SkipNet: A Scalable Overlay Network with Practical Locality Properties Nick Harvey, Mike Jones, Stefan Saroiu, Marvin Theimer, Alec Wolman Microsoft Research University of Washington

Overlay Networks Overlays have achieved several goals: Overlays have achieved several goals:  Scalable and decentralized infrastructure  Uniform and random load and data distribution But, at the price of data controllability But, at the price of data controllability  Data may be stored far from its users  Data may be stored outside its domain  Local accesses leave local organization Basic trade-off: data controllability vs. data uniformity Basic trade-off: data controllability vs. data uniformity SkipNet: SkipNet:  Traditional overlay functionality  Provides an abstraction to control this trade-off:  Constrained load balancing (CLB)

Talk Outline Practical data locality requirements Practical data locality requirements Basic SkipNet design Basic SkipNet design SkipNet locality properties SkipNet locality properties Performance evaluation Performance evaluation Conclusions Conclusions

Talk Outline Practical data locality requirements Practical data locality requirements Basic SkipNet design Basic SkipNet design SkipNet locality properties SkipNet locality properties Performance evaluation Performance evaluation Conclusions Conclusions

Key Locality Properties and Abstraction In practice, two properties are important: In practice, two properties are important:  Content Locality – ability to explicitly place data  Placement on a single node or on a set of nodes  Path Locality – ability to guarantee that local traffic remains local One abstraction is important – CLB: One abstraction is important – CLB:  SkipNet abstraction to control the trade-off  Multiple DHT scopes within one single overlay

Practical Requirements Data Controllability: Data Controllability:  Organizations want control over their own data  Even if local data is globally available Manageability: Manageability:  Data control allows for data administration, provisioning and manageability  Data center/cluster = constrained set of nodes  CLB ensures load balance across data center/cluster

Practical Requirements (cont’d) Security: Security:  Content and path locality are key building blocks for dealing with certain external attacks Data availability Data availability  Local data survives network partitions Performance Performance  Data can be stored near clients that use it

Talk Outline Practical data locality requirements Practical data locality requirements Basic SkipNet design Basic SkipNet design SkipNet locality properties SkipNet locality properties Performance evaluation Performance evaluation Conclusions Conclusions

SkipNet Key property: two address spaces Key property: two address spaces 1. Name ID space: nodes are sorted by their names (e.g. DNS names) 2. Numeric ID space: nodes are randomly distributed Combining both spaces achieves Combining both spaces achieves  Content + Path locality  Other uses could emerge: range queries [AS ’03] Scalable peer-to-peer overlay network Scalable peer-to-peer overlay network  O(log N) routing performance in both spaces  O(log N) routing state per node

SkipNet Ring Pointers at level h skip over 2 h nodes Pointers at level h skip over 2 h nodes Nodes are ordered by names Nodes are ordered by names A D M V T X Z O

SkipNet Ring Pointers at level h skip over 2 h nodes Pointers at level h skip over 2 h nodes Nodes are ordered by names Nodes are ordered by names A D M V T X Z O

SkipNet Ring Pointers at level h skip over 2 h nodes Pointers at level h skip over 2 h nodes Nodes are ordered by names Nodes are ordered by names A E F M H S Z G

SkipNet Global View A Level: L = 0 L = 1 L = 3 L = 2 Root Ring Ring 0 Ring 1 Ring 00 Ring 01 Ring 10 Ring 11 Ring 000 Ring 000 Ring 001 Ring 001 Ring 010 Ring 010 Ring 011 Ring 011 Ring 100 Ring 100 Ring 101 Ring 101 Ring 110 Ring 110 Ring 111 Ring 111 D M O T V X Z O Z AT M X D V A T M X D V Z O O Z A T M X D V

SkipNet Global View A Level: L = 0 L = 1 L = 3 L = 2 Root Ring Ring 0 Ring 1 Ring 00 Ring 01 Ring 10 Ring 11 Ring 000 Ring 000 Ring 001 Ring 001 Ring 010 Ring 010 Ring 011 Ring 011 Ring 100 Ring 100 Ring 101 Ring 101 Ring 110 Ring 110 Ring 111 Ring 111 D M O T V X Z O Z AT M X D V A T M X D V Z O O Z A T M X D V

Two Address Spaces SkipNet can route efficiently in both address spaces: SkipNet can route efficiently in both address spaces:  Name ID space (e.g. DNS names)  Numeric ID space

Routing by Name ID Level: L = 0 L = 1 L = 2 Example: route from A to V Example: route from A to V Simple Rule: Forward the message to node that is closest to dest, without going too far. Simple Rule: Forward the message to node that is closest to dest, without going too far. Ring 00 Ring 01 Ring 10 Ring 11 Ring 000 Ring 000 Ring 001 Ring 001 Ring 010 Ring 010 Ring 011 Ring 011 Ring 100 Ring 100 Ring 101 Ring 101 Ring 110 Ring 110 Ring 111 Ring 111 A Root Ring D M O T V X Z Ring 0 A M T X Ring 1 D Z V O O Z AT M X D V A T M X D V Z O L = 3 Node A’s Routing Table Node A’s Routing Table

Routing by Name ID Level: L = 0 L = 1 L = 2 Example: route from A to V Example: route from A to V Simple Rule: Forward the message to node that is closest to dest, without going too far. Simple Rule: Forward the message to node that is closest to dest, without going too far. Ring 00 Ring 01 Ring 10 Ring 11 Ring 000 Ring 000 Ring 001 Ring 001 Ring 010 Ring 010 Ring 011 Ring 011 Ring 100 Ring 100 Ring 101 Ring 101 Ring 110 Ring 110 Ring 111 Ring 111 A Root Ring D M O T V X Z Ring 0 A M T X Ring 1 D Z V O O Z AT M X D V A T M X D V Z O L = 3

Routing by Name ID Level: L = 0 L = 1 L = 2 Example: route from A to V Example: route from A to V Simple Rule: Forward the message to node that is closest to dest, without going too far. Simple Rule: Forward the message to node that is closest to dest, without going too far. Ring 00 Ring 01 Ring 10 Ring 11 Ring 000 Ring 000 Ring 001 Ring 001 Ring 010 Ring 010 Ring 011 Ring 011 Ring 100 Ring 100 Ring 101 Ring 101 Ring 110 Ring 110 Ring 111 Ring 111 A Root Ring D M O T V X Z Ring 0 A M T X Ring 1 D Z V O O Z AT M X D V A T M X D V Z O L = 3 Node T’s Routing Table Node T’s Routing Table

Routing by Name ID Level: L = 0 L = 1 L = 2 Example: route from A to V Example: route from A to V Simple Rule: Forward the message to node that is closest to dest, without going too far. Simple Rule: Forward the message to node that is closest to dest, without going too far. Ring 00 Ring 01 Ring 10 Ring 11 Ring 000 Ring 000 Ring 001 Ring 001 Ring 010 Ring 010 Ring 011 Ring 011 Ring 100 Ring 100 Ring 101 Ring 101 Ring 110 Ring 110 Ring 111 Ring 111 A Root Ring D M O T V X Z Ring 0 A M T X Ring 1 D Z V O O Z AT M X D V A T M X D V Z O L = 3 Node T’s Routing Table Node T’s Routing Table

Routing by Name ID Level: L = 0 L = 1 L = 2 Example: route from A to V Example: route from A to V Simple Rule: Forward the message to node that is closest to dest, without going too far. Simple Rule: Forward the message to node that is closest to dest, without going too far. Ring 00 Ring 01 Ring 10 Ring 11 Ring 000 Ring 000 Ring 001 Ring 001 Ring 010 Ring 010 Ring 011 Ring 011 Ring 100 Ring 100 Ring 101 Ring 101 Ring 110 Ring 110 Ring 111 Ring 111 A Root Ring D M O T V X Z Ring 0 A M T X Ring 1 D Z V O O Z AT M X D V A T M X D V Z O L = 3 Node T’s Routing Table Node T’s Routing Table

Routing by Name ID Level: L = 0 L = 1 L = 2 Example: route from A to V Example: route from A to V Simple Rule: Forward the message to node that is closest to dest, without going too far. Simple Rule: Forward the message to node that is closest to dest, without going too far. Ring 00 Ring 01 Ring 10 Ring 11 Ring 000 Ring 000 Ring 001 Ring 001 Ring 010 Ring 010 Ring 011 Ring 011 Ring 100 Ring 100 Ring 101 Ring 101 Ring 110 Ring 110 Ring 111 Ring 111 A Root Ring D M O T V X Z Ring 0 A M T X Ring 1 D Z V O O Z AT M X D V A T M X D V Z O L = 3

Routing by Numeric ID Provides the basic DHT primitive Provides the basic DHT primitive To store file “Foo.c” To store file “Foo.c”  Hash(“Foo.c”)  a random numeric ID  Find highest ring matching that numeric ID  Store file on node in that ring Log N routing efficiency Log N routing efficiency

DHT Example Store file “Foo.c” from node A Store file “Foo.c” from node A  Hash(“Foo.c”) = 101… Route from A to V in numeric space Route from A to V in numeric space Level: L = 0 L = 1 L = 2 Ring 00 Ring 01 Ring 10 Ring 11 Ring 000 Ring 000 Ring 001 Ring 001 Ring 010 Ring 010 Ring 011 Ring 011 Ring 100 Ring 100 Ring 101 Ring 101 Ring 110 Ring 110 Ring 111 Ring 111 A Root Ring D M O T V X Z Ring 0 A M T X Ring 1 D Z V O O Z AT M X D V A T M X D V Z O L = 3 Foo.c

Talk Outline Practical data locality requirements Practical data locality requirements Basic SkipNet design Basic SkipNet design SkipNet locality properties SkipNet locality properties Performance evaluation Performance evaluation Conclusions Conclusions

Constrained Load Balancing (CLB) Multiple DHTs with differing scopes using a single SkipNet structure Multiple DHTs with differing scopes using a single SkipNet structure A result of the ability to route in both address spaces A result of the ability to route in both address spaces Divide data object names into 2 parts using the ‘!’ special character Divide data object names into 2 parts using the ‘!’ special character CLB Domain CLB Suffix CLB Domain CLB Suffix microsoft.com ! skipnet.html microsoft.com ! skipnet.html Numeric RoutingName Routing

CLB Example To read file “com.microsoft ! skipnet.html” To read file “com.microsoft ! skipnet.html”  Route by name ID to “com.microsoft”  Route by numeric ID to Hash(“skipnet.html”) within the “com.microsoft” constraint com.sun edu.ucb gov.irs com.microsoft skipnet. html

SkipNet Path Locality Organizations correspond to contiguous SkipNet segments Organizations correspond to contiguous SkipNet segments  Internal routing by NameID remains internal Nodes have left / right pointers Nodes have left / right pointers com.sun edu.ucb gov.irs com.microsoft com.microsoft.research

Fault Tolerance Many failures occur along organizational boundaries: Many failures occur along organizational boundaries:  Gateway/firewall failure, BGP misconfig, physical network cut, … SkipNet handles organizational disconnect gracefully SkipNet handles organizational disconnect gracefully  Results in two well-connected, partitioned SkipNets  Efficient remerging algorithms Node independent failures Node independent failures  Same resiliency as systems such as Chord and Pastry  Similar approach to repair (Leaf Set)

Primary Security Benefit & Weakness + SkipNet + name access control mechanism:  Content locality ensures that content stays within organization  Path locality prevents:  malicious forwarders  analysis of internal traffic  external tampering - Easier to target organizations:  Someone creates one million nodes with name prefixes microsofa.com and microsort.com  Most traffic to/from Microsoft will go through a microsofa / microsort intermediate node

Talk Outline Practical data locality requirements Practical data locality requirements Basic SkipNet design Basic SkipNet design SkipNet locality properties SkipNet locality properties Performance evaluation Performance evaluation Conclusions Conclusions

Methodology Packet-level, event-driven simulator: Packet-level, event-driven simulator:  SkipNet implementation  Basic SkipNet  Full SkipNet = Basic SkipNet + network proximity  Pastry and Chord implementation Uses Mercator and GT-ITM network topologies Uses Mercator and GT-ITM network topologies Experimentally evaluated: Experimentally evaluated:  Name ID routing performance  Tolerance to organizational disconnect

Methodology Packet-level, event-driven simulator: Packet-level, event-driven simulator:  SkipNet implementation  Basic SkipNet  Full SkipNet = Basic SkipNet + network proximity  Pastry and Chord implementation Uses Mercator and GT-ITM network topologies Uses Mercator and GT-ITM network topologies Experimentally evaluated: Experimentally evaluated:  Name ID routing performance  Tolerance to organizational disconnect  Numeric ID routing performance  Effectiveness of network proximity optimizations  Effectiveness of CLB routing optimizations

Routing by Name ID Performance Benefits come at no extra cost

Surviving Organizational Disconnect Disconnected Org Size = 15% of all nodes

Conclusions SkipNet : SkipNet :  Traditional overlay functionality  Explicit control of data placement  Constrained load balancing Content + Path Locality are basic ingredients to: Content + Path Locality are basic ingredients to:  Data controllability  Manageability  Security  Data availability  Performance

Questions?