10
Scalable Adaptive Data Dissemination Under Heterogeneous Environment Yan Chen, John Kubiatowicz and Ben Zhao UC Berkeley

Scalable Adaptive Data Dissemination Under Heterogeneous Environment

  • Upload
    kynton

  • View
    40

  • Download
    3

Embed Size (px)

DESCRIPTION

Scalable Adaptive Data Dissemination Under Heterogeneous Environment. Yan Chen, John Kubiatowicz and Ben Zhao UC Berkeley. Dissemination Tree of OceanStore. Using Application-Level Multicast for Data Dissemination. Scalability Fault-tolerance Efficiency Adaptability. Our Solutions. - PowerPoint PPT Presentation

Citation preview

Page 1: Scalable Adaptive Data Dissemination Under Heterogeneous Environment

Scalable Adaptive Data Dissemination Under Heterogeneous Environment

Yan Chen, John Kubiatowicz and Ben Zhao

UC Berkeley

Page 2: Scalable Adaptive Data Dissemination Under Heterogeneous Environment

Dissemination Tree of OceanStore

Page 3: Scalable Adaptive Data Dissemination Under Heterogeneous Environment

Using Application-Level Multicast for Data Dissemination

• Scalability

• Fault-tolerance

• Efficiency

• Adaptability

Page 4: Scalable Adaptive Data Dissemination Under Heterogeneous Environment

Our Solutions• Use Tapestry, Distributed

Overlay Routing & Location Infrastructure– Randomized data structure

with search locality– Insensitive to faults, and

self-repairable– Ease-of-maintenance

http://www.cs.berkeley.edu/~ravenben/tapestry.pdf

Page 5: Scalable Adaptive Data Dissemination Under Heterogeneous Environment

Our Solutions (Cont’d)

• Dedicated Infrastructure– Complemented with intelligent replica placement

• Application-Level Semantics & Optimization– Dynamic transmission (selective dissemination)– Dynamic notification of updates

Page 6: Scalable Adaptive Data Dissemination Under Heterogeneous Environment

Dissemination Tree Construction• Client Contact Statistically Closest Server Which

Has the Data Through Tapestry– Autonomous decision - Path & load piggybacked– If client unsatisfied with QoS, server dynamically

replicate data close to client

• Model the Replica Placement as “Minimal Set Covering” Problem– Each server covers certain subset of clients (w.r.t.

certain QoS, latency, bandwidth, etc.)– Approximate the solution with greedy algorithm– Distributed load balancing

Page 7: Scalable Adaptive Data Dissemination Under Heterogeneous Environment

RealCast Tree Management Protocols

• Bi-directional Messaging– Heartbeat message stream from root to clients– Refresh message from children to parent

• Scalability– Each member only maintains states for direct children and

parent– “Join” request can be handled by any member

• Continuous Self-tuning and Auto-repair– Periodically check for better parent– Topology-aware through Wide-area Network Measurement

and Monitoring Services (WNMMS)

Page 8: Scalable Adaptive Data Dissemination Under Heterogeneous Environment

• Selective Dissemination

• Dynamic update notification– Quantitative analytic model for

dynamically choosing between poll (pull everytime), push invalidate and push update based on access/update pattern and clients’ preferences to certain metrics (e.g. average latency)

App-Level Semantics and Optimization

push invalidate

push update

Page 9: Scalable Adaptive Data Dissemination Under Heterogeneous Environment

0

10

20

30

40

50

60

0.1 0.5 1

W/R ratio

av

era

ge

la

ten

cy

(m

se

c) 100%

reducing # ofmsgs

50% reducing# of msgs,50% reducinglatencyunoptimized

–Number of messages reduced by 10 - 20%–Average response latency reduced by 30 - 40%

Preliminary Simulation•Topology generated with GT-ITM (120 nodes)

•Synthetic hot-cold pattern workload (100 objects)

•Base-line is the push invalidate from Cao/Liu paper

Page 10: Scalable Adaptive Data Dissemination Under Heterogeneous Environment

Conclusions• Dissemination Tree: Large-scale Data

Dissemination with App-level Multicast• Key Techniques

– Use distributed location services, Tapestry– “Minimal Set Covering with Load Balancing” for

replica/service placement– App-level semantics and optimization

• Preliminary Results– Feasibility of the infrastructure– Flexibility and effectiveness of app-level optimization