Distributed Algorithms

Transcription

1 Course Outline With grateful acknowledgement to Christos Karamanolis for much of the material Jeff Magee & Jeff Kramer Models of distributed comuting Synchronous message-assing distributed systems Algorithms in systems with no failures The commit roblem Consensus roblems Asynchronous message-assing distributed systems Logical time and global system snashots Imossibility of consensus Fault-tolerant broadcasts Partially synchronous message-assing distributed systems Failure detectors 2 Bibliograhy, Nancy Lynch, Morgan Kaufmann, Distributed Comuting, Hagit Attiya and Jennifer Welch, McGraw-Hill, Distributed Systems, S. Mullender (Ed.), 2nd ed., Addison- Wesley, Concurrency Control and Recovery in Database Systems, Phili Bernstein, Vassos Hadzilacos, Nathan Goodman, Addison-Wesley, Distributed Systems Distributed Systems: rovide the means for erformance, scalability, deendability loosely couled comuters, modular design introduce secial roblems regarding correctness, comlexity, failures Lamort: A distributed system is one in which the failure of a comuter you didn t even know existed can render your own comuter unusable. Fault-Tolerance: the ability of a system to rovide useful service (ossibly degraded in functionality and/or erformance), desite the fact that some of its comonents malfunction. 3 4

2 rocessors Examle: Consenting adults Bob Bob and Alice wish to meet for lunch at La Tryste. They communicate by , which is known to lose messages. Alice communication medium rocesses What kind of comutational roblems can one solve in a system? Deends on the system model: execution and interaction timeliness, failure behaviour of software and hardware comonents,... 10:00am 11:15am Does she know I got the rely? 12:15m System model: Lets meet atnoon OK Ack OK Ack Ack OK 10:45am 11:45am Does he know I got the ack? Unreliable message transmission No known bound on transmission time No commonly accessible medium (eg. Whiteboard) 5 6 Examle: Consenting adults Bob System model: Let s meet at noon. Sure, see you then. one arty hears what the other says within a bounded delay, or the existence of roblem is known within a bounded delay Alice Ref: The Many Faces of Consensus in Distributed Systems, J. Turek and D. Shasha, IEEE Comuter, June Three dimensions to consider. synchrony considerations comunication medium failure tyes 7 8

3 Communication medium Point-to-oint networks oint-to-oint network : modelled as a grah rocesses : grah nodes communication medium message-assing network oint-to-oint secific toologies eg. broadcast (Ethernet), ring, star,... communication links : grah edges (uni- or bi-directional) Processes connected by a link communicate via send/recv rimitives. send(m) recv(m) transmission out buffer in buffer shared memory send:non-blocking; necessary for fault-tolerance! Often assume comlete communication grah. Links are virtual, not necessarily direct hysical connections Point-to-oint networks Tyes of failures Proerties of failure-free oint-to-oint networks Process secifications: If a rocess has not reached a final state, eventually it will execute another ste. Liveness Process failures: crash...a rocess stos taking stes before reaching a final state. faulty rocess: violatesrocess secifications correct rocess: satisfies rocess secifications Communication secifications: Process receives message m from at most once and only if has reviously sent m to. Note: If sends m to and takes infinitely many stes, then eventually receives m from. In general, do not assume FIFO links; easy to imlement, if needed. Exercise: How? Safety Liveness 11 Link (communication) failures: message loss a message sent from to is never received by, even though takes infinitely many stes. faulty link: violates liveness of communication secifications correct link: satisfies liveness of communication secifications Benign failures - other tyes of failures introduced later on. 12

4 A. Synchronous network model: Known uer bound on time reuired for a rocess to execute a local ste. Known uer bound on message transmission delay. Can assume that rocesses have erfectly synchronised hysical clocks. In ractice, when the two revious roerties hold, aroximately synchronised (with a known bounded drift > 0, from each other or from real time) clocks can be imlemented -- they are more realistic; erfectly synchronised clocks are simler for models. Conseuences: can usetimeouts to detect rocess or link failures send(,m) wait for timeout max{t 1 + t 2 + t 3 } m m recv(m) send(,m ) t 1 t 2 t 3 can organise comutation in rounds send messages to a set of rocesses P recv message of that round from all rocesses in P change state m 1 m 3 m 2 recv(m ) t t r round 1 round 2 round B. Asynchronous network model: No bound on time to execute a local rocess ste; however, time to execute a local ste is finite. No bound on message transmission delay. Cannot assume the existence of erfectly or aroximately synchronised hysical clocks (that measure real time). Note: may have logical clocks. the the most general model -- --an algorithm designed for for asynchronous systems also works in in synchronous systems. 15 Unfortunately, some very basic comutational roblems cannot be solved in asynchronous systems in a fault-tolerant manner (in the resence of failures) and with a deterministic algorithm Thus, for certain roblems we have to resort to synchronous systems or randomised (robabilistic) algorithms - not discussed in this course Synchronous Partially Synchronous Asynchrounous 16

5 Summary To describe a distributed system, must secify: communication grah (often: comlete) rocess failures (e.g. crash failures) link failures (e.g. message loss) assumtions on the number (usually max) of rocess or link failures degree of synchrony for rocesses and communication It is crucial to be clear and recise about these matters as they affect whether: an algorithm works in a given system a comutational roblem is solvable in a given system 17