1. Efficient, Proximity-Aware Load Balancing for DHT-Based P2P Systems Yingwu Zhu, Yiming Hu Appeared on IEEE Trans. on Parallel and Distributed Systems, vol. 16, no. 4, April 2005 Presented by Ki

2. 2 Outline Introduction System Design Proximity-Aware Load Balancing Experimental Evaluations Conclusions

3. 3 Introduction DHT-based P2P systems like Chord, Pastry, Tapestry, CAN, …  Provide a distributed hash table abstraction for object storage and retrieval  Assume nodes are homogeneous Two main drawbacks  Imbalance load  Node heterogeneity

4. 4 Introduction Existing solutions for DHT load balancing  Some ignore node heterogeneity  Some ignore proximity information This work proposes a proximity-aware load balancing scheme which considered the above two aspects

5. 5 System Design - Basic Virtual Server (VS)  Act like an autonomous peer  Responsible for a contiguous portion of the DHT’s identifier space  Responsible for certain amount of load Physical Peer  Can host multiple virtual servers Heavily loaded physical peer  Moves some of its virtual servers to other lightly loaded physical peers to achieve load balancing  Movement of virtual server can be viewed as a leave operation followed by a join operation

6. 6 System Design – Four Phrases The load balancing scheme proceed in 4 phrases  Load balancing information (LBI) aggregation Collect load and capacity information of whole system  Node classification Classify nodes into HEAVY, LIGHT, NEUTRAL  Virtual server assignment (VSA) Determine the VS assignment to make HEAVY nodes becomes LIGHT Proximity-aware  Virtual server transferring (VST)

7. 7 System Design – k-ary Tree on DHT A k-ary tree (KT) is built on top of the DHT  Occupying the same identifier space KT root node is responsible for the entire identifier space Each child node is responsible for a portion of their parent’s identifier space

8. 8 System Design – k-ary Tree on DHT A KT node is responsible for identifier space region  key & host is obtained by procedure plant_KT_node()  Keeps track of its parent and children KT node, X, is planted into a virtual server which responsible for X.key Example (KT Node X, VS S):  X.region = (3,5]  X.key = 4  S’s region = (3, 6]  X is planted in S

9. 9 System Design – k-ary Tree on DHT For KT node X, its region  is further divided into k parts, then taken by its k children Until…  X’s region is completely covered by its hosting VS Each KT node periodically check if its region is completely covered by its VS  Yes  delete the existing children  No  keep k children

10. 10 Load Balancing Information Aggregation Load Balancing Information (LBI) of node i  L i  total load of VS in node i  C i  capacity of node i  L i,min  minimum load of VS in node i X.host randomly responds to one of its VS only

11. 11 Load Balancing Information Aggregation KT root node obtains the system-wide LBI  L  Total load  C  Total capacity  L min  minimum load of all VS in system KT root node distribute the system-wide LBI  Along the tree, back to the leaf nodes, VS and finally the DHT node

12. 12 Node Classification System-wide utilization = L / C Utilization of node i = L i / C i Define T i = (L / C + ε) * C i  ε is a parameter for trade off between amount of load movement and quality of load balance Classification  HEAVY node if L i > T i  LIGHT node if (T i – L i ) >= L min  NEUTRAL node if 0 <= (T i – L i ) < L min

13. 13 Virtual Server Assignment Each HEAVY DHT node i  Randomly choose a subset of its VS that minimizes s.t.  Minimized the amount of load movement  VSA info.: … Each LIGHT DHT node j  VSA info.: This VSA information propagates upward along the KT

14. 14 Virtual Server Assignment Proximity Ignorant Each KT node i  Collects the VSA information until the a pairing_threshold  Uses a best-fit heuristic to reassign VS Reassign VS in HEAVY nodes to LIGHT nodes And minimize load movement  DHT nodes of reassigned VS get notified while the rest of VSA information propagate to i’s parent

15. 15 Virtual Server Assignment Proximity Aware Use landmark clustering  Measure distance to a number of landmark nodes  Obtain a landmark vector which represent point in a m-dimensional space  Nodes with close landmark vectors are in general physically close Transform the landmark m-dimensional space to the DHT identifier space and obtains a DHT key, LM i  By Hilbert curve, i.e. N m  N  Proximity preserving

16. 16 Virtual Server Assignment Proximity Aware Each node i independently determines its landmark vector and its corresponding DHT key, LM i Node publish its VSA information in the DHT network with the DHT key LM i  Node j that responsible for the region contains LM i receive the VSA information  Node j propagate the received VSA information into the KT

17. 17 Virtual Server Transferring Upon receiving the reassigned VSA info.  HEAVY node transfer the reassigned VS to the LIGHT node The transfer of VS would cause KT to reconstruct Lazy migration  Reconstruction of KT only after the completion of all transfers

18. 18 Experimental Evaluations Load Balancing Information aggregation and virtual server assignment latency

19. 19 Experimental Evaluations Underlying network  Generated by GT-ITM with about 5000 nodes Underlying DHT overlay  Chord with 4096 nodes  5 virtual servers in each node, exponential identifier space k-ary tree  k = 2 Pairing threshold  50 Landmark node count  15

20. 20 Experimental Evaluations Nodes carry loads proportional to their capacities by reassigning virtual servers

21. 21 Experimental Evaluations Cumulative distribution of moved load  Proximity-aware 36% within 2 hops, 57% within 10 hops  Proximity-ignorant 17% within 10 hops Proximity-aware  Reduce load balancing cost  Fast and efficient load balancing

22. 22 Experimental Evaluations Effect of node churn (join & leave) Overhead = (M d – M s ) / M s  M d : number of VSA messages with node churn  M s : number of VSA messages without node churn

23. 23 Conclusions This work focuses on an efficient, proximity-aware load balancing scheme  Align load distribution and node capacity  Use proximity information to guide load reassignment and transferring