Presentation is loading. Please wait.

Presentation is loading. Please wait.

SkipNet: A Scaleable Overlay Network With Practical Locality Properties Presented by Rachel Rubin CS294-4: Peer-to-Peer Systems By Nicholas Harvey, Michael.

Similar presentations


Presentation on theme: "SkipNet: A Scaleable Overlay Network With Practical Locality Properties Presented by Rachel Rubin CS294-4: Peer-to-Peer Systems By Nicholas Harvey, Michael."— Presentation transcript:

1 SkipNet: A Scaleable Overlay Network With Practical Locality Properties Presented by Rachel Rubin CS294-4: Peer-to-Peer Systems By Nicholas Harvey, Michael Jones, Stephan Saroiu, Marvin Theimer and Alec Wolman

2 P2P Systems 2003©2003 Rachel Rubin/UC BerkeleySkipNet:2 Philosophy of SkipNet General purpose, scalable, fault tolerant overlay that allows for explicit control over content availability and placement Preserve useful content and path locality Enable load balancing over subset of nodes Provide resiliency against common internet failures

3 P2P Systems 2003©2003 Rachel Rubin/UC BerkeleySkipNet:3 Added Functionality SkipNet adds content locality –Explicitly place data on specific overlay nodes or across nodes within an organization Adds path locality –Guarantee that message traffic is routed between two overlay nodes within the same organization stays within it –Important for availability and security This provides advantages for availability, performance, manageability and security Provides node and data management

4 P2P Systems 2003©2003 Rachel Rubin/UC BerkeleySkipNet:4 Why not a DHT? Controlling Data Location is not the goal of a DHT DHT’s provide load balancing at the price of where data is stored –May be stored far away –May be stored out of the domain Destroy Locality Discard useful application-specific information But what are the benefits of a DHT?

5 P2P Systems 2003©2003 Rachel Rubin/UC BerkeleySkipNet:5 Structure of SkipNet: Perfect Skip Lists Dictionary-like structure stored in memory Sorted linked list with pointers Height of the i’th node is the exponent of the largest power- of-two that divides i The pointers have length of 2 h Searches in O(log N) Perfect Skip List 1 2 3 4 5 6 7 8 9 10 0 1 0 2 0 1 0 3 0 1

6 P2P Systems 2003©2003 Rachel Rubin/UC BerkeleySkipNet:6 Structure of SkipNet: Probabilistic Skip List Each node chooses a height with a probability of choosing height h is 1/2 h Maintains O(log N) Probabilistic Skip List

7 P2P Systems 2003©2003 Rachel Rubin/UC BerkeleySkipNet:7 Structure of SkipNet: Perfect Lexicographic Address Space Maintain a sorted list of all data records as well as “skip” pointers Doubly-linked ring Each node stores2*log N pointers

8 P2P Systems 2003©2003 Rachel Rubin/UC BerkeleySkipNet:8 Structure of SkipNet: Lexicographic Address Space (2) The pointers maintained at each level if a nodes routing table are “rings” Each level skips over 2 h nodes Routing possible in O(log N) Ring membership determined by their unique random ID number – probabilistic optimization Duplicate pointer elimination from the rou\ting table improves performance by around 20%

9 P2P Systems 2003©2003 Rachel Rubin/UC BerkeleySkipNet:9 Numeric Space vs. Lexicographic Space Node names and content identifier strings are mapped into the Lexicographic space –Distributed generalization of Skip Lists –Support content placement and path localities Hashes of node names and content identifiers are mapped into the numeric space –Supports DHT functionality –Routing between overlay nodes –Allows the functionality of Chord

10 P2P Systems 2003©2003 Rachel Rubin/UC BerkeleySkipNet:10 Structure of SkipNet: Lexicographic Searching Follow the pointers that route closest to the intended destination Route along the highest-level pointer that doesn’t point past the destination Terminate when the message arrives at the node closest to the lexicographic name

11 P2P Systems 2003©2003 Rachel Rubin/UC BerkeleySkipNet:11 Structure of SkipNet: Node Join Find the top-level ring that corresponds to the random ID by searching through rings Join the top-level ring and the the next lower ring, etc. until the zero-level ring –Nodes do not acknowledge the newcomer until this is complete

12 P2P Systems 2003©2003 Rachel Rubin/UC BerkeleySkipNet:12 Structure of SkipNet: Node Departure Can route correctly as long as the bottom level ring is maintained Repair upper-level rings lazily Each node maintains a leaf-set that points to additional nodes along the bottom ring

13 P2P Systems 2003©2003 Rachel Rubin/UC BerkeleySkipNet:13 Useful Properties: Locality Incorporating a peer’s address into a content name guarantees the content will be hosted there –Naming is not arbitrary Routing locality can be guaranteed if you use naming prefixes

14 P2P Systems 2003©2003 Rachel Rubin/UC BerkeleySkipNet:14 Useful Properties: Constrained Load Balancing Dividing a data object’s name into two parts –Specifies the set of nodes over which DHT leaf balancing should be performed –Input to the DHT hash function Searches for the name prefix in lexicographic space Switches to numeric space to search for the hash of the name suffix Limitations –Can be performed over any naming subtree, not over an arbitrary subset of the nodes –Domain of load-balancing is encoded in the name, so transparent remapping is impossible

15 P2P Systems 2003©2003 Rachel Rubin/UC BerkeleySkipNet:15 Useful Properties: Fault Tolerance Generally node failures in overlay systems are not independent –Nodes belonging to the same organization tend to fail together Independent failures –Maintain level zero –Lazy stabilization Organization Boundaries –Since you preserve locality, this results in partitions instead of scattered link failures –Partitions remerged via the leaf sets

16 P2P Systems 2003©2003 Rachel Rubin/UC BerkeleySkipNet:16 Useful Properties: Range Queries Inherits the functionality and flexibility in supporting efficient range queries from skip lists Keys are stored in lexicographic order –Documents sharing common prefixes are stored over contiguous ring segments Range query answering is routing along a ring segment

17 P2P Systems 2003©2003 Rachel Rubin/UC BerkeleySkipNet:17 Results: Methodology Measured –Relative Delay penalty –Physical Network Hops –Number of failed lookups Uniform, clustered and Zipf-clustered

18 P2P Systems 2003©2003 Rachel Rubin/UC BerkeleySkipNet:18 Results: Basic Routing Costs ChordBasic SkipNet Full SkipNet Pastry 16.341.773.563.2 Unique Routing Entries Relative Delay Penalty vs. Network Size

19 P2P Systems 2003©2003 Rachel Rubin/UC BerkeleySkipNet:19 Results: Locality of Placement Absolute lookup request latency vs. Access Locality

20 P2P Systems 2003©2003 Rachel Rubin/UC BerkeleySkipNet:20 Results: Fault Tolerance Failed Lookup Requests vs. Data Access Locality

21 P2P Systems 2003©2003 Rachel Rubin/UC BerkeleySkipNet:21 Results: CLB RDP of CLB data vs. Network Size

22 P2P Systems 2003©2003 Rachel Rubin/UC BerkeleySkipNet:22 Additional Optimizations: Efficient Routing Proximity Table –Entry points to a node chosen from the ring segment whose endpoints are defined by two consecutive basic routing table entries Search in this table except when violates the CLB search constraints

23 P2P Systems 2003©2003 Rachel Rubin/UC BerkeleySkipNet:23 Additional Optimizations: Virtual Nodes Many virtual nodes on a physical node Routing tables can be shared

24 P2P Systems 2003©2003 Rachel Rubin/UC BerkeleySkipNet:24 Design Alternatives Virtualize the overlay nodes per data object Two-part naming scheme –Virtual node name –Node-relative name Path locality added to the DHT –Not guaranteed

25 P2P Systems 2003©2003 Rachel Rubin/UC BerkeleySkipNet:25 Discussion Questions What does locality buy? What are the disadvantages to SkipNet? How does this differ from Chord and Pastry?


Download ppt "SkipNet: A Scaleable Overlay Network With Practical Locality Properties Presented by Rachel Rubin CS294-4: Peer-to-Peer Systems By Nicholas Harvey, Michael."

Similar presentations


Ads by Google