Emerald - et OO-språk for distribuerte applikasjoner Review – Concurrency - Mobility

1. Emerald - et OO-språk for distribuerteapplikasjonerReview – Concurrency - Mobility Eric Jul OMS Professor II

2. Emerald Review We’ll start with a review of Emerald from the first lecture series (F1): • Everything is an object • Algol-60/Simula/Smalltalk heritage • Object constructors • Conformity based type system – think interfaces (e.g., as in Java)

3. People

4. Main Contributions • Distribution: Mobile objects (Eric/Hank) Anyobjectcanmove at any time. Fullon-the-fly • objectmobility • threadmobility • heterogeneousmobility • Conformitybased type system (Norm/Andrew) Type system basedonconformityprinciple Well-definedsemantics (e.g., NIL makessense!) • Clean OO language (betterthansuccesors?) including uniform object model

5. What does it look like? • In a nutshell: Java with an Algol-like syntax • Heavily inspired by Algol/Simula and Smalltalk • ”Clean” OO language – ”everything” is an object: data, integers, strings, arrays, classes, types as in Smalltalk • Language constructs are NOT objects – for compilability and speed • No pointers: just object & object references

6. ClasslessObject Construction Objectconstructors: objectseqno var prev: Integer = 0 Integer operation getSeqNo[] prev <- prev +1 returnprev end getSeqno end seqno The above is an executableexpression!

7. Object Constructors • Execution results in a new object • Execute again – and get yet another object • No class! Want classes?

8. An Objectthat is a Class objectseqnoclass operation create[] return objectseqno var prev: Integer = 0 Integer operation getSeqNo[] prev <- prev +1 returnprev end getSeqno end seqno end create end seqnoclass

9. Class done by SyntaticSugaring The followingturnsinto the previous double objectconstructor: classseqno var prev: Integer = 0 Integer operation getSeqNo[] prev <- prev +1 returnprev end getSeqno end seqno

10. Types Types are abstract descriptions of the operations required of an object (think: Java Interfaces – they are close to identical to types in Emerald). Collection of operation signatures. Signature is name & types of each parameter

11. type BankAccount operation deposit[Integer] operation withdraw[Integer] ->[Integer] functionfetchBalance[] -> [Integer] end BankAccount type DepositOnlyBankAccount functionfetchBalance[] -> [Integer] operation deposit[Integer] end DepositOnlyBankAccount Conformityobject-to-type and type-to-type BankAccount conforms to DepositOnlyBankAccount because it support all the required operations – and the parameters also conform What is conformity?

12. Conformity informally An object is said to conform to a type, if • It has the operations specified by the type • For each operation in the type: • The number of parameters is the same in the object as in the type • Each input parameter of the object conforms to the corresponding param of the type • Each output parameter of the type conforms to the corresponding param of the object (contra variant)

13. Conformity between types Conformity is a mathematicalrelationship If T is to conform to S: • T must have all the operations required by S • For each operation in T the corresponding operation in S: • in-parameters must conform • out-parameters must conformin oppositeorder Contravariance: not in Simula nor Eiffel necessary to makesemanticsense of programs

14. Conformity details • Conformity is implicit • No ”implements” as in Java • Operation namesimportant • Parameter names do not matter, just their type • Aritymatters: foo(char) different from foo(char, float)

15. Conformity more formally NOT IMPORTANT HERE

16. Lattice of types • Types form a lattice • Top is type Any end Any • Bottom is Noone(it has ALL operations”) • NILconforms to Noone • NILcanthusbeassigned to any variable! (Read ”MuchAdoAboutNIL.)

17. Array • Just an object • Supports “of” which returns an object • Array.of[Integer] • This is a type (!) • But it is also an object – that supports create (!) • Creates an EMPTY array. • Addupper, addlower

18. TODAY (F2) • Process concept (Emerald process = thread) • Synchronization using Monitors

19. Process Concept A process is a thread of execution. Every object can have ONE process. Process section in object constructor

20. Process section object A initially … initializestuff end initially process … do something end process end A

21. Process execution After the end of the initially section, the process (if present) is started and executes in parallel with all other processes.

22. Concurrent Updates object A process m <-ba.withdraw[100] end process end A object B process m <- ba.withdraw[100] end process end B

23. Synchronization Required Classic Monitors as described by Tony Hoare Example: hi – ho program (synch.m)

24. Alternatives • Semaphore – go thru example (sem.m) • Rendez-vous • Barrier

25. Distribution • Sea of objects (draw) • Sea is dividedintodisjunct parts called Nodes • An object is onone and onlyone Node at a time • Each node is represented by a Nodeobject

26. Location Primitive • Locate X returns the node where X is (was!) • Note that the objectmayalready have moved to another node (actuallyanynumber of moves)

27. Mobility Primitive Basic primitive is move X to Y The object X is moved where Y is. More formally: The object denoted by the expression X is move to the node where the object denoted by expression Y was! If the move cannot be done, it is ignored. NOTHING is guaranteed – nothing may happen.

28. Strong Move: Fix Basic primitive is fix X at Y The object X is moved where Y is & stays there. More formally: The object denoted by the expression X is move to the node where the object denoted by expression Y was! Either the move happens – or it fails. Strong guarantees; potentially expensive

29. Immutable Objects Immutable objects are omnipresent so moving them does nothing and ALWAYS succeeds.

30. objectBoss process var w: Worker var n: Node n <- …find usablenode moveself to n w <- Worker.create[ ] end process end Boss classWorker process do work… end process end Worker Mobility Example

31. object Boss var w: Worker var n: Node n <- …find usable node w <-Worker.create[ ] move w to n w.StartWork[ ] end Boss classWorker op StartWork slave <- object slave process work… work end process end StartWork end Worker Mobility Example

32. Why two different moves? • Fast efficient – mobility hint • Slow but sure for when location is part of the semantics of the application.

33. Why Mobility • Localcallsaretypically 1,000 – 10,000 times faster thanremotecalls • Mobility for: • performance • availability

34. Mobility and Location Concepts locate X returns(one of) the object X’s locations move X to Y movethe object X to the node whereY is (orratherwas) fix X at Y as move but disregard subsequentmoves refix X at Y as fix but for fixedobjects unfix X allownormal moves

35. var B: some data object … X.F[move B] … object X operation F[arg:T] loop arg.g[…] exit after manyloops end loop end X Call-by-move

36. var B: some data object … X.F[move B] … … X.F[visit B] … object X operation F[arg:T] loop arg.g[…] exit after manyloops end loop end X Call-by-visit

37. How Many Calls of B? Questions: given a normal PC enviroment, say 2 GHz CPU, 100 Mbit/s Ethernet, howmanycalls of a small (say 100 bytes) argument B beforebreakeven? • 1 • 10 • 100 • 1,000 • 10,000 • 100,000 • 1,000,000

38. Return-by-move When an operation creates a result object and knows it is for the caller’s use only, it can choose to return the parameter by move. Return-by-move is not necessary – but increases efficiency – why??

39. objectKillroy process var myNode <- locateself var up: array.of[Nodes] up <- myNode.getNodes[] foreach n in up moveself to n end foreach end process end Killroy Objectmovesitself to all available nodes On the original MicroVAX( 1987) implementation: 20 moves/second! Note: the thread (called a process in Emerald) movesalong Killroy

40. Attachment Problem: move an object but its internal data structure does not move along! Classic example: A tree

41. Tree class TreeClass var left, right: TreeClass var data: … end TreeClass

42. Attached Tree class TreeClass attached var left, right: TreeClass var data: … end TreeClass

43. Attachment: can it be decided automatically? Tree example TreeNode left, right Mail message To From Subject Body

44. Attachment costs Attachment has NO run-time cost! Just a bit in the DESCRIPTOR for an object. One bit for each variable. Better: compiler sorts by attached bit – then merely two integers, e.g., 5 attached variables 4 non-attached variables

45. Dynamic Attachment var X: … <- something attached varaX: … … Join: aX <- X Leave: aX <- NIL

46. Conclusion Emerald has • concurrency with Hoare monitors • fully integrated distribution facilities • has full on-the-fly mobility • a novel attachment language feature Many novel implementation techniques (more talks to come!)