Songqing Chen, Bo Shen, Yong Yan, Sujoy Basu, and Xiaodong Zhang

SRB: Shared Running Buffers SRB: Shared Running Buffers in Proxy to Exploit Memory Locality in Proxy to Exploit Memory Locality

of Multiple Streaming Media Sessionsof Multiple Streaming Media Sessions

Songqing Chen, Bo Shen, Yong Yan, Sujoy Basu, and Xiaodong Zhang

IEEE Proceedings of the 24th International Conference on Distributed Computing Systems (ICDCS’04)

July 13th, 2006Seo, Dongmahn

Media Service Team 2/30

ContentsContents Introduction Shared Running Buffers (SRB) Media Caching

Algorithm Performance Evaluation Conclusion Presenter’s Opinion


ContentsContentsIntroductionIntroduction Shared Running Buffers (SRB) Media Caching

Algorithm Performance Evaluation Conclusion Presenter’s Opinion


IntroductionIntroduction Web content delivery network

Server-proxy-client systemproxy

reduce network load and end-to-end latencycache, static

The delivery of streaming media content larger sizecontinuous and timely deliverya lot of resources have to be reserved

high bandwidth of disk and network


Related worksRelated works Partial caching

cache either a prefix or segments of a media object in the disk storage of the proxy

Dynamic caching of streaming data memory-based running buffer based caching tech.

1. the fixed-sized running buffer caching fixed-sized buffer to cache media data fetched by the proxy hoping that subsequent requests could reuse

2. the interval caching consider the arrival intervals of all request pairs The smaller intervals, the more requests with given memory


Related works Related works (cont.)(cont.)

Above bothdo NOT take consideration of user access patterns


Related works Related works (cont.)(cont.)

Other related workspatching algorithms

take advantage of the client-side storage resourcethe greedy patching

patch to the existing complete stream

the grace patching restarts a new complete stream at some appropriate points in time

the optimal patching consider how to reuse the limited storage on the client side to receive

as many data as possible while listening to as many channels as possible


Works of this paperWorks of this paper propose Shared Running Buffers (SRB)

a new memory-based caching algorithm for streaming media objects

based on user access patterns and the requested media objectsdifferent buffers are fully shareduseful when requests are highly temporally localizedefficiently reclaim the idle memory spaceefficient near-optimal buffer replacement at runtime when

requests are terminated

propose Patching SRB (PSRB) improves the performance of the media data delivery

without the necessity of caching


ContentsContents Introduction

Shared Running Buffers (SRB) Media Shared Running Buffers (SRB) Media Caching AlgorithmCaching Algorithm

Performance Evaluation Conclusion Presenter’s Opinion


SRB Media Caching AlgorithmSRB Media Caching Algorithm SRB (Shared Running Buffer)

Motivationthe limits of current memory buffering approaches

based caching algorithm for streaming media with the aim to maximize the memory utilization


SRB Related ConceptsSRB Related Concepts Concepts

Interval SeriesAverage Request Arrival Interval (ARAI)

three buffer states and three timing conceptsConstruction State & Start-TimeRunning State & Running-Distance Idle State & End-Time


SRB AlgorithmSRB Algorithm working flow

If the latest running buffer of the object i is caching the prefix of the object i, the request will be served directly from all the existing running buffers of the object

OtherwiseIf there is enough memory, a new running buffer of a

predetermined size T is allocatedIf there is no enough memory, the SRB buffer replacement

algorithm is called to either re-allocate an existing running buffer to the request or serve this request without caching

Update the End-Times of all existing buffers of the object i


SRB Algorithm SRB Algorithm (cont.)(cont.)

SRB Buffer Lifecycle Management Starting from the Construction State a three-state lifecycle management process

Case 1 : in the Construction State. The proxy makes a decision at the end of T If ARAI = ∞, initial buffer enters the Idle State (case 3) immediately If In (waiting time) > ARAI, the initial buffer is shrunk If In ≤ ARAI, the initial buffer maintains the construction state and continues to grow buffer expansion is bounded by the available memory



Case 2 : in the Running State.

Case 3 : in the Idle State. When a buffer enters the Idle State, it is ready for reclamation.

time span T (the initial buffer size) is determinedbuffer has upper bound and lower bound



SRB Buffer Dynamic Reclamation Two types of session termination

the complete session termination If the resident buffer is the only running buffer for the media object, the

resident buffer enters the idle state If the resident buffer is not only running buffer, the buffer enters the idle state

and maintains its content until its End-Time the premature session termination

If one request is served from the middle of it resident buffer, the resident buffer maintains its current state and the request get deleted from all its associated running buffers.

If one request is served from the head of its resident buffer, the request is deleted from all of its associated running buffers.

If one request is served at the tail of a running buffer, deleting the request from the request list of its resident buffer, if the

request list is not empty, do nothing If the request is at the tail of the last running buffer, the buffer will be

shrunk to the extent that previous R is the last request served from the buffer.



SRB Buffer Replacement Policypropose popularity based replacement policies

for the SRB media caching algorithmbasic idea

ordering on-going streams according to their popularities calculated in a certain past time period

if demanded object has a higher popularity that the least popular object, deallocated and re-allocated


Patching SRB (PSRB)Patching SRB (PSRB)Media Delivering AlgorithmMedia Delivering Algorithm SRB is extended to a PSRB

for sharing of such sessions using more client-side storage to maximize the data sharing among concurrent sessions to minimize the server-to-proxy traffic



Algorithm

Performance EvaluationPerformance Evaluation Conclusion Presenter’s Opinion


Performance EvaluationPerformance Evaluation event-driven simulator to model a proxy’s memory

caching behavior Both synthetic workloads and a real workload from

enterprise media server logs real workload as REAL

HP Corporate Media Solutions April 1 to April 10, 2001 403 total objects unique object size accounts to 20G 9000 requests, run for 916427 seconds 83% requests only view the objects for less than 10 minutes 56% requests only view the objects for less than 10% Only about 10% requests view the whole objects


Evaluation MetricsEvaluation Metrics the server traffic reduction

neither object hit ratio nor hit ratio measure the traffic

between proxy and clients

HP workstation x4000 with 1GHz CPU and 1GB memory running Linux Redhat 7.1

simulation a set of seven algorithms in three groups

buffering schemes the running buffer caching and the interval caching

patching algorithms the greedy patching, the grace patching and the optimal patching

two shared running buffer algorithms proposed in this paper


Performance ResultsPerformance Results


Performance Results Performance Results (cont.)(cont.)



Algorithm Performance Evaluation

ConclusionConclusion Presenter’s Opinion


ConclusionConclusion propose two new algorithms for efficiently

delivering streaming media objectsShared Running Buffers (SRB) caching algorithm

to dynamically cache media objects in the proxy memoryPatching SRB (PSRB) algorithm

to enhance the memory utilization in the proxyshow simulation results



Algorithm Performance Evaluation Conclusion

Presenter’s OpinionPresenter’s Opinion


Presenter’s OpinionPresenter’s Opinion in the paper

Server-Proxy-Client system modelsimulation VS implementationmedia data characteristicsserver topology

How about to studyadapt to VCR-like functions

Fast Forward, Fast Rewind, Pause, Resume and so on

Songqing Chen, Bo Shen, Yong Yan, Sujoy Basu, and Xiaodong Zhang

Documents

Transcript of Songqing Chen, Bo Shen, Yong Yan, Sujoy Basu, and Xiaodong Zhang