Wide Area Placement of Data Replicas for Fast and Highly Available - - PowerPoint PPT Presentation

▶
wide area placement of data replicas
SMART_READER_LITE
LIVE PREVIEW

Wide Area Placement of Data Replicas for Fast and Highly Available - - PowerPoint PPT Presentation

Mar 23, 2024 175 likes •636 views

Wide Area Placement of Data Replicas for Fast and Highly Available Data Access Fan Ping Xiaohu Li, Christopher McConnell Rohini Vabbalareddy, Jeong-Hyon Hwang State University of New York - Albany Outline Background Network Coordinate

slide-1
SLIDE 1

Fan Ping

Xiaohu Li, Christopher McConnell Rohini Vabbalareddy, Jeong-Hyon Hwang

State University of New York - Albany

Wide Area Placement of Data Replicas for Fast and Highly Available Data Access

slide-2
SLIDE 2

Outline

  • Background
  • Network Coordinate System
  • Data Replication
  • Data Replication for Performance
  • Data Replication for Performance and Availability
  • Conclusion
slide-3
SLIDE 3

Outline

  • Background
  • Network Coordinate System
  • Data Replication
  • Data Replication for Performance
  • Data Replication for Performance and Availability
  • Conclusion
slide-4
SLIDE 4

Data Intensive Distributed Systems

  • Google, Amazon, Facebook, Microsoft…
slide-5
SLIDE 5

Data Intensive Distributed Systems

  • Google, Amazon, Facebook, Microsoft…
  • Dynamo, Cassandra, PNUTS…
slide-6
SLIDE 6

Data Replica Placement

  • Given a replication degree (e.g., 3), where

should we put those data replicas in order to effectively improve the overall data access speed and availability?

  • Challenges
  • Scalability
  • Certain SLA
slide-7
SLIDE 7

Outline

  • Background
  • Network Coordinate System
  • Data Replication
  • Data Replication for Performance
  • Data Replication for Performance and Availability
  • Conclusion
slide-8
SLIDE 8

Network Coordinate Systems

  • Based on the network latencies between each other,

nodes are embedded into a virtual space so that their distances in this virtual space are close to the network latencies.

  • E.g., Vivaldi, RNP
slide-9
SLIDE 9

Network Coordinate Systems

slide-10
SLIDE 10

Network Coordinate Systems

slide-11
SLIDE 11

Network Coordinate Systems

slide-12
SLIDE 12

Network Coordinate Systems

slide-13
SLIDE 13

Network Coordinate Systems

slide-14
SLIDE 14

Network Coordinate Systems

slide-15
SLIDE 15
slide-16
SLIDE 16
slide-17
SLIDE 17
slide-18
SLIDE 18
slide-19
SLIDE 19
slide-20
SLIDE 20
slide-21
SLIDE 21

Outline

  • Background
  • Network Coordinate System
  • Data Replication
  • Data Replication for Performance
  • Data Replication for Performance and Availability
  • Conclusion
slide-22
SLIDE 22

Outline

  • Background
  • Network Coordinate System
  • Data Replication
  • Data Replication for Performance
  • Data Replication for Performance and Availability
  • Conclusion
slide-23
SLIDE 23

Servers on the map

slide-24
SLIDE 24

Servers in the coordinate system

slide-25
SLIDE 25

Clients in the coordinate system

slide-26
SLIDE 26

Cluster the clients in the coordinate system

slide-27
SLIDE 27

Cluster the clients in the coordinate system

slide-28
SLIDE 28

Centroids of the clusters

slide-29
SLIDE 29

Servers near centroids of the clusters

slide-30
SLIDE 30

Simulation Settings

  • Java simulator
  • ~200 Planetlab-node trace as input
  • A certain number of nodes are selected as

servers

  • The other nodes are used as clients
slide-31
SLIDE 31

Performance VS. Number of Replicas

slide-32
SLIDE 32

Outline

  • Background
  • Network Coordinate System
  • Data Replication
  • Data Replication for Performance
  • Data Replication for Performance and Availability
  • Conclusion
slide-33
SLIDE 33

Conditional Failure vs. Angle

C R S1 S2

slide-34
SLIDE 34

Conditional Failure vs. Angle

C R S1 S2

0.05% 0.05%

slide-35
SLIDE 35

Conditional Failure vs. Angle

C R S1 S2

0.05% 0.05%

The conditional probability of the failure of (C, R, S2) given the failure of (C,R,S1) is more than

50%!!

slide-36
SLIDE 36

Conditional Failure vs. Angle

C S1 S2

Ɵ

(0,-20) (-10,50) (30,35)

slide-37
SLIDE 37

Conditional Failure vs. Angle

slide-38
SLIDE 38

Conditional Failure vs. Angle

slide-39
SLIDE 39

Conditional Failure vs. Angle

Ɵ

slide-40
SLIDE 40

Conditional Failure vs. Angle

slide-41
SLIDE 41

Estimations for Latency and Availability

  • Per-client latency

L(c, S) = dist(c,s)

  • Per-client availability

A(c,S) = 1- ( F(c,S1)*F(c,S2|S1)*…*F(c,Sr|Sr-1) )

  • Utility function to combine latency and

availability

U =

𝑩 𝑴

slide-42
SLIDE 42

Simulation Settings

  • Java simulator
  • Traceroute and ping data collected from ~

100 PlanetLab nods for a month

  • Randomly select some nodes as servers
  • The rest are clients
slide-43
SLIDE 43

Unavailability vs. Number of Replicas

slide-44
SLIDE 44

Outline

  • Background
  • Network Coordinate System
  • Data Replication
  • Data Replication for Performance
  • Data Replication for Performance and Availability
  • Conclusion
slide-45
SLIDE 45

Conclusion and Future Work

  • Improves the average user access latency by 35%
  • Improves the overall availability
  • Designs the utility function to take into account both

latency and availability

  • Needs more realistic dataset
  • Better utility function
  • Non-exponential algorithm (Greedy…)
  • Take inter-datacenter cost into account
slide-46
SLIDE 46

THANK YOU!

Questions?