Synchronization II

1. Synchronization II CSE5306 Lecture Quiz 17 due at 5 PM Tuesday, 7 October 2014

2. 6.3.5 A Token Ring Algorithm • A token ring network can make a resource mutually exclusive. • A token circulates around the ring till a needy process can briefly hold it, while using the resource. • The protocol is simple, fast, and it never starves a process. When a crashed process cannot acknowledge its token passing message, the sender throws the token over it to the next process. • But the token can be lost when its holder crashes, and using a simple expiration timer risks the possibility that the holder actually is still using the token.

3. 6.3.6. Comparing the Algorithms • The centralized algorithm is simple and efficient, requiring only 3 messages to enter and exit a critical region. Assuming brief resource usage, access is delayed by only 2 messages. Crashes can bring down system. • The decentralized algorithm makes k attempts to send/rcv messages to/from m coordinators. Its access delay is 3mk. It is immune to crashes, but starvation and efficiency are problems. • The distributed algorithm sends request and grant messages to its n-1 peers, with access delay 2(n-1). Crashes can bring it down. • The token ring algorithm may waste hours sending token-passing messages, while no one is using the resource. Access delay is from 1 to n-1. Crashes can bring down system.

4. 6.4 Global Positioning Nodes • Geometric overlay networks organize large numbers of communicating (possibly mobile) nodes by their physical positions: • A client may be redirected to a server’s closest replicate, assuming everyone’s position is known. • Optimal replica placement minimizes average client-to-replica response time. • Position-based routing uses only local information in forwarding messages to waypoints closest to the destination.

5. Global Positioning (continued) • Contacting three nodes can fix a mobile client’s position, from their measured message latencies, d = √(xi-xp)2+(xi-xp)2, i=1..3 • If inconsistent distances violate the triangle inequality, use only two (landmark) nodes at a time, and minimize their cumulative error. • Or model equal-tension springs from every node to the client.

6. R U O K ? • Which mutual exclusion algorithm is most reliable? __ • Centralized. • Decentralized. • Distributed. • Token ring.

7. R U O K ? 2. A client contacts three nearby nodes at (1,2),(2,3) and (3,1), measuring messages latencies of 6, 4 and 2 microseconds. What is the client’s position? __ • (11,3) b. (-11,3) c. (-11,-3) d. (11,-3)

8. R U O K ? Match the following algorithms with their definitions below. 3. Centralized algorithm. __ 4. Decentralized algorithm. __ 5. Distributed algorithm. __ 6. Token ring algorithm. __ • Others grudgingly grant resource access, if my timestamp is earlier than theirs. • Many replicated resources and their DHT-based coordinators allocated around a Chord ring. • Ask the coordinator process’ permission to use the resource. • Message circulates till a needy process briefly holds it, while using the resource.

9. R U O K ? Match the following terms with their definitions below. 7. Starved. __ 8. Geometric overlay network. __ 9. Optimal replica placement. __ 10. Position-based routing. __ 11. Triangle inequality. __ 12. Landmarks. __ • Uses only local information in forwarding messages to waypoints closest to the destination. • Many pairs of nodes, which may be used to triangulate position. • A process with reduced access a resource. • Side c < a + b. • Large numbers of communicating (possibly mobile) nodes organized by their physical positions. • Minimizes average client-to-replica response time.

10. 6.5 Election Algorithms • An election algorithm chooses one of many nodes to serve as coordinator. • All processes know each other’s node identification numbers, but they do not know which nodes are currently active. • Traditionally the bully and ring algorithms were most popular. • Wireless environments and large-scale systems call for more sophistication.

11. 6.5.1 Traditional Election Algorithms • Bully algorithm: • When node P finds its coordinator is unresponsive, it sends an ELECTION message to all higher numbered nodes. • Each receiving node answers “OK” and recursively performs steps 1-3. • If nobody answers, P sends a victory (COORDINATOR) message and becomes the new coordinator. • Crashed nodes also hold elections, when they restart. • The name comes from the biggest numbered node always winning.

12. Traditional Elections (continued) • Ring Algorithm: • When node P finds its coordinator is unresponsive, it circulates an ELECTION message around the ring. • Each receiving node responds and adds its number to the circulating message. • When P receives its own ELECTION message, it circulates a COORDINATOR message around the ring to inform everyone that it is the new coordinator. • If two nodes do this simultaneously, messages circulated by the lower-numbered node will have the last word.

13. 6.5.2 WiFi Elections • Traditional elections don’t work in wireless networks, where message passing is unreliable and topologies change. WiFi elections choose the best candidate: • A node that receives an ELECTION message for the first time calls the sender its parent. • It Acks its parent’s message, telling its battery life and other resources. • And it relays the parent’s message to its other neighbors. • When other nodes call it their parent and Ack its message, it returns their resource reports to its parent. • Eventually the initiator designates the most resourceful node as their new coordinator. • Each node joins only one of many concurrent elections.

14. 6.5.3 Large-Scale System Elections • Massive distributed systems elect superpeers: • Low-latency access to normal nodes. • Even distribution across the overlay network. • Predefined ratio to the total number of nodes. • Serve more than a fixed number of normal nodes. • In structured DHT-based systems, we can identify a superpeer as number 11100000 and its nodes as 111ΦΦΦΦΦ, where Φ=don’t care. • In unstructured gossip-based systems, we can randomly distribute tokens. The tokens repel each other, such that each token forces nearby tokens away. Every token eventually settles on a superpeer.

15. 6.6 Summary • Synchronization causes communicating processes to “do the right thing at the right time,” without a globally shared clock. • Processes synchronize clocks by exchanging clock values and accounting for propagation delays. • Latencies can be estimated from separation distances. • Lamport showed that ordering is more important than real time. His timestamps extend to vector timestamps that order events by causality. • A coordinator can ensure a resource’s mutual exclusion; i.e., only one process’ access at a time. • Election algorithms safely replace crashed coordinators and select superpeers.

16. R U O K ? Match the following election algorithms with their definitions below. 13. Bully. __ 14. Ring. __ 15. WiFi. __ 16. Superpeer. __ • Highest numbered node always wins. • Lowest numbered node always has the last word. • Initiator chooses most resourceful node as coordinator. • Zeroth node identifier in each series marks DHT-based coordinators. Repulsive tokens settle on multiple coordinators in randomly unstructured systems.