Christian Queinnec. Locality, Causality and Continuations. In Proceedings of the 1994 ACM Conference on Lisp and Functional Programming, Orlando, Florida, June 1994.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....by an implementation with transparent annotations for parallelism. Thanks to the proof of correctness, our semantics satisfies these criteria. There exist other approaches to parallelism which do not preserve the sequential meaning of programs. Among others, let us cite Queinnec s ICSLAS [19, 20], and Ito s PAILISP [10, 11] which both define new semantics of continuations in a parallel framework. Let us observe that their goals differ from ours because their primitives for parallelism add expressiveness to the sequential core, like for instance Queinnec s pcall or Ito s parallel or . ....

Christian Queinnec. Locality, Causality and Continuations. In Proceedings of the 1994 ACM Conference on Lisp and Functional Programming, Orlando, Florida, June 1994.

....of an annotated program with effects requires synchronisations between computations. The goal of this paper is to design a semantics that introduces as few synchronisations as possible so that parallelism can exist. It has been observed that mostly functional programs use effects in a local way [24]: e.g. continuations are captured and invoked inside a function, or side effects are performed locally in a function. The semantics that we propose offers parallelism when effects remain local: an expression e 2 is allowed to run speculatively in parallel with an expression e 1 as long as e 2 ....

....approach should first be extended to a Scheme like language. Halstead [11, page 19] proposed three criteria for the semantics of parallel constructs and continuations in parallel Scheme. Not only does our semantics satisfy them, but also can these criteria be extended to side effects. Queinnec [24] proposed a coherency protocol for shared mutable variables in a distributed dialect of Scheme, called ICSLAS [25] The protocol ensures that a sideeffect perceived by a thread cannot be ignored by the threads of its continuations. However, Queinnec s goal differs from ours because parallelism is ....

Christian Queinnec. Locality, Causality and Continuations. In Proceedings of the 1994 ACM Conference on Lisp and Functional Programming, Orlando, Florida, June 1994.

....flowed from a task with legitimacy 2 to a task with legitimacy 1 . As we want future to be an annotation, every program should return the result that it would produce when evaluated sequentially in the absence of future. The solution adopted in our semantics is to perform causally dependent [22] box accesses in the same order as in a sequential implementation; the solution relies on legitimacies. The translation of the primitive deref, x: deref (sync (touch x) touches and then applies sync on its argument. The primitive sync behaves as the identity function if the legitimacy of the ....

....of our rules could make this strategy explicit. Though rule (migrate task) does not enforce any migration strategy, we think that task stealing [1, 2] would be appropriate; according to this strategy, a processor that becomes idle steals a task from a heavily loaded processor. Queinnec s ICSLA [25, 22, 24, 23] is a dialect of Lisp offering primitives for parallelism (essentially like a fork) transparent migration of objects, and maintenance of their cache coherence over the network. Queinnec s purpose is different from ours: as he does not rely on transparent annotations, he does not preserve the ....

Christian Queinnec. Locality, Causality and Continuations. In Proceedings of the 1994 ACM Conference on Lisp and Functional Programming, Orlando, Florida, June 1994.

....flowed from a task with legitimacy 2 to a task with legitimacy 1 . As we want future to be an annotation, every program should return the result that it would produce when evaluated sequentially in the absence of future. The solution adopted in our semantics is to perform causally dependent [24] box accesses in the same order as in a sequential implementation; the solution relies on legitimacies. The translation of the primitive deref, x: deref (sync (touch x) touches and then applies sync on its argument. The primitive sync behaves as the identity function if the legitimacy of the ....

....of our rules could make this strategy explicit. Though rule (migrate task) does not enforce any migration strategy, we think that task stealing [1, 3] would be appropriate; according to this strategy, a processor that becomes idle steals a task from a heavily loaded processor. Queinnec s ICSLA [27, 24, 26, 25] is a dialect of Lisp offering primitives for parallelism, transparent migration of objects, and maintenance of their cache coherence over the network. Queinnec s purpose is different from ours: as he does not rely on transparent annotations, he does not preserve the sequential meaning of ....

Christian Queinnec. Locality, Causality and Continuations. In Proceedings of the 1994 ACM Conference on Lisp and Functional Programming, Orlando, Florida, June 1994.

....threads [22] or higher level constructs like pcall [34, 45] or future [18, 31] In the sequel, we use the term task to denote an evaluation thread created by the constructs for parallelism. Similarly, message passing style primitives [12, 14, 13] or higher level forms of remote function calls [30, 32, 42] may be adopted for distribution. 1 Other names found in the literature for a similar concept are fuel [20] computron [46, p. 102 103] teleclick [27] or metapill [1] primitives : call with group(F; e; e ; t ) j pause(g; p ) j awake(g; e) g 2 Group e 2 Energy F : Group Theta ....

....Our permanent concern when designing Quantum was to be able to compute in a distributed framework. Hence, we decided that Quantum would be independent of the memory model: so, real shared memory, shared memory simulated over a distributed memory [30] distributed causally coherent memory [42] are memory models that may be adopted with Quantum. 3 Implementation Description This section outlines the current implementation. First, it presents a solution in a non distributed setting, and then describes specific problems encountered when modifying the thread library PPCR [52] Second, it ....

[Article contains additional citation context not shown here]

Christian Queinnec. Locality, Causality and Continuations. In Proceedings of the 1994 ACM Conference on Lisp and Functional Programming, Orlando, Florida, June 1994.

....the remotely read object. Therefore a cache is invalid if its monitoring clock is known to be greater. This implements a lazy invalidation protocol and ensures causality i.e. when a site receives information from another site, it cannot ignore the mutations this other site was aware of. See [Que94b, Que94a] for more details. Causality seems to be the weakest coherency protocol that still keeps a decent semantics for a distributed memory viewed from a language, that is, offered by ae oe ae oe Gamma Gamma Delta Delta T T Delta Delta T T 6 an Object . ....

....3.6 Clocks DMeroon s API offers to specify explicitly the clock that monitors an object. On all sites there exists a general clock which, by default, monitors all exported objects not already monitored by a clock. Because a clock monitors numerous objects, and since the lazy invalidation protocol [Que94b, Que94a] when incrementing a clock, ipso facto invalidates all its cached objects, one may create clocks and explicitly specify the objects they monitor. To have numerous clocks allows to invalidate less objects when ticking but increases the transmission overhead since more clocks are (lazily) spread ....

Christian Queinnec. Locality, causality and continuations. In LFP '94 -- ACM Symposium on Lisp and Functional Programming, pages 91--102, Orlando (Florida, USA), June 1994. ACM Press.

....of functions allowing to share objects over Internet in a coherent way among multiple readers and, at most, one writer. Management is distributed and coherency is causal. This Section will focus on the implementation of objects and remote pointers. Coherency is out of the scope of this paper [Que95, Que94] and so are the details of DMeroon [Que98] The DMeroon project started as the memory layer for a distributed language [QD93] but soon tried to multi lingual i.e. be a common layer for di erent languages (currently C and Scheme) to exchange, copy or share typed structured values. 1.1 Objects ....

Christian Queinnec. Locality, causality and continuations. In LFP '94 { ACM Symposium on Lisp and Functional Programming, pages 91-102, Orlando (Florida, USA), June 1994. ACM Press.

....Our permanent concern when designing Quantum was to be able to compute in a distributed framework. Hence, we decided that Quantum would be independent of the memory model: so, real shared memory, shared memory simulated over a distributed memory [19] distributed causally coherent memory [25] are memory models that may be adopted with Quantum. However, in the programs that we wrote in Quantum, we needed primitives to synchronise computations and to exchange information between them. We observed that asynchronous unbounded communication channels [15] offered the appropriate level of ....

Christian Queinnec. Locality, Causality and Continuations. In Proceedings of the 1994 ACM Conference on Lisp and Functional Programming, Orlando, Florida, June 1994.

....a group implicitly suspends all its subgroups which can be resumed independently. We propose a coherency protocol to achieve the distributed management of these groups. Section 1 informally presents the essential features of the Icslas language, already described in more details in [Que92, QD93, Que94] The basic operational machinery we use to define the language appears in Section 2. We also exercise it to define the functional kernel and the concurrent features of the language. Distribution aspects as well as the protocol to manage shared mutable objects are dealt with in Section 3. ....

....An essential concurrent and distributed language This section informally describes the main features of a concurrent and distributed toy functional language. This language (see figure 1) is based on Scheme [CR91] and is an essential version of the language described with examples in [Que92, QD93, Que94] quote ) lambda ( if 1 2 3) begin 1 2) new box ) box ref ) box set 1 2) fork 1 2) suicide) remote 1 2) sponsor ) pause ) awake ) Fig. 1. Syntax of the essential language The value of a variable is obtained via its name ( Constants ....

[Article contains additional citation context not shown here]

Christian Queinnec. Locality, causality and continuations. In LFP '94 -- ACM Symposium on Lisp and Functional Programming, pages 91--102, Orlando (Florida, USA), June 1994. ACM Press.

....of functions allowing to share objects over Internet in a coherent way among multiple readers and, at most, one writer. Management is distributed and coherency is causal. This Section will focus on the implementation of objects and remote pointers. Coherency is out of the scope of this paper [Que95, Que94] and so are the details of DMeroon [Que98] The DMeroon project started as the memory layer for a distributed language [QD93] but soon tried to multi lingual i.e. be a common layer for different languages (currently C and Scheme) to exchange, copy or share typed structured values. 1.1 Objects ....

Christian Queinnec. Locality, causality and continuations. In LFP '94 -- ACM Symposium on Lisp and Functional Programming, pages 91--102, Orlando (Florida, USA), June 1994. ACM Press.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC