Time and Coordination in Distributed Systems. Operating Systems

Size: px

Start display at page:

Download "Time and Coordination in Distributed Systems. Operating Systems"

Alisha Clarke
6 years ago
Views:

1 Time and Coordination in Distributed Systems Oerating Systems Clock Synchronization Physical clocks drift, therefore need for clock synchronization algorithms Many algorithms deend uon clock synchronization Often we need to know the order in which two events occurred on two different comuters Clock synch. Algorithms Christian, NTP, Berkeley algorithm, etc. However, since we cannot erfectly synchronize clocks across comuters, we cannot use hysical time to order events

2 Skew between comuter clocks in a distributed system Network Clock synchronization using a time server m r m t Time server,s

3 Clock synchronization algorithms Cristian s algorithm should set its time to t + T round / Earliest time at which S could have laced its time in m t was min after disatched m r Latest oint at which it could do so was min before m t arrived at Time by S s clock when message arrives at is in range [t + min, t + T round min] Accuracy ±(T round /-min) 5 An examle synchronization subnet in an NTP imlementation Note: Arrows denote synchronization control, numbers denote strata. 6

4 Logical time & clocks Lamort roosed using logical clocks based uon the haened before relation If two events occur at the same rocess, then they occurred in the order observed Whenever a message is sent between rocesses, the event of sending occurred before the event of receiving X haened before Y denoted by X Y 7 Events occurring at three rocesses a b m c d m Physical time e f 8

5 Lamort s algorithm Each rocess has its own logical clock LC: C is incremented before each event at rocess LC:. When rocess sends a message it iggybacks on it the value C. On receiving a message (m,t) a rocess q comutes C q = max(c q, t) and then alies LC before timestaming the receive event 9 Lamort timestams for the events a b m c d m Physical time e f 5 0 5

6 Vector timestams for the events (,0,0) (,0,0) a b m (,,0) (,,0) c d m Physical time (0,0,) e f (,,) Totally ordered logical clocks Logical clocks only imose artial ordering For total order, use (T a,p a ) where P a is rocessor id (T a,p a ) < (T b,p b ) if and only if either T a < T b or (T a = T b and P a < P b ) 6

7 Distributed mutual exclusion Central server algorithm Ricart and Agrawal algorithm A distributed algorithm that uses logical clocks Ring-based algorithms NOTE: the above algorithms are not fault-tolerant and not very ractical. However, they illustrate issues in the design of distributed algorithms Several other mutual exclusion algorithms have been roosed Quorum consensus algorithms Maekawa s algorithm Server managing a mutual exclusion token for a set of rocesses Queue of requests Server. Grant token. Request token. Release token 7

8 A ring of rocesses transferring a mutual exclusion token n Token 5 Ricart and Agrawala s algorithm On initialization state := RELEASED; To enter the section state := WANTED; Multicast request to all rocesses; request rocessing deferred here T := request s timestam; Wait until (number of relies received = (N )); state := HELD; On receit of a request <T i, i > at j (i j) if (state = HELD or (state = WANTED and (T, j ) < (T i, i ))) then queue request from i without relying; else rely immediately to i ; end if To exit the critical section state := RELEASED; rely to any queued requests; 6 8

9 Multicast synchronization Rely Rely Rely 7 Maekawa s algorithm Every node needs ermission from the other nodes in its quorum before it enters the critical section Quorums are constructed in such a way that no two nodes can be in their critical section at the same time The size of each nodes quorum is O(sqrt(N)), which can roved to be otimal 8 9

10 Construction of coteries Consider a system with 9 nodes The quorum for any node includes the other nodes in the same row and column Node s quorum = {,,,,7} Node s quorum = {,,,5,8} Node 6 s quorum = {,5,6,,9} The quorum of any two nodes have a non-null intersection. This ensures that two nodes cannot get ermission to enter their critical section at the same time 9 Maekawa s algorithm On initialization state := RELEASED; voted := FALSE; For i to enter the critical section state := WANTED; Multicast request to all rocesses in V i { i }; Wait until (number of relies received = (K )); state := HELD; On receit of a request from i at j (i j) if (state = HELD or voted = TRUE) then queue request from i without relying; else send rely to i ; voted := TRUE; end if 0 0

11 Maekawa s algorithm cont d For i to exit the critical section state := RELEASED; Multicast release to all rocesses in V i { i }; On receit of a release from i at j (i j) if (queue of requests is non-emty) then remove head of queue from k, say; send rely to k ; voted := TRUE; else voted := FALSE; end if Election Algorithms An election is a rocedure carried out to chose a rocess from a grou, for examle to take over the role of a rocess that has failed Main requirement: elected rocess should be unique even if several rocesses start an election simultaneously Algorithms: Bully algorithm: assumes all rocesses know the identities and addresses of all the other rocesses Ring-based election: rocesses need to know only addresses of their immediate neighbors

12 A ring-based election in rogress Note: The election was started by rocess 7. The highest rocess identifier encountered so far is. Particiant rocesses are shown darkened The bully algorithm election The election of coordinator, after the failure of and then Stage election answer answer election C Stage election answer election C Stage timeout Eventually... Stage coordinator C

Distributed Systems (5DV147)

Distributed Systems (5DV147) Mutual Exclusion and Elections Fall 2013 1 Processes often need to coordinate their actions Which rocess gets to access a shared resource? Has the master crashed? Elect a new