machine

4. N is the number of used machines

Lateness-first Mapping

Lateness-first Mapping(LFM):- Compute the minimal number of N based on the deadline of each

job and minimize τ for the given N- Deciding the minimum N:

Tpic(M)*R < SG *N

- Minimizing τ given N: For the i-th job in every N jobs, compute its adjusted processing time

p’i=pi – (di – d1)

Sort the n jobs by the reverse order of p’I

Schedule the job with the largest p’I to the first available compute node,

the second largest one to the second available node

Test

SVC: JSVM

Environment:- Input: 64 480p video GOPs- GOP: 8 pictures- Picture: 4 temporal layers, 2 spatial layers, 1 quality layer- Up tp 4 cores on each compute node- Slices number corresponding to cores

Performance

Average encoding time and speedup using up to 4 cores in intra-node parallelism

LFM

HM

Comparing LFM & HM

HM can successfully decide the appropriate compute node number and limit the transcoding jitters

HM may require greater N in order to achieve the same level of lateness constraint than LFM

Cloud Download

Using Cloud Utilities to achieve high-quality content distribution for unpopular videos

Motivation:- Video content distribution dominates Internet traffic- High-quality video content distribution is of great significance

-1. high data health-2. high data transfer rate

Motivation of Cloud Download

High data health- Data health: number of available full copies of the shared file in a

BitTorrent swarm- Data health < 1.0 is unhealthy- Use data health to represent data redundancy level of a video file

High data transfer rate- Enables online video streaming- Live & VoD

State-of-the-art Techniques: CDN

CDN(Content Distribution Network)- Strategically deploying edge servers- Cooperate to replicate or move data according to data popularity

and server load- User obtains copy from a nearby edge server

CDN: limited storage and bandwidth- Not cost-effective for CDN to replicate unpopular videos the edge

servers- Charged facility only serving the content providers who have paid

State-of-the-art Techniques: P2P

P2P(Peer-to-Peer)- End users forming P2P data swarms- Data directly exchanged between peers- Real strength shows for popular file sharing

P2P: poor performance for unpopular videos- Too few peers

- Low data health- Low data transfer rate

Neither of CDN and P2P work well in distributing unpopular videos, due to low data health or low data transfer rate

Worldwide deployment of cloud utilities provides a novel perspective to solve the problem:Cloud Download!

Cloud Download

Cloud

High data rate !

Cloud Download

Firstly, a user sends video request to the cloud

Subsequently, the cloud downloads the requested video from the file link and stores it in the cloud cache

User retrieve the requested video from the cloud with hight data rate via the intra-cloud data transfer acceleration

User-side energy Efficiency

Commonly download an unpopular video- A common user keeps his computer (& NIC) powered-on for long

hours- Much Energy is wasted while waiting

Cloud download an unpopular video- The user can just be “offline”- When the video is ready, quickly retrieve it in short time- User-side energy efficient!

Cloud Download: View Startup Delay

The only drawback of Cloud Download:- For some videos, the user must wait for the cloud to download it:

View startup delay

This drawback is effectively alleviated- By the implicit and secure data reuse among users- The cloud only downloads a video when it is requested for the first

time: Cloud cache!

- Subsequent requests directly satisfied- Secure because oblivious to users- Data reuse rate -> 87%

System Architecture

Video request

Data download

Data store/cache

Data transfer(high data rate)

Check cache

Component Function

ISP Proxy: receive & restrict requests in each ISP

Task Manager: check cache

Task Dispatcher: load balance

Downloaders: download data

Cloud Cache: store and upload data

Hardware Composition

Building Block # of servers Memory Storage Bandwidth

ISP Proxy 6 8 GB 250 GB 1 Gbps (Intranet), 0.3 Gbps (Internet)

Task Manager 4 8 GB 250 GB 1 Gbps (Intranet)

Task Dispatcher 3 8 GB 460 GB 1 Gbps (Intranet)

Downloaders 140 8 GB 460 GB1 Gbps (Intranet),

0.325 Gbps (Internet)

Cloud Cache400 chunk servers93 upload servers

3 index servers8 GB

4 TB (chunk server), 250

GB (upload server)

1 Gbps (Intranet), 0.3 Gbps (Internet)

Cache Capacity Planning & Replacement Strategy

Handel 0.22M daily requests- Average video size: 379MB- Video cache duration: <7 days- Thus, C=372MB*0.22M*7= 584TB

Cache replacement strategies- 17 days trace-driven simulations- FIFO vs. LRU vs. LFU- FIFO worst, LFU best!

Performance Evaluation

Dataset- Complete running log of the VideoCloud system in 17 days: Jan.1,2011 – Jan. 17, 2011- 3.87M video requests, around 1.0M unique videos

Metrics- Data transfer rate- View startup delay- Energy efficiency

Data transfer rate & View startup delay

Energy Efficiency

User-side energy efficiency- E1: users’ energy consumption using common download- Eu: users’ energy consumption using cloud download- User-side energy efficiency =(E1 - Eu)/E1 = 92%

Overall energy efficiency- Ec: the cloud’s energy consumption- E2: the total energy consumption of the cloud and users, so E2 = Ec

+ Eu

- Overall energy efficiency = (E1 – E2)/E1 = 86%

Cloud Download application

Cloud Transcoding for mobile users- http://xf.qq.com- Mobile user submits a video linnk and the transcoding parameters

to the cloud- The cloud downloads the video from Internet via cloud download- The cloud transcodes the downloaded video and transfers the

transcoded video back to user

References

① Huang et al., Cloudstream: Delivering highquality streaming videos through a cloud-based svc proxy,  INFOCOM 2011

② Huang et al., Cloud download: using cloud utilities to achieve high-quality content distribution for unpopular videos, ACM Multimedia 2011

③ http://www.slideshare.net/AmazonWebServices/aws-for-media-content-in-the-cloud-miles-ward-amazon-web-services-and-bhavik-vyas-aspera

④ The QQCyclone platform. http://xf.qq.com.

Thank you !