Efficient, ProximityAware Load Balancing for DHTBased P2P Systems

1
Efficient, Proximity-Aware Load Balancing for
DHT-Based P2P Systems
2
Motivations

• DHT's offer uniform distribution
• But resources on servers are not uniform
• Ignores heterogeneous nature of P2P systems
• Previous load-balancing techniques
• Ignored heterogeneity
• Ignored proximity
• Sometimes both

3
Goals

• Achieve load balance by allowing nodes with
  higher capacity to be responsible for more load
• Use proximity to assign load (in terms of virtual
  servers) in order to minimize transfer time
• Perform load assignments and load transfer in an
  efficient manner

4
Overview of System

• Four phases
• Load balancing information (LBI) aggregation
• Node classification
• Virtual server assignment
• Virtual server transfer

5
Distributed Tree

• LBI distribution is performed by building a
  distributed k-ary tree
• Each node is responsible for a space of the DHT
• Each node is stored on the virtual server
  responsible for the center of the region for
  which the node is responsible
• This region is further divided until the host of
  a node completely covers the node's region

6
Tree Functions

• Each node performs routine functions to maintain
  its state
• check_KT_node
• delete_KT_children
• add_KT_children

7
LBI Aggregation

• Each node requests LBI from its children
• If it is a leaf-node, then it asks its virtual
  server to report the LBI
• LBI ltTotal Load, Total Capacity, Min. Loadgt
• This LBI is propagated up the tree
• The root contains the system-wide LBI
• The system-wide LBI is sent back down the tree

8
Node Classification

• Target load (L/C e)Ci
• Node classification
• Heavy if Load gt Target load
• Light if (Target - Load) gt Minimum load
• Neutral if 0 lt (Target Load) lt Minimum load

9
Virtual Server Assignment

• Heavy node
• Chooses a minimal subset of its virtual servers
  that need to be moved in order to make the server
  light
• Reports this information to host node
• Light node
• Reports change in load necessary to reach target
  load

10
Virtual Server Assignment

• VSA information is collected into a pool
• If pool reaches a threshold, the current node
  will perform the virtual server assignments
• The assignments are given to children
• Remaining VSA information is propagated to host

11
Virtual Server Transfer

• Heavy node receives transfer information
• Virtual server is transferred
• Tree is restructured after transfer
• VSA and VST can overlap

12
Proximity

• Oops what happened to proximity?
• Current DHT is proximity-ignorant
• We need to add proximity information
• How?

13
Adding Proximity

• Use proximity to produce DHT keys
• Use landmark nodes to find proximity
• Use space-filling curves to map m-dimensional
  space vector to 1-dimensional key
• VSA information is first mapped into this DHT
• VSA information is now mapped into physically
  close locations

14
Gaussian Distribution, Gnutella-like Capacity
15
Pareto Distribution, Zipf-like Capacity
16
Proximity ts5k-large
Pareto/Zipf
Gaussian/Gnutella
17
Proximity ts5k-small
Pareto/Zipf
Gaussian/Gnutella