Unit 7: RPC and Indirect Communication

Transcription

1 SR (Systèmes Répartis) Unit 7: RPC and Indirect Communication François Taïani

2 Outline n Remote Procedure Call è First Class RPC è Second Class RPC (RMI) n Indirect Communication è Group Communication è Publish Subscribe è Message Queues è Tuple Spaces è Shared Memory Reference: Chapter 6 of Distributed Systems: Concepts and Design (5ed edition, 2012), George Coulouris, Jean Dollimore, Tim Kindberg, Gordon Blair, Publisher: Addison-Wesley F. Taiani 2

3 What is RPC? n Higher level mechanism supporting the construction of distributed applications (see also discussion on transparency later) n Supports the calling of a procedure in a separate address space (process) as if it exists in the local address space; the process may or may not be on the same machine n But what is the semantics of a local procedure call? F. Taiani 3

4 RPC and Middleware Application RMI, RPC and indirect comm. Request reply protocol External Data Representation Middleware layers Operating System F. Taiani 4

5 Styles of RPC n First class RPC è Integrated into the language => normal language mechanisms can be used for exceptions etc è Examples include Java RMI, Ada and Argus n Second Class RPC è A special Interface Definition Language (IDL) is used to define communications (see later) è Language independence è Examples include Sun RPC, CORBA and DCE F. Taiani 5

6 Asynchronous RPC 2-12 a) The interconnection between client and server in a traditional RPC b) The interaction using asynchronous RPC F. Taiani 6

7 Programming with Interfaces n Separation of interface and implementation n Client software does not need to know the details of the implementation, cf. abstraction n Important for platform and language independence n Also important to support the evolution of software someinterface myobject:myclass F. Taiani 7

8 Example: CORBA IDL interface inventory! {! // attributes and type definitions!! const long MAX_STRING = 30;! typedef long part_num;! typedef long part_price;! typedef long part_quantity;! typedef string part_name<max_string+1>;!! struct part_stock {!! part_quantity max_threshold;! part_quantity min_threshold;! part_quantity actual;!! }; //end of part_stock! p.t.o. F. Taiani 8

9 CORBA IDL (continued) // operations!! boolean is_part_available(in part_num number);!! void get_price (in part_num nmb, out part_price price );!! void get_stock (in part_num number,! out part_quantity quantity);!! long order_part(in part_num number, inout part_quantity quantity,! in account_num account );!! }; //end of interface inventory!! F. Taiani 9

10 The Issue of Transparency n Definition è Transparency = hiding aspects of distribution n Transparency in RPC = ultimate goal, but è Overall cost of an RPC è Extra exceptions that are raised è Parameter passing is different, e.g. pointers F. Taiani 10

11 Implementing RPC Communications modules Object A Proxy Request Dispatcher Remote Object B Client Stub Reply Server Stub N.B. Proxies, stubs, dispatchers are generated automatically by an appropriate IDL compiler F. Taiani 11

12 Key Components: Client Side n Proxies (see also smart proxies) è Masquerades as a local version of the remote interface è Redirects calls to client stubs è May perform other actions (smart proxies) n Client stubs è Carries out marshalling (flattening) of calls and the requests the retransmission of the message è Also must unmarshal returning replies F. Taiani 12

13 Key Components: Server Side n Dispatchers è Receive incoming messages and direct them to an appropriate server stub n Server stubs (skeletons) è Unmarshals message and then invokes appropriate code body è Must also marshal reply values and initiate transmission back to the client F. Taiani 13

14 Essential Characteristics of RMI n Full integration with OO programming language è Ability to exploit objects, class and inheritance è Integration with built in mechanisms e.g. exceptions è Presented as object references (transparency) n More sophisticated options for parameter passing è Pass by (object) reference è Pass by value (exploiting serialisation) n Often integrated with code (object) mobility è E.g. java.rmi.server.codebase property in RMI Java RMI F. Taiani 14

15 Advanced Features n Static vs. dynamic invocation è Static inv. (default) = interfaces compile time è Any changes of interface re-compilation è Dynamic inv. interface / class run-time (related to reflection, e.g. java.lang.reflect) n Object adapters è wrapper around an object or set of objects è goal 1 = common abstraction / hide heterogeneity è goal 2 = additional services (lifecyle, threading) n Mostly available in CORBA (less in RMI) è e.g. Dynamic Inv. Interface, Dynamic Skeleton Interface F. Taiani 15

16 Additional References n Remote Object Activation: Doc è rmi-activation.html n Remote Object Activation: Tutorials è activation.html è rmi/activation/overview.html F. Taiani 16

17 Outline n Remote Procedure Call è First Class RPC è Second Class RPC (RMI) n Indirect Communication è Group Communication è Publish Subscribe è Message Queues è Tuple Spaces è Shared Memory Reference: Chapter 6 of Distributed Systems: Concepts and Design (5ed edition, 2012), George Coulouris, Jean Dollimore, Tim Kindberg, Gordon Blair, Publisher: Addison-Wesley F. Taiani 17

18 Indirect Communication? n Communication in DS è through intermediary è no direct coupling between sender(s) & receiver(s) n Intermediary = è Groups è Pub/Sub (distributed events) è Message queues è Shared memory abstractions (e.g. DSM, tuple spaces) n Often = multiparty communication F. Taiani 18

19 A Closer Look at Coupling n RPC / RMI = direct link between client and server n Indirect communication = uncoupling è Space uncoupling: participants not known to each other è Time uncoupling: independent lifecycles n Space and time uncoupling good for è mobility (participants come and go) è dependability (transparent replication, durability) è event dissemination (dynamic senders / receivers) F. Taiani 19

20 A Few Words on Indirection All problems in computer science can be solved by another level of indirection Roger Needham et al There is no performance problem that cannot be solved by eliminating a level of indirection Jim Gray F. Taiani 20

21 Example 1: Group Comm. n What is group communication? è Concept of a group abstraction (join, leave) è Messages sent to the whole group (one invocation) è Additional guarantees on ordering and reliability n Uses in distributed systems è Important in supporting fault-tolerance è Also used heavily in dissemination of events è Key examples: JGroups and Isis See lecture on group communication F. Taiani 21

22 Example 2: Publish-Subscribe n What is publish-subscribe? è Publishers publish an event e: publish(e) è Subscribers express interest through filter f: subscribe(f) è Events are delivered asynchronously: notify(e) è Publishers optionally advertise what published: advertise(f) è System = broker to deliver events to the right subscribers n Uses in distributed systems è financial information systems / news feeds applications è Cooperative working è Ubiquitous computing/ monitoring (smart homes, ) è Examples: JMS F. Taiani 22

23 Publish-Subscribe (continued) F. Taiani 23

24 Subscription Model n Topic-based (also referred to as subject-based) è Each notification belong to a topic è subscriptions defined in terms of topic of interest n Content-based è subscriptions = logical predicate on event è more expressive but difficult to implement efficiently! F. Taiani 24

25 Implementing Publish-Subscribe n Choices: centralised / distributed / fully P2P n Most systems è network of brokers F. Taiani 25

26 Realisations n Flooding è All events to everyone (reverse: all subscriptions) è Simple but significant message overload n Filtering è Filters propagated back through the broker network è notifications only forwarded if path to a valid subscriber n Advertisements è With filtering = filters need to be go back to all publishers è Overhead can be reduced by use of advertisements F. Taiani 26

27 Realisations (Cont) n Rendezvous è Set of all possible events = an event space è Partition this event space è Each broker = responsible for part of the space n Two distributed functions available on each node: è SN(s) for sub. s = rendezvous node(s) in charge of s è EN(e) for event, e = rendezvous node(s) in charge of e è Can lead to highly scalable implementations F. Taiani 27

28 Message Queues n What is a message queue? è messages are sent to a queue = distributed entity è reception: either blocking or non blocking interface è Additional properties: persistence / reliability (ack/nacks) n Uses in distributed systems è Enterprise Application Integration (EAI) è Commercial transaction processing systems è Examples: JMS, IBM s Websphere MQ, Microsoft s MSMQ and Oracle s Streams Advanced Queuing AQ F. Taiani 28

29 Message Queues (continued) F. Taiani 29

30 Distributed Shared Memory n Key ideas è Illusion of shared memory in a distributed system è If data is not available locally = must be fetched è Hides distribution entirely from the programmer è Challenges: consistency / cost n Uses in distributed systems è High performance computing, Cluster Computing è Some of these systems developed in Rennes e.g. F. Taiani 30

31 Distributed Shared Memory (continued) F. Taiani 31

32 Comparison of Approaches source: Distributed Systems: Concepts F. Taiani and Design (5ed edition, 2012) 32

33 Summary n Not one single interaction paradigm in dist. sys. è RPC / RMI one of them è Induce strong coupling between clients / server n Many approaches around indirect communication è usually = distribution no longer entirely transparent è but better space / time uncoupling è enriched with additional properties (order, reliability, ) F. Taiani 33