A Method for Designing Proximity-Aware Regular Graph-Based Structured Overlay Networks

2019 IEEE 4th International Conference on Computer and Communication Systems, Singapore, February 23-25, 2019

Transcript

1. A Method for Designing Proximity-Aware Regular Graph-Based Structured Overlay Networks Youki Shiraishi†, Akiko Manada‡, Yuzo Taenaka†, Youki Kadobayashi† † Nara Institute of Science and Technology ‡ Shonan Institute of Technology 4th International Conference on Computer and Communication Systems Singapore, February 23-25, 2019 0VUHSPX
   ZPVS
   MJNJUT ແ ݶ ͷ Մ ೳ ੑ ɺ ͕͜͜ ࠷
2. • An application layer network built on top of an existing network (e.g., IP network) ‣ Routes a query to its responsible node ‣ Enables scalable distributed database systems - e.g., distributed hash tables (DHT) Structured Overlay Network 2 Underlying Network Overlay Network Responsible node Query
3. Practical Problem of Overlay Networks • Neighbors are determined without considering proximity ‣ It leads high-delay communication with high probability - e.g., intercontinental communication ‣ It increases lookup latency • Existing solution: Proximity neighbor selection (PNS) ‣ Each node selects their neighbors based on proximity • Issue: It is difficult to apply to regular graph-based overlay networks 3 Overlay Network Singapore USA Underlying Network
4. Contribution • A method of constructing proximity-aware regular graph- based structured overlays in order to reduce lookup latency ‣ Our scheme: Proximity-aware distributed line graphs (PDLG) - Selects an overlay topology from the candidates based on proximity when a node joins/leaves - Simulation result: Lookup latency is improved by 10% 4 … Select from all the possible topology candidates
5. Distributed Line Graphs (DLG) • Directed graphs generated from an initial d-regular graph ‣ Enable to construct a overlay network having a constant routing table size d ‣ The overlay topology is expanded when a new node joins: 1. Generate d virtual nodes if necessary; 2. Split and share them within two physical nodes 5 Responsible node DL iteration Joining node Node splitting
6. Proximity-Aware Distributed Line Graphs (PDLG) • Construct a proximity-aware regular graph-based overlay • Node joins are done by proximity-aware node splitting: ‣ It considers all the possibilities of the expanded topology and construct the best topology in terms of proximity ‣ It locally optimizes the neighbors of the joining node and its responsible node 6 · · · <latexit sha1_base64="zuJOmAyo7rjOtrCBPhBytYpLsOA=">AAAB7XicdVDLSsNAFL3xWeur6tLNYBFchaQIuiy6cVnBPqAtZTKZtGMnmTBzI5TQf3DjQhG3/o87/8ZJW8HngWEO59zLvfcEqRQGPe/dWVpeWV1bL22UN7e2d3Yre/stozLNeJMpqXQnoIZLkfAmCpS8k2pO40DydjC+LPz2HddGqOQGJynvx3SYiEgwilZq9Vio0AwqVc+teQXIb+K7s9+rwgKNQeWtFyqWxTxBJqkxXd9LsZ9TjYJJPi33MsNTysZ0yLuWJjTmpp/Ptp2SY6uEJFLavgTJTP3akdPYmEkc2MqY4sj89ArxL6+bYXTez0WSZsgTNh8UZZKgIsXpJBSaM5QTSyjTwu5K2IhqytAGVLYhfF5K/ietmut7rn99Wq1fLOIowSEcwQn4cAZ1uIIGNIHBLdzDIzw5ynlwnp2XeemSs+g5gG9wXj8AtxWPNQ==</latexit> <latexit sha1_base64="zuJOmAyo7rjOtrCBPhBytYpLsOA=">AAAB7XicdVDLSsNAFL3xWeur6tLNYBFchaQIuiy6cVnBPqAtZTKZtGMnmTBzI5TQf3DjQhG3/o87/8ZJW8HngWEO59zLvfcEqRQGPe/dWVpeWV1bL22UN7e2d3Yre/stozLNeJMpqXQnoIZLkfAmCpS8k2pO40DydjC+LPz2HddGqOQGJynvx3SYiEgwilZq9Vio0AwqVc+teQXIb+K7s9+rwgKNQeWtFyqWxTxBJqkxXd9LsZ9TjYJJPi33MsNTysZ0yLuWJjTmpp/Ptp2SY6uEJFLavgTJTP3akdPYmEkc2MqY4sj89ArxL6+bYXTez0WSZsgTNh8UZZKgIsXpJBSaM5QTSyjTwu5K2IhqytAGVLYhfF5K/ietmut7rn99Wq1fLOIowSEcwQn4cAZ1uIIGNIHBLdzDIzw5ynlwnp2XeemSs+g5gG9wXj8AtxWPNQ==</latexit> <latexit sha1_base64="zuJOmAyo7rjOtrCBPhBytYpLsOA=">AAAB7XicdVDLSsNAFL3xWeur6tLNYBFchaQIuiy6cVnBPqAtZTKZtGMnmTBzI5TQf3DjQhG3/o87/8ZJW8HngWEO59zLvfcEqRQGPe/dWVpeWV1bL22UN7e2d3Yre/stozLNeJMpqXQnoIZLkfAmCpS8k2pO40DydjC+LPz2HddGqOQGJynvx3SYiEgwilZq9Vio0AwqVc+teQXIb+K7s9+rwgKNQeWtFyqWxTxBJqkxXd9LsZ9TjYJJPi33MsNTysZ0yLuWJjTmpp/Ptp2SY6uEJFLavgTJTP3akdPYmEkc2MqY4sj89ArxL6+bYXTez0WSZsgTNh8UZZKgIsXpJBSaM5QTSyjTwu5K2IhqytAGVLYhfF5K/ietmut7rn99Wq1fLOIowSEcwQn4cAZ1uIIGNIHBLdzDIzw5ynlwnp2XeemSs+g5gG9wXj8AtxWPNQ==</latexit> <latexit sha1_base64="zuJOmAyo7rjOtrCBPhBytYpLsOA=">AAAB7XicdVDLSsNAFL3xWeur6tLNYBFchaQIuiy6cVnBPqAtZTKZtGMnmTBzI5TQf3DjQhG3/o87/8ZJW8HngWEO59zLvfcEqRQGPe/dWVpeWV1bL22UN7e2d3Yre/stozLNeJMpqXQnoIZLkfAmCpS8k2pO40DydjC+LPz2HddGqOQGJynvx3SYiEgwilZq9Vio0AwqVc+teQXIb+K7s9+rwgKNQeWtFyqWxTxBJqkxXd9LsZ9TjYJJPi33MsNTysZ0yLuWJjTmpp/Ptp2SY6uEJFLavgTJTP3akdPYmEkc2MqY4sj89ArxL6+bYXTez0WSZsgTNh8UZZKgIsXpJBSaM5QTSyjTwu5K2IhqytAGVLYhfF5K/ietmut7rn99Wq1fLOIowSEcwQn4cAZ1uIIGNIHBLdzDIzw5ynlwnp2XeemSs+g5gG9wXj8AtxWPNQ==</latexit> Current topology Node splitting All the possibilities of the expanded topology Construct the topology which minimizes communication delay Joining node Responsible node
7. Proximity-Aware Node Splitting: Example (1) 7 20 10 Joining node Responsible node Candidate A 10 20 Joining node Responsible node Candidate B Responsible node (merged nodes: 10, 20) Joining node Node
 splitting Current topology (B) (A) 20 10 10 20 Focus on in-neighbors for each merged nodes
8. Proximity-Aware Node Splitting: Example (2) 8 Step 1. Measure the average communication delay of in-neighbors of a merged node from the responsible node Step 2. Measure the average communication delay of in-neighbors of a merged node from the joining node Step 3. Sort merged nodes in ascending order of the difference of the average delay and merge the first half to the responsible node and the other half to the joining node delay = 10 delay = 20 delay = 20 Candidate A delay = 30 delay = 15 delay = 25 Candidate B delay = 10 delay = 20 delay = 30 delay = 20 delay = 15 delay = 25 Constructed! avg. 15 avg. 30 avg. 20 avg. 20 10 20 10 20 diff(10) = 15 − 20 = − 5, diff(20) = 30 − 20 = 10, 20 10 10 20
9. Simulation Preparation • Use the transit-stub model as an underlay topology ‣ L3 network with 5,050 routers ‣ Link communication delays are determined by GT-ITM • Construct two overlays on the underlay topology ‣ DLG-Kautz: Original DLG-enabled overlay ‣ PDLG-Kautz: DLG-enabled overlay based on our scheme 9 Transit node transit node transit domain stub domain stub node Transit domain Transit node Stub domain Stub node (is attached some overlay nodes)
10. Lookup Latency and Merged Nodes (1) • Routing table size: d = 2 • 1,000,000 lookup queries are sent between two nodes 10 Correlation between lookup latency and number of merged nodes (d = 2)
11. Lookup Latency and Merged Nodes (1) 11 Correlation between lookup latency and number of merged nodes (d = 2) Lookup latency is improved by 10% regardless of the number of nodes • Routing table size: d = 2 • 1,000,000 lookup queries are sent between two nodes
12. Lookup Latency and Merged Nodes (2) 12 Correlation between lookup latency and number of merged nodes (d = 16) • Routing table size: d = 16 • 1,000,000 lookup queries are sent between two nodes
13. Lookup Latency and Merged Nodes (2) 13 Correlation between lookup latency and number of merged nodes (d = 16) Lookup latency is improved by 10% when the number of merged nodes is the maximum Lookup reduction rate decreases with decreasing the number of merged nodes • Routing table size: d = 16 • 1,000,000 lookup queries are sent between two nodes
14. Conclusions • Issue: How can we consider proximity on regular graph- based structured overlays? ‣ Our scheme: Consider how node splitting should be done based on proximity on distributed line graphs ‣ We demonstrated that our scheme reduces lookup latency • Future direction ‣ Examine a method of optimizing topology at any time other than node joins/leaves for global optimization • Any questions? 14