Attiya H. and Welch J.L., Sequential Consistency versus Linearizability. ACM Transactions on Computer Systems, 12(2):91-122, 1994.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....memory systems providing different semantics. It seems clear that strong memory models make it simpler to write programs, since the return value of each read operation is more predictable. However, it is widely accepted that strong memory models do not scale well with the number of processes [1, 7]. On the Most of this work was performed while the author was visiting the Laboratory for Computer Science, MIT. other hand weaker memory models can be more efficiently implemented, since they allow more possible return values for each read operation. This presents a tradeoff between simplicity ....

H. Attiya and J.L. Welch. Sequential consistency versus linearizability. ACM Transactions on Computer Systems, 12(2):91--122, 1994.

....els. The strong memory models are close in behavior to a centralized memory, which makes simple to write programs with them (the return value of each read operation is rather predictable) However, it is widely accepted that strong memory models do not scale well with the number of processes [3, 6]. On the other hand, weaker memory models can be more eciently implemented, since they require less consistency overhead. This implies more possible return values for each read operation [2, 5, 8] which makes harder to write programs for these models. Hence, we are faced with a tradeo between ....

H. Attiya and J. Welch. Sequential consistency versus linearizability. ACM Transactions on Computer Systems, 12(2):91-122, 1994.

....consistency. Sequential consistency [13] guaranteed by systems such as Orca [3] allows operations to be re ordered as long as they remain consistent with the view of individual clients. An inherent disparity in the performance of atomic and sequentially consistent objects has been established [2]. Other systems provide even weaker guarantees to the clients [6, 5, 7] in order to get better performance. Improving performance by providing weaker guarantees results in more complicated semantics. Even when the behavior of the replicated objects is specified unambiguously, it is more difficult ....

H. Attiya and J. Welch. Sequential consistency versus linearizability. ACM Transactions on Computer Sys- tems, 12(2), 1994.

....objects. Speci cally, under these execution constraints, it is necessary and sucient to ensure legality of reads to guarantee m sequential consistency (and m linearizability) Finally, we provide algorithms for ensuring proposed consistency conditions in a distributed system. Several papers [2, 4, 18, 20] have proposed sequentially consistent implementations for read write objects. Attiya and Welch [4] provide sequentially consistent and linearizable implementations for read write objects, FIFO queues and stacks. In addition, they also give an analysis of the response time of their ....

....of reads to guarantee m sequential consistency (and m linearizability) Finally, we provide algorithms for ensuring proposed consistency conditions in a distributed system. Several papers [2, 4, 18, 20] have proposed sequentially consistent implementations for read write objects. Attiya and Welch [4] provide sequentially consistent and linearizable implementations for read write objects, FIFO queues and stacks. In addition, they also give an analysis of the response time of their implementations. But their implementation for linearizability assumes that clocks are perfectly synchronized and ....

[Article contains additional citation context not shown here]

Hagit Attiya and Jennifer L. Welch. \Sequential Consistency versus Linearizability". ACM Transactions on Computer Systems, 12(2):91-122, May 1994. 28

....the first of two inc operations is completed before the second is initiated, the first gets a lower counter value than the second. For many applications, the use of a linearizable counting scheme as a basic primitive considerably simplifies arguing about correctness. As explained in [HW90, AW94] linearizability is related to (but not the same as) other correctness criteria for asynchronous distributed systems [Lam79, Pap79] The quest for finding an efficient scheme for distributed counting is strongly related with the quest for finding an appropriate measure of efficiency. The ....

....Pyramid are the only counting schemes known to be applicable and efficient when processors initiate the inc operation very frequently. Linearizability Besides performance, linearizability is a key property of a counting scheme since it considerably simplifies arguing about correctness [HW90, AW94] We have seen that both, the Diffracting Tree and the Bitonic Counting Network are not linearizable (see Lemma 4.10 for the Bitonic Counting Network and [HSW96] for Counting Networks in general, see Lemma 5.5 for the Diffracting Tree) The Central Scheme and the Combining Tree (Counting ....

Hagit Attiya and Jennifer L. Welch. Sequential consistency versus linearizability. ACM Transactions on Computer Systems, 12(2):91--122, May 1994.

....We present an algorithm for hnearizable read write objects in the clock automaton distributed system model. We consider the simple algorithm for this problem in the timed automaton distributed system model given by Mavronicolas in [10] the algorithm is a generalization of an algorithm in [2]. We then transform the algorithm using our simulation techniques to arrive at a simple algorithm in the clock automaton system. We proceed by defining superlinearizabilit (a property stronger than hnearizability) and modifying the algorithm for hnearizability in the timed automaton model, to ....

H. Attiya and J. Welch, Sequential Consistency Versus Linearizability, Proceedings of 3rd A CM Symposium on Parallel Algorithms and Architectures, July 1991.

....close to what programmers expect from a shared memory, while the later is considered to be powerful enough to allow easy programming, but at the same time allows for inexpensive implementations. As a consequence, a number of algorithms have been proposed in the literature implementing sequential [1, 3, 4] and causal consistency [2, 7, 8] A third consistency model proposed in the literature is the cache model [5] for which to our knowledge, no algorithm has been proposed. An interesting property of any algorithm implementing a consistency model is the time a memory operation takes. If a memory ....

....that issued it, it is said to be fast, which is a very desirable feature. All the above mentioned algorithms for causal consistency are fast. This work is partially supported by the CICYT under grant TEL99 0582 and the Comunidad Aut onoma de Madrid under grant CAM 07T 00112 1998. However, in [3], Attiya and Welch have shown that no sequential consistency algorithm can guarantee the fast executions of all its operations. This impossibility result restricts the eciency of any implementation of sequential consistency. In general, in order to increase concurrency, most DSM protocols support ....

[Article contains additional citation context not shown here]

H. Attiya and J.L. Welch. Sequential consistency versus linearizability. ACM Transactions on Computer Systems, 12(2):91-122, 1994.

....what programmers expect from a shared memory, while the later because it is considered to be powerful enough to allow easy programming, but at the same time allows for inexpensive implementations. As a consequence, a number of algorithms have been proposed in the literature implementing sequential [2, 4, 6] and causal consistency [3, 11, 12] A third consistency model proposed in the literature is the cache model [7] However, to our knowledge, no algorithm to implement the cache consistency model has been proposed. An interesting property of any algorithm implementing a consistency model is how ....

....does not need to wait for any communication to finish, and can be completed based only on the local state of the process that issued it, it is said that the operation is fast, which is a very desirable feature. All the above mentioned algorithms for causal consistency are fast. However, in [4], Attiya and Welch have shown that no sequential algorithm can guarantee the fast executions of all its operations. This impossibility result restricts the efficiency of any implementation of sequential consistency. In general, in order to increase concurrency, most DSM protocols support ....

[Article contains additional citation context not shown here]

H. Attiya and J.L. Welch. Sequential consistency versus linearizability. ACM Transactions on Computer Systems, 12(2):91--122, 1994.

....of consistency. Sequential consistency [16] guaranteed by systems such as Orca [3] allows operations to be reordered as long as they remain consistent with the view of isolated clients. An inherent disparity in the performance of atomic and sequentially consistent objects has been established [2]. Other systems provide even weaker guarantees to the clients [9, 5, 10] in order to get better performance. Improving performance by providing weaker guarantees results in more complicated semantics. Even when the behavior of the replicated objects is specified unambiguously, it is more ....

H. Attiya and J. Welch. Sequential consistency versus linearizability. ACM Transactions on Computer Systems, 12(2), 1994.

....guarantees with respect to message delivery, but not with respect to request processing. When multiple requests to an object overlap, we say that they execute concurrently. There exist several correctness conditions for concurrent execution, like serializability [14] sequential consistency [2], or linearizability[9] These correctness conditions specify what happens in the presence of interleaved or overlapping accesses to an object (or to different copies of the same object) They ensure that the concurrent execution of multiple requests are equivalent to some valid non concurrent ....

H. Attiya and J. Welch. Sequential consistency versus linearizability. ACM Transactions on Computer Systems, 12(2):91--122, May 1994.

....programmability for performance. Sequential consistency was first defined by Lamport [19] in this paper, we use an alternative formulation proposed by Afek et al. 2] based on I O automata. Other papers exploring correctness conditions for shared memory and algorithms that implement them include [1, 3, 5, 8, 9, 11, 12, 13, 14, 15, 16, 21, 24]. In most of this work, memory is modeled as a collection of items that are accessed through read and write operations. The study of correctness for shared memory with more general data types was initiated by Herlihy and Wing [17] Sequential consistency and other consistency conditions for ....

....of items that are accessed through read and write operations. The study of correctness for shared memory with more general data types was initiated by Herlihy and Wing [17] Sequential consistency and other consistency conditions for general data types has been studied by Attiya and Welch [5] and Attiya and Friedman [4] The rest of the paper is organized as follows. Section 2 introduces basic terminology that is used in the rest of the paper. Section 3 contains the definition of a sequentially consistent shared memory and introduces our new method for proving sequential consistency. ....

H. Attiya and J. Welch. Sequential consistency versus linearizability. ACM Transactions on Computer Systems, 12(2):91--122, 1994.

....programmability for performance. Sequential consistency was first defined by Lamport [19] in this paper, we use an alternative formulation proposed by Afek et al. 2] based on I O automata. Other papers exploring correctness conditions for shared memory and algorithms that implement them include [1, 3, 5, 8, 9, 11, 12, 13, 14, 15, 16, 21, 24]. In most of this work, memory is modeled as a collection of items that are accessed through read and write operations. The study of correctness for shared memory with more general data types was initiated by Herlihy and Wing [17] Sequential consistency and other consistency conditions for ....

....of items that are accessed through read and write operations. The study of correctness for shared memory with more general data types was initiated by Herlihy and Wing [17] Sequential consistency and other consistency conditions for general data types has been studied by Attiya and Welch [5] and Attiya and Friedman [4] The rest of the paper is organized as follows. Section 2 introduces basic terminology that is used in the rest of the paper. Section 3 contains the definition of a sequentially consistent shared memory and introduces our new method for proving sequential consistency. ....

H. Attiya and J. Welch. Sequential consistency versus linearizability. ACM Transactions on Computer Systems, 12(2):91--122, 1994.

....els. The strong memory models are close in behavior to a centralized memory, which makes simple to write programs with them (the return value of each read operation is rather predictable) However, it is widely accepted that strong memory models do not scale well with the number of processes [3, 6]. On the other hand, weaker memory models can be more eciently implemented, since they require less consistency overhead. This implies more possible return values for each read operation [2, 5, 8] which makes harder to write programs for these models. Hence, we are faced with a tradeo between ....

H. Attiya and J. Welch. Sequential consistency versus linearizability. ACM Transactions on Computer Systems, 12(2):91-122, 1994.

.... of the work described in [13] The proposed protocol can easily be tailored to implement the ROWA strategy (Read One replica, Write All) In that case, it encompasses distributed data management protocols based on an atomic broadcast primitive that have been designed in the no failure model (e.g. [2, 16]) Let us finally remark that votes can be dynamically modified by adopting techniques described in [7] 6 Conclusion Weighted voting techniques are used to increase the availability of replicated objects. This paper has shown that an atomic broadcast primitive can be used to implement weighted ....

H. Attiya and J. Welch, Sequential Consistency versus Linearizability. ACM TOCS, 12(2):91-122, 1994.

....programmability for performance. Sequential consistency was first defined by Lamport [20] in this paper, we use an alternative formulation proposed by Afek et al. 2] based on I O automata. Other papers exploring correctness conditions for shared memory and algorithms that implement them include [1, 3, 5, 8, 9, 11, 12, 13, 15, 16, 17, 21, 24]. In most of this work, memory is modeled as a collection of items that are accessed through read and write operations. The study of correctness for shared memory with more general data types was initiated by Herlihy and Wing [18] Sequential consistency and other consistency conditions for ....

....of items that are accessed through read and write operations. The study of correctness for shared memory with more general data types was initiated by Herlihy and Wing [18] Sequential consistency and other consistency conditions for general data types has been studied by Attiya and Welch [5] and Attiya and Friedman [4] The rest of the paper is organized as follows. Section 2 introduces basic terminology that is used in the rest of the paper. Section 3 contains the definition of a sequentially consistent shared memory and introduces our new method for proving sequential consistency. ....

H. Attiya and J. Welch. Sequential consistency versus linearizability. ACM Transactions on Computer Systems, 12(2):91--122, 1994.

....implementation. Our specifications and resulting implementation have the following distinct features: Rigor: We provide a formal definition of basic consistency conditions, given from the application point of view, as requirements on possible ordering of clients local histories. As discussed in [5], this implementation independent approach yields more rigorous definitions, and it is easier to prove program correctness with such definitions than with operational definitions. Modularity: Our conditions can be combined in various ways to yield guarantees with different levels of strength and ....

....than with operational definitions. Modularity: Our conditions can be combined in various ways to yield guarantees with different levels of strength and complexity. This approach allows the known tradeoff between the strength of the consistency semantics and the overhead it imposes (cf. [5]) to be taken into consideration when configuring the set of consistency guarantees for a particular application. For users of our service, this means that they have more freedom in choosing the exact consistency semantics they need. From the implementation standpoint, this yields a more modular ....

[Article contains additional citation context not shown here]

H. Attiya and J. Welch. Sequential Consistency versus Linearizability. ACM Transactions on Computer Systems, 12(2):91--122, May 1994.

....generally, the consistency model specifies what event orderings are legal when several processes are accessing a common set of objects. 23 Several consistency models for multiprocessor shared and cached memory systems have been proposed in the literature. Examples include sequential consistency [Lamp79, Feke95, Atti94], processor consistency [Good89] and weak consistency [Dubo86, Adve98, Aham95, Ghar98, Atti92, Yehu93, Atti98 and Torr99] In database field, there were also similar models for assuring consistency in distributed database models. The conditions of sequential consistency in shared memory systems ....

....mechanism [Atti92] 1. what conditions can be implemented efficiently, 2. what conditions can be used conveniently, 3. and what conditions support faster programs. Most of the research on this subject addressed strong consistency conditions like sequential consistency and linearizability [Atti92, Atti94, Lamp78, Birm91, and Yehu93]. These conditions guarantee that operations appear to be executed atomically, in some sequential order that is consistent with the order seen at individual sites. Sequential consistency and linearizability provide a clean semantics for the executions of operations; programming conditions is ....

[Article contains additional citation context not shown here]

Hagit Attiya, Jennifer L. Welch. "Sequential consistency versus linearizability. " Transactions on Computer Systems Vol. 12, No. 2 (May 1994), Pages 91-122.

....objects. Specifically, under these execution constraints, it is necessary and sufficient to ensure legality of reads to guarantee sequential consistency (and linearizability) Finally, we provide algorithms for ensuring proposed consistency conditions in a distributed system. Several papers [2, 4, 17, 20] have proposed sequentially consistent implementations for read write objects. Attiya and Welch [4] provide sequentially consistent and linearizable implementations for read write objects, FIFO queues and stacks. In addition, they also give an analysis of the response time of their ....

....of reads to guarantee sequential consistency (and linearizability) Finally, we provide algorithms for ensuring proposed consistency conditions in a distributed system. Several papers [2, 4, 17, 20] have proposed sequentially consistent implementations for read write objects. Attiya and Welch [4] provide sequentially consistent and linearizable implementations for read write objects, FIFO queues and stacks. In addition, they also give an analysis of the response time of their implementations. But their implementation for linearizability assumes that clocks are perfectly synchronized and ....

Hagit Attiya and Jennifer L. Welch. "Sequential Consistency versus Linearizability". ACM Transactions on Computer Systems, 12(2):91--122, May 1994.

....database protocols use serializability adapted to replicated scenarios: one copy serializability [BHG87] It is possible to use other correctness criteria[KA98] but, in all cases, the basis for correctness are data dependencies. Distributed systems use linearisability and sequential consistency [AW94] Linearisability is strictly stronger than sequential consistency. Linearisability is based on real time dependencies, while sequential consistency only considers the order in which operations are performed on every individual process. Sequential consistency allows, under some conditions, to ....

H. Attiya and J. Welch. Sequential consistency versus linearizability. ACM Transactions on Computer Systems, 12(2):91--122, May 1994.

....Computer Science University of Calgary Calgary, AB, T2N 1N4 Canada (e mail: fhigham,jolantag cpsc.ucalgary.ca) Abstract. Atomic broadcast is a powerful communication primitive, which is applied in a natural way in sequentially consistent implementations of various data structures (see [1]) Unfortunately the atomic broadcast algorithm as presented is incorrect. Here some ways of correcting it are analyzed under assumptions of blocking versus nonblocking communication. Two atomic broadcast algorithms are proposed: one correct if it is used in a blocking manner, the other ....

....of blocking versus nonblocking communication. Two atomic broadcast algorithms are proposed: one correct if it is used in a blocking manner, the other correct unconditionally. The latter algorithm, when used in the sequentially consistent implementations of data structures proposed in [1], ensures the same time complexity as claimed in the original application. Keywords: Sequential consistency, Atomic broadcast, Blocking execution. 1. Introduction The overhead of ensuring linearizability versus sequential consistency of shared objects in a message passing system under various ....

[Article contains additional citation context not shown here]

H. Attiya and J. Welch. Sequential consistency versus linearizability. ACM Transactions on Computer Systems, 12(2):91--122, 1994.

.... multiple consistency conditions The plethora of shared memory consistency conditions mentioned in the introduction can be confusing [23] While linearizability and sequential consistency provide the clearest semantics for the programmer, they are also limited in their achievable efficiency [6] and fault tolerance [14, 17] The weaker consistency conditions can be implemented more efficiently and more robustly, but suffer from less clear semantics. Programming methodologies and proof systems have been designed for the strong conditions, but are lacking for the weak conditions (but see ....

Hagit Attiya and Jennifer L. Welch. Sequential consistency versus linearizability. ACM Transactions on Computer Systems, 12(2):91--122, May 1994.

....do not overlap in time is preserved. Sequential consistency also provides atomic operations, but the global time order is not necessarily preserved. These strong shared memories can have a significant negative impact on the performance of applications, and can limit the scalability of a system [6]. Weaker conditions have been defined in an attempt to ameliorate such effects, such as release consistency [13] causal memory [3] processor consistency [14] pipelined RAM [25] and mixed consistency [1] The capabilities and characteristics of the various memory consistency conditions has ....

Hagit Attiya and Jennifer L. Welch. Sequential consistency versus linearizability. ACM Transactions on Computer Systems, 12(2):91--122, May 1994.

....causal consistency [3, 4] lazy release consistency [29] entry consistency [14] and hybrid consistency [22] In contrast to our service, such systems are geared towards high performance computing, and generally assume non faulty environments and fast local communication. We refer the reader to [7, 8, 22, 53] for other applied and theoretical studies of consistency strategies. The Globe system [49] follows an approach similar to CASCADE by providing a flexible framework for associating various replication coherence models with distributed objects. Among the coherence models supported by Globe are the ....

....system concurrently maintains several copies of the same object, it should also guarantee mutual consistency of these copies. Different levels of consistency are widely known, e.g. 13, 14, 23, 28, 33] In general, the stronger the consistency level, the higher latency its implementation incurs [8]. In order to make our service as flexible as possible, we support several consistency policies. The set of supported policies is motivated by the guarantees presented in [48] and [30] Following [48] we present 6 consistency guarantees that can be combined together to form a consistency model. ....

H. Attiya and J. Welch. Sequential consistency versus linearizability. ACM Transactions on Computer Systems, 12(2):91--122, May 1994.

....exclusive) lock: such locks can easily be implemented by read and write quorums [28, 9] 6 Actually, if there are no query transactions, Rule R1 implements linearizability [11] on a set of non duplicated objects. Linearizability is a consistency criterion more constraining than serializability [4]; it is equivalent to strict serializability. INRIA From Serializable to Causal Transactions for Collaborative Applications 15 ffl Rule R2. This rule ensures one copy equivalence for the set of all objects. It requires dissemination of update messages be consistent, i.e. all update messages be ....

....is used by the underlying protocol (described in Section 4.1) that has to be modified accordingly. An update transaction executed by process P i finishes when P i is delivered the corresponding update message it sent with the ABCAST primitive (a similar protocol, called fast read, is used in [4] to implement sequential consistency) It is important to note that, due to the underlying causal consistency layer, query transactions are not constrained by acquisition of read tokens, and so are wait free as in previous protocols. Moreover, let us note that classical lock based protocols ....

H. Attiya and J. L. Welch. Sequential Consistency versus Linearizability. ACM Trans. on Computer Systems, 12(2):91--122, May 1994.

....database protocols use serialisability adapted to replicated scenarios: one copy serialisability [6] It is possible to use other correctness criteria[17] but, in all cases, the basis for correctness are data dependencies. Distributed systems use linearisability and sequential consistency [5]. Linearisability is strictly stronger than sequential consistency. Linearisability is based on real time dependencies, while sequential consistency only considers the order in which operations are performed on every individual process. Sequential consistency allows, under some conditions, to read ....

H. Attiya and J. Welch. Sequential consistency versus linearizability. ACM Transactions on Computer Systems, 12(2):91--122, May 1994.

....and the processor orders are satisfied. If each processor recognizes (a part of) the ww list, local read operations may be processed at each processor. One way to achieve this is to provide each processor with local processor memory which stores copies of DSM objects. Protocols presented in [2, 5, 7] 3 An execution is said to be legal if each read operation reads the value written by the immediately preceding write operation on the same object. are classified in this category. In these protocols, the WW synchronization is achieved by atomic broadcasting of write operations. Legality is ....

....This implies that read operations may be processed locally at each processor. This approach requires that each processor holds (a part of) the ww list. One way to achieve this is to provide each processor with a local processor memory which stores copies of DSM objects. Protocols presented in [2, 5, 7] fall in this category. In order for every processor to recognize the ww list, each write operation needs to be propagated to other processors. In implementations of sequential consistency, however, processors do not have to recognize the current ww list in real time. Instead, at any time, ....

[Article contains additional citation context not shown here]

Attiya, H. and Welch, J. Sequential consistency versus linearizability. ACM TOCS, 12(2):91--122, 1994.

....are linearizability [21, 30] and sequential consistency [29] which guarantee that the DSM behaves like a single physical memory module. However, such strong shared memories can have a significant negative impact on the performance of applications, and can limit the scalability of a system [8]. Weaker conditions have been defined that allow optimizations that potentially provide better performance, but can lead to nonsequential behavior. These conditions include processor consistency [2, 18] release consistency [17] pipelined RAM (PRAM) 33] hybrid consistency [7] mixed consistency ....

.... some weak memories [6] 3 For data race free programs, many weak consistency conditions are equivalent to sequential consistency [1, 3, 17, 41] For linearizability and sequential consistency, the sum of the worst case read and write response times is at least as great as the network delay [8]. In a partially synchronous model, where message delays are bounded, linearizability incurs higher latency than sequential consistency in the worst case [8, 37, 38] Although some latency bounds are known for the synchronous and partially asynchronous cases, many real distributed systems are ....

[Article contains additional citation context not shown here]

H. Attiya and J. L. Welch. Sequential consistency versus linearizability. ACM Trans. Comput. Syst., 12(2):91--122, May 1994.

.... protocols (e.g. 32] and the references therein) work on computing with faulty shared memories [4] work on determining minimal ordering constraints needed to preserve sequential consistency [33] and work on comparing implementations of sequential consistency versus linearizability (e.g. [8]) However, none of this work addresses the general testing questions considered in this paper. 1.1. The testing problems. During an execution of a parallel program on a given multiprocessor, processors request to read or write particular shared memory 2 S 1 : write(a; 0) write(b; 1) read(a; ....

....As discussed in Section 1, linearizability is more restrictive than sequential consistency, in that a legal schedule must respect the time intervals for the operations. This added constraint makes implementations of linearizability provably slower than implementations of sequential consistency [8]. In the interest of memory system performance, shared memory multiprocessors such as the Kendall Square KSR1 [12] support sequential consistency instead of linearizability. On the other hand, linearizability has the advantage over sequential consistency that each address, or more generally each ....

H. Attiya and J. L. Welch, Sequential consistency versus linearizability, ACM Trans. on Computer Systems, 12 (1994), pp. 91--122.

....issued by P 1 . It is also easy to see that the history c H 2 (Figure 2) is not sequentially consistent, as no legal linear extension of c H 2 can be built. 3. 2 Protocols Various cache based protocols implementing sequential consistency have been proposed in the context of parallel machines [2, 7, 27]. In most of these protocols, 3 In his definition, Lamport assumes that the process order relation defined by the program (see point (2) of the definition) is maintained in the equivalent sequential execution, but not necessarily in the execution itself. As we do not consider programs but only ....

....1 (x)0 r 1 (y)2 r 2 (x)1 w 2 (x)1 w 2 (y)2 Figure 1: A sequentially consistent history c H 1 every processor local memory contains a copy of the whole shared memory. Read operations are local, while write operations issued by processes are totally ordered. In [2] and in the fast read protocol of [7], this total order is built by an underlying atomic broadcast primitive (messages sent with this primitive are delivered in the same order to each procesor [14] Read operations issued by a process are appropriately scheduled by its processor in order to ensure their legality. In the protocol ....

[Article contains additional citation context not shown here]

H. Attiya and J.L. Welch. Sequential consistency versus linearizability. ACM Transactions on Computer Systems, 12(2):91--122, 1994.

....24] seems to have launched investigations into relaxations of sequential consistency. Since then there has been substantial research on a large selection of memory consistency models. Most closely related to our research are several papers presenting formal descriptions of specific memory models [22, 2, 1, 4, 5, 8, 14, 13, 16, 20, 27, 30] and papers that present various formalisms for describing memories and reasoning about them [3, 6, 7, 4, 11, 15, 25, 26, 17, 29, 33] Our work has benefited from all of these papers. For an extensive bibliography on memory consistency models see the online listing at the University of Alberta ....

H. Attiya and J. L. Welch. Sequential consistency versus linearizability. In Proc. 3rd ACM Symp. on Parallel Algorithms and Architectures, pages 304--315, 1991.

....some mechanism for merging the changes made by the multiple writers. Systems that require annotations provide good performance, but the application writer has the onus of providing the protocol specification for the data in the program. 2.2. 2 Weakly Consistent Systems Several researchers [9, 41] have shown that a sequentially consistent memory cannot scale. This is because any implementation of sequentially consistent memory cannot avoid either a read or write operation from incurring the network latency cost of communicating with some other processor. Weakly consistent systems improve ....

H. Attiya and J. L. Welch. Sequential consistency versus linearizability. ACM Transactions on Computer Systems, 12(4), August 1994.

....enjoys a number of nice properties, like, e.g. locality 1 , and this has made it quite attractive for different applications, such as concurrent programming, multiprocessor operating systems, distributed file systems, etc. where concurrency is of primary interest. Attiya and Welch ([3]) initiated a comparative study of the impact of the strength of correctness guarantees provided by sequential consistency and linearizability on the cost of supporting them. Specifically, they considered caching implementations of read write objects in non bused distributed systems, and took as ....

....In this paper, we continue this study and present upper bounds on the same cost for new, efficient linearizable implementations in order to further enlighten the advantages of linearizability over other, seemingly cheaper , correctness conditions, such as sequential consistency. We follow [3] and consider a model consisting of a collection of application programs running concurrently and communicating through virtual shared memory, which consists of a collection of read write objects. These programs are running in a distributed system consisting of a collection of processes located at ....

[Article contains additional citation context not shown here]

H. Attiya and J. Welch, "Sequential Consistency versus Linearizability," in Proceedings of the 3rd ACM Symposium on Parallel Algorithms and Architectures, pp. 304--315, 1991.

....etc. Strong conditions, such as linearizability and sequential consistency, give the DSM the clear semantics of a single physical memory module. However, such conditions can have a significant negative impact on the performance of applications. For example, the scalability of the system is limited [1], and some operations must suffer network latencies in order to guarantee correctness [1, 2] Weak conditions ameliorate these effects to some degree by offering opportunities for optimizations. However, the semantics of weak memories are less clear than those of strong memories. We have extended ....

....the clear semantics of a single physical memory module. However, such conditions can have a significant negative impact on the performance of applications. For example, the scalability of the system is limited [1] and some operations must suffer network latencies in order to guarantee correctness [1, 2]. Weak conditions ameliorate these effects to some degree by offering opportunities for optimizations. However, the semantics of weak memories are less clear than those of strong memories. We have extended the existing analyses of DSM consistency conditions by investigating achievable levels of ....

ATTIYA, H., AND WELCH, J. L. Sequential consistency versus linearizability. ACM Trans. Comput. Syst. 12, 2 (May 1994), 91--122.

....with prede ned crash patterns. The proposed protocol can easily be tailored to implement the ROWA strategy (Read One replica, Write All) In that case, it encompasses distributed data management protocols based on an atomic broadcast primitive that have been designed in the no failure model (e.g. [2, 17]) Let us nally remark that votes can be dynamically modi ed by adopting techniques described in [8] 6 Related Problems 6.1 Consensus vs Atomic Broadcast In this paper we have shown how to transform a Consensus algorithm for the asynchronous crash recovery model into an atomic broadcast ....

Attiya H. and Welch J., Sequential Consistency versus Linearizability. ACM TOCS, 12(2):91-122, 1994.

....of items that are accessed through read and write operations. The study of correctness for shared memory with more general data types was initiated by Herlihy and Wing [22] Sequential consistency and other consistency conditions for general data types has been studied by Attiya and Welch [8] and Attiya and Friedman [7] The algorithms that provide each layer in this paper are closely related to some in the literature. In the lower layer, context multicast is provided by placing sequence numbers in messages, and delaying processing until after the receipt of messages that should be ....

....used by Welch [36] and Neiger and Toueg [31] which delay point to point messages based on sequence numbers. The algorithm in the upper layer updates all copies, and reads any copy; this is folklore from the database community, where operation ordering is managed by locking. Attiya and Welch [8] proved that this algorithm provides sequential consistency when run over a totally ordered broadcast communication service. Our proof technique for sequential consistency based on a partial order is similar to a method used by Attiya and Friedman [7] to prove hybrid consistency. 1.2 Overview of ....

[Article contains additional citation context not shown here]

H. Attiya and J. Welch. Sequential consistency versus linearizability. ACM Transactions on Computer Systems, 12(2):91--122, 1994.

....to using linearizability rather than sequential consistency, since a straightforward way of turning a sequentially consistent implementation into a linearizable one is to add delays, making each operation wait until the effect of the operation is observable by every other thread. Attiya and Welch [AW91] prove that linearizable memory is, for a particular abstract machine model, more expensive than sequentially consistent memory. For these reasons we assume that a machine provides sequentially consistent built in operations, but require that software operations built on top are linearizable. The ....

Hagit Attiya and Jennifer L. Welch. Sequential consistency versus linearizability. In Symposium on Parallel Algorithms. ACM SIGACT-SIGARCH, 1991.

....in the context of 2 A linear extension b S = S; S ) of a partial order b H = H; H ) is a topological sort of this partial order, i.e. i) S = H, ii) op 1 H op 2 ) op 1 S op 2 ( b S maintains the order of all ordered pairs of b H) and (iii) S defines a total order. parallel machines [2, 5, 14]. In most of these protocols, every processor local memory contains a copy of the whole shared memory. Read operations are local, while write operations issued by processes are totally ordered. In [2] and in the fast read protocol of [5] this total order is built by an underlying atomic broadcast ....

....(iii) S defines a total order. parallel machines [2, 5, 14] In most of these protocols, every processor local memory contains a copy of the whole shared memory. Read operations are local, while write operations issued by processes are totally ordered. In [2] and in the fast read protocol of [5], this total order is built by an underlying atomic broadcast primitive (messages sent with this primitive are delivered in the same order to each procesor [6] Read operations issued by a process are appropriately scheduled by its processor in order to ensure their legality. 4 Atomic ....

[Article contains additional citation context not shown here]

H. Attiya and J.L. Welch. Sequential consistency versus linearizability. ACM Transactions on Computer Systems, 12(2):91--122, 1994.

No context found.

Attiya H. and Welch J.L., Sequential Consistency versus Linearizability. ACM Transactions on Computer Systems, 12(2):91-122, 1994.

No context found.

Attiya H. and Welch J.L., Sequential Consistency versus Linearizability. ACM Transactions on Computer Systems, 12(2):91-122, 1994.

No context found.

H. Attiya and J. Welch. Sequential Consistency versus Linearizability. ACM Transactions on Computer Systems, 12(2):91--122, May 1994.

No context found.

Attiya, H. and Welch, J. L. 1994. Sequential consistency versus linearizability. ACM Transactions on Computer Systems 12, 2 (May), 91--122.

No context found.

H. Attiya and J.L. Welch. Sequential consistency versus linearizability. ACM Transactions on Computer Systems, 12(2):91-122, 1994.

No context found.

H. Attiya and J.L. Welch. Sequential consistency versus linearizability. Technical Report 674, Department of Computer Science, The Technion, October 1991.

No context found.

H. Attiya and J.L. Welch. Sequential consistency versus linearizability. ACM Trans. on Computer Sys., 12(2):91--122, May 1994.

No context found.

Attiya, H. and Welch, J. L. (1991) Sequential consistency versus linearizability. In Proc. 3rd Ann. ACM Symp. on Parallel Algorithms and Architectures, pp. 304--315.

No context found.

Hagit Attiya and Jennifer L. Welch. Sequential consistency versus linearizability. ACM Transactions on Computer Systems, 12(2):91-- 122, May 1994.

No context found.

ATTIYA, H. AND WELCH, J. L. 1994. Sequential consistency versus linearizability. ACM Trans. Comput. Syst. 12, 2 (May), 91--122.

No context found.

H. Attiya and J. Welch. Sequential consistency versus linearizability. ACM Transactions on Computer Systems, 12(2):91--122, May 1994.

No context found.

H. Attiya and J. L. Welch. Sequential consistency versus linearizability. ACM Trans. Comput. Syst., 12(2):91--122, May 1994.

No context found.

Hagit Attiya and Jennifer L. Welch. Sequential consistency versus linearizability. ACM Transactions on Computer Systems, 12(2), pages 91--122, May 1994.

First 50 documents  Next 50

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