G.T.J. Wuu and A. Bernstein, "Efficient Solutions to the Replicated Log and Dictionary Problems," in Proc. 3rd ACM Symposium on Principles of Distributed Computing, 1984.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....must be included in every message until the sender knows that the recipient of the message has received that record. In addition to tolerating an unreliable network, this communication method tolerates a network that experiences congestion and dynamically changing topology. Wuu and Bernstein [14] combine logs and vector clocks to solve the distributed dictionary problem efficiently. Each site S i keeps a two dimensional time table T i , which corresponds to S i s most recent knowledge of the vector clocks at all sites. Each time table ensures the following time table property: if T i [k; ....

....S i knows that S k has received the records of all events at S j up to time v (which is the value of S j s local clock) To reduce communication it is desirable for a site to know which sites have received the record of a particular event. To this end a predicate HasRecvd, is defined as follows [14]: HasRecvd(T i ; t; S k ) T i [k; Site(t) time(t) where t is an event, Site(t) is the site at which t occurred, and time(t) is the local time at Site(t) when t occurred. When HasRecvd(T i ; t; S k ) is true S i knows that S k has received a record of event t. When a site S i performs an ....

[Article contains additional citation context not shown here]

G. T. Wuu and A. J. Bernstein. Efficient Solutions to the Replicated Log and Dictionary Problems. In Proceedings of the Third ACM Symposium on Principles of Distributed Computing, pages 233--242, August 1984. 19

....must be included in every message until the sender knows that the recipient of the message has received that record. In addition to tolerating an unreliable network, this communication method tolerates a network that experiences congestion and dynamically changing topology. Wuu and Bernstein [12] combine logs and vector clocks to solve the distributed dictionary problem efficiently. Each site S i keeps a two dimensional time table T i , which corresponds to S i s most recent knowledge of the vector clocks at all sites. Each time table ensures the following timetable property: if T i [k; ....

....S i knows that S k has received the records of all events at S j up to time v (which is the value of S j s local clock) To reduce communication it is desirable for a site to know which sites have received the record of a particular event. To this end a predicate HasRecvd, is defined as follows [12]: HasRecvd(T i ; t; S k ) T i [k; Site(t) time(t) where t is an event, Site(t) is the site at which t occurred, and time(t) is the local time at Site(t) when t occurred. When HasRecvd(T i ; t; S k ) is true S i knows that S k has received a record of event t. When a site S i performs an ....

[Article contains additional citation context not shown here]

G. T. Wuu and A. J. Bernstein. Efficient Solutions to the Replicated Log and Dictionary Problems. In Proceedings of the Third ACM Symposium on Principles of Distributed Computing, pages 233--242, Aug. 1984.

....While vector clocks provide a mechanism to represent causality information, the vector clock representation is not readily usable in optimistic protocols due to the forward and backward motion of time. In matrix clock representation, a process maintains a n n matrix of non negative integers [18,21]. A process maintains this matrix as its clock value. This representation has all the properties of vector clocks. In addition, a process P i knows the minimum time value of process P k that is known to every other process P j . This allows processes to discard obsolete information received from ....

G. Wuu and A.J.Bernstein. Efficient solutions to the replicated log and dictionary problems. In Proceedings of 3rd ACM Symposium on Principles of Distributed Computing (PODC), pages 233--242. ACM Press, 1984.

....weak consistency are Internet news, air traffic control, airline reservation and stock exchanges. Grapevine [SBN84] and the Global Name Service [Lam86] were among the first systems to use weak consistency. Other weak consistency protocols were presented in [DGH 87, DGP90, Gol92, LLS90, QP93, WB84] These protocols are useful and interesting; however, they require a node to communicate with all other nodes. This would be acceptable for small networks with a few replicas, but it will incur a large communication overhead for wide area networks like the Internet. This paper evaluates the ....

G. Wuu and A. Bernstein. Efficient Solutions to the Replicated Log and Dictionary Problems. In Proceedings of the third ACM Symposium on Principles of Distributed Computing, pages 233--242, August 1984.

....working in the areas of distributed computing and databases, respectively. We believe that such interactions are highly beneficial to both communities. Interestingly, when looking in the past, we can observe such beneficial influences. We cite a few: 1) the area of epidemic communication [8, 18, 24] that first appeared in the distributed computing community and has then been successfully applied to databases (e.g. 3] 2) the domain of atomic broadcast multicast and group communications [1, 15, 16, 21] and (3) as a problem not directly related to databases, publish subscribe systems [5, ....

Wuu G. T. and Bernstein A. J., Efficient Solutions to the Replicated Log and Dictionary Problems. Proc. of the 3rd ACM Symposium on Principles of Distributed Computing, pp. 233-242, August 1984.

....from a replica s log once that write has fully propagated to all other replicas. Determining which writes have fully propagated can be done by running a distributed snapshot algorithm to establish a cutoff timestamp [17] or by having replicas maintain an acknowledgment vector [4] or timetable [1, 12, 21] of which replicas have received what writes, The problem with these approaches is that a single, longdisconnected replica can cause the write logs at all other repllcas to grow indefinitely. Satin and Lynch noted this problem and proposed forcibly removing such sites from the replica set [17] ....

....[8] File systems like Coda [18] and Ficus [7] exchange updated files between servers or between clients and servers. The notion of reconciling logs of update operations held at various replicas, as is done in Bayou via the anti entropy protocol, has been discussed for some time in the literature [1, 17, 21] and Is used in some commercial database systems [5, 16] Oracle7, for instance, uses asynchronous RPCs to propagate transactions between a master and its snapshots or other masters [16] it does not, however, allow these transaction to propagate through intermediary servers. Rover also uses ....

G. T J. Wuu and A. J. Bernstein. Efficient solutions to the replicated log and dictionary problem. Proceedings Third ACM Symposium on Principles of Distributed Computing, Vancouver, B.C., Canada, August 1984, pages 233-242. 301

....during a network partition. For example, suppose a partition divided the network into sides A and B and that r is known at all replicas in A and also at all replicas in B. If no replica in A knows that r is known in B, there is nothing we can do. However, as was pointed out by Wuu and Bernstein [34], progress can be made if replicas include their copy of rs table in gossip messages and receivers merge this information with their own rs tables. In this way, each replica would get a more recent view of what other nodes know. 11 3.5. Analysis In this section we argue informally that our ....

....interacts. 6. Related Work Our work builds on numerous previous results in the area of highly available distributed systems and algorithms, including general replication techniques such as voting [11, 13, 2] and the primary copy method [1, 27, 26] Our work is also related to gossip schemes [12, 10, 34]. In this section, we focus on the most closely related work, namely providing high availability for applications where operations need not be ordered identically at all replicas. In this light, we compare our method with the relevant gossip schemes and with the work on ISIS. We also consider ....

[Article contains additional citation context not shown here]

Wuu, G. T. J., and Bernstein, A. J. Efficient Solutions to the Replicated Log and Dictionary Problems. In Proc. of the Third Anntl Symposium on Principles of Distributed Computing, pages 233-242. ACM, August, 1984.

.... Push NIS[73] Wide area directory Pull Push Oracle[58] Sybase Database Pull Roam[65] Palm[66] Coda[43] Mobile DB file system Push Usenet[69] Active Directory[49] Porcupine[67] Wide area data service Pull Bayou[61] Golding[25] Mobile DB file system Push Agrawal[2] Wuu[79] Wide area directory service Figure 1. Classification of existing optimistic replication systems. The space efficiency of a system is often as important as the performance. Replication algorithms must maintain several data structures as well as the object contents themselves the log of ....

....before the application update on all the replicas, but we 7 may mix the two with updates to unrelated modules. Such a constraint cannot be expressed using total update ordering mechanisms. Several systems address the shortcomings of total ordering by taking advantage of update commutativity [79, 63, 36, 32, 43]. For example, additions and subtractions on a numeric value or file creations in the same directory can be applied in any order to produce the same result. The idea of commutative updates can be extended to the concept of causality. Causality is a partial ordering, defined over updates, which ....

[Article contains additional citation context not shown here]

Gene T. J. Wuu and Arthur J. Bernstein. Efficient solutions to the replicated log and dictionary problems. In 3rd ACM Symp. on Princ. of Distr. Computing (PODC), pages 233-- 242, Vancouver, Canada, August 1984.

.... Wide area, Multi master Replication Services The replication algorithm developed in this dissertation is most closely related to wide area, multimaster replicated services, including Xerox Clearinghouse [46] Usenet [81, 92, 137] Active Directory [103] and several other experimental systems [2, 3, 163]. These algorithms achieve fault tolerance by being optimistic, and they alleviate the performance problem of mobile database systems by proactively pushing changes to peer replicas rather than polling them periodically. Pushing also supports many small objects replicated on a diverse set of ....

Gene T. J. Wuu and Arthur J. Bernstein. Efficient solutions to the replicated log and dictionary problems. In 3rd ACM Symp. on Princ. of Distr. Computing (PODC), pages 233--242, Vancouver, Canada, August 1984. 3.6.3

....order [12] In many applications, causal order based on logical time is not enough to capture the dependencies needed by the applications logic. Vector clocks bring progress over logical (scalar) timestamps by inducing an order that exactly reflects causal precedence [15] 23] Matrix clocks [4][24] extend vector clocks by capturing a global property, namely what A knows about what B knows about C. However, matrix clocks change propagation require O(n 2 ) control information and message size for a n node system [13] This precludes the use of matrix clocks for large scale systems (several ....

G.T.J. Wuu and A.J. Bernstein. Efficient Solutions to the Replicated Log and Dictionary Problems, Proc. 3nd ACM Symposium on PODC, 99. 233-242, 1984.

....(meaning they are strongly consistent) a b , C(a) C(b) This means that given two vector clocks C(a) and C(b) it is possible to deduce the causal relation between the two operations a and b. It is possible to augment the dimension of the logical clock even more, resulting in matrix clocks [Wuu84, Sari87]. Matrix clocks provide second order information to a process. It is a list of vector clocks, namely per process the last vector clock that was communicated to the current process. This information can be used to discard obsolete information in distributed systems. 5 Execution replay methods for ....

G. Wuu and A. Bernstein. Efficient Solutions to the Replicated Log and Dictionary Problems. Proc. 3rd ACM Symp. Principles Distributed Computing, 233--242, New York, 1984. ACM Press.

....must be sent to other sites during propagation. The timing information can also be used to garbage collect the update records when they are no longer needed. Most prior applications of epidemic communication have only supported commutative operations, e.g. a dictionary with insert and delete [13] or single item update operations [5, 7] We have recently developed an approach that uses epidemic update to support multi operation transactions with one copy serializability guarantees [1, 11] In this approach, a transaction, t, executes locally under a local concurrency control mechanism, ....

....date when it reconnects. 6. S j resumes sending and receiving epidemic messages and acts as a proxy for S i . In order to explain the handling of transactions when a proxy has been given, we will represent the timing information with vector clocks and the time table defined by Wuu and Bernstein [13]. Time is recorded using a vector clock [8] that captures the causal order of events and ensures the following property: 8e; f 2 E e f iff Time(e) Time(f) Note that if Time(e) and Time(f) are incomparable, denoted Time(e) Time(f) then events e and f are concurrent. Each member S i ....

[Article contains additional citation context not shown here]

G. T. Wuu and A. J. Bernstein. Efficient Solutions to the Replicated Log and Dictionary Problems. In Proceedings of the Third ACM Symposium on Principles of Distributed Computing, pages 233--242, Aug. 1984.

....process. In the first step, a replica learns about the state of the other replicas. In the second step, the replicas gossips this information to the other replicas, who now learn what other replicas know about themselves. See Guy [3] for further details. Acknowledgments have also been used by Wuu [17] and Ladin [9] The difference between their work and ours is that they use acknowledgments for garbage collection at the replica servers while we use them at the clients also. 6 Experiments and Evaluation We have implemented view consistency as a stackable file system layer over the Ficus file ....

Gene T. J. Wuu and Arthur J. Bernstein. Efficient solutions to the replicated log and dictionary problems. In Proceedings of the Third Annual ACM Symposium on Principles of Distributed Computing, August 1984.

....same way by all processes; put another way, two concurrent write operations can be perceived in distinct orders by different processes. Several implementations of CC in various contexts and for different types of applications (such as CSCW and distributed dictionary management) have been explored [1, 5, 8, 16]. Systems that support ordered group communication provide users with an easy way to implement CC. For example, a communication primitive such as the ISIS cbcast primitive [4] which ensures all messages are delivered according to causal order) can be used in the following way. Each write ....

....size of the control information carried by this message is O(jY j n) The set Y has been defined at line (4.1) in Section 3. 5) In the worst case, this size is O(m n) and is similar to the one encountered in protocols that ensure causal delivery of all the messages exchanged during a computation [4, 5, 12, 16]. This price seems to be inherent to causality. Reduction of the size of the control information requires to add more synchronization but then we get a stronger consistency criterion 4 . So, the advantage of our protocol over classical causal delivery protocols lies in the possibility to discard ....

G.T. Wuu and A.J. Berntsein. Efficient Solutions to the Replicated Log and Dictionary Problems. In Proc. 3rd ACM Symposium on Principles of Distributed Computing, Montreal, 1984, pp. 233-242. Irisa

....the receiver local state systematically entails a change to a new local state for the receiver process; so in our context Gamma , i.e. w , is acyclic) PI n796 8 E. Fromentin and M.Raynal 4 Detecting precedences 4. 1 Vector clock Sketched by several authors to solve particular problems [15], vector clocks have been formally introduced as a general mechanism by Fidge [4] and Mattern [11] They constitute an operational device to encode causal precedence and concurrency of events. We use here such vector clocks with a slight modification: only relevant events are accounted. More ....

G.T. Wuu and A.J. Bernstein. Efficient solutions to the replicated log and dictionnary problems. In Proc. 3rd ACM Symposium on PODC, pages 233-- 242, 1984.

....in this paper we explore a weaker consistency criterion for distributed services, namely causal consistency. This consistency criterion takes into account only causality relations among accesses to objects. It appears that it has two main advantages: 1) it is meaningful for several applications [1, 5, 11, 22], and (2) protocols that implement it require only weak synchronization that permits the design of efficient implementations. One aim of this paper is the design of such a protocol for causally consistent distributed services. We consider a distributed model in which three sets of nodes are ....

....We will see in Section 4.1 how this dependence can be tracked in the case of causal consistency. PI n1039 10 M. Raynal 4. 1 Tracking Causality The most fundamental data structure to track causality relations and implement causally consistent objects is a two dimensional matrix of integers [5, 22, 3, 17]. Although a version vector is sufficient for tracking causality in the full replication case that was discussed previously, partial replication requires that updates to objects be recorded separately. For example, consider server node s i with copies of objects x a and x b , and server node s j ....

G.T. Wuu and A.J. Berntsein. Efficient Solutions to the Replicated Log and Dictionary Problems. In Proc. 3rd ACM Symposium on Principles of Distributed Computing, Montreal, 1984, pp. 233-242. PI n1039

....) c t 5 b(X 0 )b(Y 0 )r(Z 3 ) w(Z 5 ) c Figure 2. Relaxed Check out Mode Using Browse locks Each member maintains a vector clock that captures the causal order of events (transaction pre commits) In addition, each member S i keeps a 2 dimensional time table T i as defined by Wuu and Bernstein [13], which corresponds to S i s most recent knowledge of the events at all members. This timing information is included in all epidemic messages. The time table ensures the following time table property: if T i [k; j] v then S i knows that S k has received the records of all events at S j up to ....

G. T. Wuu and A. J. Bernstein. Efficient Solutions to the Replicated Log and Dictionary Problems. In Proceedings of the Third ACM Symposium on Principles of Distributed Computing, pages 233--242, Aug. 1984. 9

....approach that was first promoted by the networks and distributed operating systems community is to execute operations locally without synchronization with other sites, and after termination, the This work was partially supported by NSF grant CCR97 12108. updates are propagated to other copy sites [20, 14, 17]. In this approach, changes are propagated throughout the network using an epidemic approach [17] where updates are piggy backed on messages, thus ensuring that eventually all updates are propagated throughout the system. The epidemic approach or its variant, referred to as asynchronous logging, ....

....that eventually all sites incorporate all the operations that have occurred in the system. Due to the unreliable nature of the communication medium, a record must be included in every message until the sender knows that the recipient of the message has received that record. Wuu and Bernstein [20] combine communication logs and vector clocks to solve the distributed dictionary problem efficiently. Each site S i keeps a two dimensional time table T i , which corresponds to S i s most recent knowledge of the vector clocks at all sites. Each time table ensures the following time table ....

[Article contains additional citation context not shown here]

G.T. Wuu and A.J Bernstein, "Efficient Solutions to the Replicated Log and Dictionary Problems," Proceedings, Third ACM Symposium on Principles of Distributed Computing, p.233-343, August 1984. 18

....hash routing to let each node determine the exact change in the hash map. The replication mechanism used in Porcupine can be viewed as a variation of optimistic replication schemes, in which timestamped updates are pushed to peer nodes to support multi master replication [Agrawal et al. 1997; Wuu and Bernstein 1984]. Porcupine s total object update property allows it to use a single timestamp per object, instead of timestamp matrices, to order updates. In addition, since updates are idempotent, Porcupine can retire updates more aggressively. These differences make Porcupine s approach to replication simpler ....

WUU, G. T. J. AND BERNSTEIN, A. J. 1984. Efficient solutions to the replicated log and dictionary problems. In Proceedings of the 3rd Symposium on Principles of Distributed Computing. ACM, Vancouver, Canada, 233--242.

....selected sites until it has tried to share the update with too many sites that have already seen them. Unlike anti entropy, rumor mongering is not reliable since it is possible for a rumor to die out before all sites have been informed. A related communication mechanism is gossip messages [48]. Like anti entropy methods, every site eventually learns of each update. However, the gossip algorithm requires that each site maintain information about the states of other sites. Consequently, it can guarantee the causality property that if two updates are causally related, then if a site knows ....

G.T. Wuu and A.J. Bernstein. Efficient solutions to the replicated log and dictionary problems. In Proceedings of the Third ACM Symposium of Principles of Distributed Computing, pages 233--242, August 1984. 25

....maintenance problem as a replicated dictionary problem, with slightly (but significantly) different semantics. A timestamp vector was used to infer from a comparison of two dictionary replicas which entries had been inserted and which had been deleted. Allchin [1] and Wuu and Bernstein [14] expanded upon Fischer and Michael s approach to use two dimensional timestamp matrices to reduce the number of messages exchanged, with small variations in semantics. None of these works addressed the general problem of reclaiming resources of named replicated objects; they were concerned with ....

Gene T. J. Wuu and Arthur J. Bernstein. Efficient solutions to the replicated log and dictionary problems. In ACM Symposium on Principles of Distributed Computing, August 1984. 17

....Agreed queue. Simple optimizations can minimize the amount of state to be transfered. For instance, since the associated gossip messages carries the current round number of the late process, the state message can be made to carry only those messages that are not known by the recipient (see [12, 19]) 5.4 Sending Message Batches For better throughput, it may be interesting to let the application propose batches of messages to the Atomic Broadcast protocol, which are then proposed in batch to a single instance of Consensus. Unfortunately, the definition of Atomic Broadcast implies that ....

G. Wuu and A. Bernstein, Efficient Solutions to the Replicated Log and Dictionary Problems. Proc. 3rd Int. ACM Symposium on Principles of Distributed Computing (PODC'84), pp. 233-242, 1984. 20

....They recast the problem as a replicated dictionary problem, to focus on the basic insert and delete operations common to each. Unsynchronized, concurrent directory modifications are resolved via timestamps. Various inefficiencies in Fischer s work were addressed by Allchin [All83] Wuu [WB84] offered further improvements. Unfortunately, the successive improvements reduced communications complexity at the expense of storage complexity: each replica in Wuu s scheme is required to maintain a version matrix. Bloch [BDS84] utilized weighted voting in a serializable approach to directory ....

....flows around the ring three times. Phase one of the algorithm begins for all nodes in the first round trip. Phase one completes and phase two begins for all nodes during the second round trip. Phase two completes for all nodes during the third round trip. 98 Allchin [All83] and Wuu and Bernstein [WB84] expanded upon Fischer and Michael s approach to use two dimensional timestamp matrices to reduce the number of messages exchanged, with small variations in semantics. None of these works addressed the general problem of reclaiming resources of named replicated objects; they were concerned with ....

Gene T. J. Wuu and Arthur J. Bernstein. "Efficient Solutions to the Replicated Log and Dictionary Problems." In Proceedings of the Third Annual ACM Symposium on Principles of Distributed Computing, August 1984.

....The protocols using this approach range from the simple read one write all to variations on majority or quorum approaches. The second approach referred to as asynchronous or lazy replication has been used extensively with much success in the area of network and distributed operating systems [5]. In this approach, transactions (in the case of operating systems, these are typically single operations) are executed on the This work was partially supported by NSF grants CCR97 12108, EIA98 18320, and IIS98 17432. local replica. The effects of locally executed transactions are propagated ....

....will be applied in that order at all sites. Event logs and vector clocks are combined as follows. Each site S i keeps a two dimensional time table T i , which corresponds to S i s most recent knowledge of the vector clocks at all sites. Each time table ensures the following time table property [5]: if T i [k; j] v then S i knows that S k has received the records of all events at S j up to time v (which is the value of S j s local clock) In the epidemic update protocol [1] for supporting multi operation transactional semantics, each site maintains a local clock, a time table and an ....

G.T. Wuu and A.J Bernstein, "Efficient Solutions to the Replicated Log and Dictionary Problems," Proceedings, Third ACM Symposium on Principles of Distributed Computing, p.233-343, August 1984.

....specific operations. Sites exchange their respective event logs to keep each other informed about the operations that have occurred in the system. This information exchange ensures that eventually all sites incorporate all the operations that have occurred in the system. Wuu and Bernstein [15] combine logs and vector clocks as follows. Each site S i keeps a twodimensional time table T i , which corresponds to S i s most recent knowledge of the vector clocks at all sites. Each time table ensures the following time table property: if T i [k; j] v then S i knows that S k has received ....

G. T. Wuu and A. J. Bernstein. Efficient Solutions to the Replicated Log and Dictionary Problems. In Proceedings of the Third ACM Symposium on Principles of Distributed Computing, pages 233--242, August 1984.

....nodes. However, unlike previous load balancing studies that assumed complete independence of incoming tasks, we also balance the write traffic, taking message affinity into consideration. The replication mechanism used in Porcupine can be viewed as a variation of optimistic replication schemes [1, 27], in which timestamped updates are pushed to peer 13 nodes to support multi master replication. Porcupine s total object update property allows it to use a single timestamp per object, instead of timestamp matrices, to order updates. In addition, since updates are idempotent, Porcupine can retire ....

Gene T. J. Wuu and Arthur J. Bernstein. Efficient solutions to the replicated log and dictionary problems. In Proceedings of the 3rd Symposium on Principles of Distributed Computing, pages 233--242, Vancouver, Canada, August 1984. 15

....server directories are maintained redundantly on multiple sites and updates are incorporated in a lazy This research was partially supported by LANL under grant number 6863V0016 3A, by CALTRANS under grant number 65V250A, and by the NSF under grant numbers CDA94 21978 and CCR95 05807. manner [12, 33] through gossip messages [21, 17] epidemic propagation, and anti entropy [11] In this paper we use the epidemic communication model as the basis for a fault tolerant algorithm that supports transaction processing in replicated databases. In an epidemic system, sites perform update operations and ....

....exchange ensures that eventually all sites incorporate all the operations that have occurred in the system. Due to the unreliable nature of the communication medium, a record must be included in every message until the sender knows that the recipient of the message has received that record [33]. Wuu and Bernstein [33] combine logs and vector clocks to solve the distributed dictionary problem efficiently. Each site S i keeps a two dimensional time table T i , which corresponds to S i s most recent knowledge of the vector clocks at all sites. Each time table ensures the following ....

[Article contains additional citation context not shown here]

G. T. Wuu and A. J. Bernstein. Efficient Solutions to the Replicated Log and Dictionary Problems. In Proceedings of the Third ACM Symposium on Principles of Distributed Computing, pages 233--242, August 1984.

....a logical tree structure. Each of these subtransactions commits locally, with the provision that if the root of the tree commits, all its children commit too. In this approach, persistent transmission of messages is required, for instance, by mean of stable queues [7] or gossip mechanisms [22] [37]. Another approach that avoids two phase commit is given in [29] transactions commit at a primary site and are propagated asynchronously. A partial order among transactions is defined for controlling update propagation; the use of a partial order is also proposed in [27] and in [15] in the ....

....events that originated in the system. However, only the insert and delete operations are allowed, and the vector does not give any indication of how up to date other sites of the system are. Therefore, this approach requires a site to send its entire copy of a dictionary at each message. In [37], the approach of [16] was extended by storing at each site a matrix, instead of a vector. The matrix at each site indicates how up to date all sites are, thus limiting the communication requirements among sites. As we will see, 37] has therefore many features in common with our paper; however ....

[Article contains additional citation context not shown here]

G.T.J. Wuu and A. Bernstein, "Efficient Solutions to the Replicated Log and Dictionary Problems," in Proc. 3rd ACM Symposium on Principles of Distributed Computing, 1984.

....and that at communication time, processes exchange time stamped copies of the global auxiliary variables. In this way, the state of the system, as represented by the global auxiliary variables, is diffused throughout the processes in the system. This concept is similar to the gossip messages of [14] except that, here, only a restricted history of size 1 is disseminated, whereas, in [14] an entire event history is treated. This section, mathematically describes how the global auxiliary variables are updated in CCSP and that the method of updating preserves the property of noninterference. ....

....variables. In this way, the state of the system, as represented by the global auxiliary variables, is diffused throughout the processes in the system. This concept is similar to the gossip messages of [14] except that, here, only a restricted history of size 1 is disseminated, whereas, in [14] an entire event history is treated. This section, mathematically describes how the global auxiliary variables are updated in CCSP and that the method of updating preserves the property of noninterference. 4.1 Definitions of Functions In this section, the method of updating the global auxiliary ....

G. Wuu and A. Bernstein. Efficient solutions to the replicated log and dictionary problems. In Proc. of the 3rd ACM Symposium on Principles of Distributed Computing, pages 233--242, 1984.

....result provides better performance: it requires many fewer messages than the process group approach, the messages are smaller, and it can tolerate network partitions. Our technique is based on the gossip approach first introduced by Fischer and Michael [9] and later enhanced by Wuu and Bernstein [36] and our own earlier work [16] We have extended the earlier work in two important ways: by supporting causal ordering for updates as well as queries and by incorporating forced and immediate operations to make a more generally applicable method. The implementation of causal operations is novel ....

....counter, and the initial (zero) timestamp contains zero in each part. Timestamps are partially ordered in the obvious way: t s ( t s . t s ) 1 1 n n Two timestamps t and s are merged by taking their component wise maximum. Multipart timestamps were used in Locus [30] and also in [9, 13, 16, 22, 36]. Both uids and labels are represented by multipart timestamps. Every update operation is assigned a unique multipart timestamp as its uid. A label is created by merging timestamps; a label timestamp t identifies the updates whose timestamps are less than or equal to t. The dependency relation is ....

[Article contains additional citation context not shown here]

Wuu, G. T. J., and Bernstein, A. J. Efficient Solutions to the Replicated Log and Dictionary Problems. Proc. of the Third Annual Symposium on Principles of Distributed Computing, ACM, August, 1984, pp. 233-242. i

.... s knowledge of the value of vtnc at server S j ; i.e. S i knows that versions with version numbers less than or equal to V i [j] vtnc cannot be created at S j . Similarly, V i [j] ctnc represents S i s knowledge of the value of ctnc at server S j . In addition, each server S i , maintains a log [19], L i , which contains the update history of the data items maintained in the system. The log is composed of a sequence of records hid; vn; vali, where id identifies the data item, vn identifies the version number of the data item and val is the value associated with the version of the data item. ....

G. T. Wuu and A. J. Bernstein. Efficient Solutions to the Replicated Log and Dictionary Problems. In Proceedings of the Third ACM Symposium on Principles of Distributed Computing, pages 233--242, August 1984.

....from both own history and out histories. The problem of choosing the time moment for the clipping to be performed is a well known problem is distributed systems that of establishing a consistent cut within the system s execution[16] The algorithm we adopt is the gossip message technique [17] which offers an asynchronous broadcast facility with a causal order of delivery, allowing the information to disseminate within the system at arbitrary times. In the gossip technique, beside functional messages, the processes also exchange gossip messages, containing information about events that ....

G. Wuu and A. Bernstein. Efficient solutions to the replicated log and dictionary problem. In Proceedings of the Third ACM Symposium on Principles of Distributed Computing, pages 233--242, 1984.

....directory maintenance problem as a replicated dictionary problem, with slightly (but significantly) different semantics. A timestamp vector was used to infer from a comparison of two dictionary replicas which entries had been inserted and which had been deleted. Allchin [1] and Wuu and Bernstein [12] expanded upon Fischer and Michael s approach to use two dimensional timestamp matrices to reduce the number of messages exchanged, with small variations in semantics. None of these works addressed the general problem of reclaiming resources of named replicated objects; they were concerned with ....

Gene T. J. Wuu and Arthur J. Bernstein. Efficient solutions to the replicated log and dictionary problems. In Proceedings of the Third Annual ACM Symposium on Principles of Distributed Computing, August 1984.

....generated. The resolution is done either automatically or manually depending on the semantics of the data type. The integration and resolution of data at other sites ensures that replicas eventually reach identical states or become mutually consistent if no new updates are generated in the system [WB84, HHW89]. This consistency approach, which guarantees eventual consistency of data, is called eventual data consistency. Eventual data consistency provides high availability by allowing updates to be made in any order. This consistency approach is attractive when it is essential to provide access to data, ....

....replicated copies [Tho79, Gif79] However, availability is still limited to a single partition when strong consistency is used for replica control. 8. 1 Typed Consistency Several researchers realized that accesses do not have to be serialized for maintaining data consistency for certain data types [FM82, All83, DS83, WB84]. The semantics of the data type can be used to define an alternative consistency cri86 terion. Much of this research was aimed at providing consistency for the directory data structure that allows the insert, delete and list operations. Many applications such as mail and calendar use a variation ....

[Article contains additional citation context not shown here]

Gene T. J. Wuu and Arthur J. Bernstein. "Efficient Solutions to the Replicated Log and Dictionary Problems." In Proceedings of the Third Annual ACM Symposium on Principles of Distributed Computing, August 1984.

....could limit the scalability of the implementations. Since version vectors only need to be sent with synchronization variables, we believe that this is not a problem for data race free programs. Furthermore, there exist techniques that can be used to limit the information carried in timestamps [41] by maintaining at each processor how far the clocks at other processors have advanced. The performance benefits due to better scalability of causal memory or message passing are not easily seen for the applications and their sizes that we used in the study. This is because in all of them, the ....

Gene T. J. Wuu and Arthur J. Bernstein. Efficient solutions to the replicated log and dictionary problems. In Proceedings of the 3rd ACM Symposium on Principles of Distributed Computing, 1984.

....weak consistency are internet news, air traffic control, air line reservation and stock exchanges. Grapevine [SBN84] and the Global Name Service [Lam86] were among the first systems to use weak consistency. Other weak consistency protocols were presented in [DGH 87, DGP90b, Gol92, LLS90, QP93, WB84] However, these protocols assumed that any node could communicate with any other node. This assumption, though appropriate for small networks with a few replicas, is unrealistic for wide area networks like the Internet. This paper presents a Hierarchical Asynchronous Replication Protocol (HARP) ....

....mongering for fast, unreliable propagation, while anti entropy provides a reliable backup if the other methods fail. However, anti entropy is an expensive procedure since it involves comparing the contents of two copies of the database, one of them sent over the network. Wuu and Bernstein in [WB84] present a propagation scheme for maintaining a replicated dictionary using logs. They use a two dimensional table (2DDT) of size On 2 similar to the union of AckM and V V which allows only missing updates to be exchanged rather than the whole copy of the database. Periodically, each node x ....

[Article contains additional citation context not shown here]

G. Wuu and A. Bernstein. Efficient Solutions to the Replicated Log and Dictionary Problems. In Proceedings of the third ACM Symposium on Principles of Distributed Computing, pages 233--242, August 1984.

....protocol for managing replicated data asynchronously. This method is especially suited for the bulletin board application mentioned above. Grapevine [17] and the Global Name Service [15] were among the first systems to use weak consistency. Other weak consistency protocols were presented in [7, 8, 12, 14, 16, 18]. These protocols are useful and interesting; however, they require a node to send messages to all other nodes. This would be acceptable for small networks with a few replicas, but it will incur a large communication overhead for wide area networks like the Internet. Moreover, a node sends not ....

G. Wuu and A. Bernstein. Efficient Solutions to the Replicated Log and Dictionary Problems. In Proceedings of the third ACM Symposium on Principles of Distributed Computing, pages 233--242, August 1984.

....can tolerate some inconsistency and reconciliation methods should be available to resolve conflicts. Typical applications that have used weak consistency are internet news, air traffic control, airline reservation and stock exchanges. Several weak consistency protocols were presented in [10, 11, 15, 19, 22, 23]. However, they require a node to communicate with all nodes which will incur a large communication overhead for wide area networks like the Internet. This paper evaluates the performance of HARP, a hierarchical replication protocol that scales well for thousands of replicas while ensuring ....

G. Wuu and A. Bernstein. Efficient Solutions to the Replicated Log and Dictionary Problems. In Proceedings of the third ACM Symposium on Principles of Distributed Computing, pages 233--242, August 1984.

....by this than performance of Lotus Notes (since reconciliation may run less frequently because most updates will be propagated by update notification mechanism) our approach would still be beneficial by improving performance of update propagation when it does run. In Wuu and Berntein s protocol [15], anti entropy is done by nodes exchanging gossip messages [9, 7] A gossip message from j to i contains log records of updates that j believes are missed by i, and version vector information describing the state of j as well as the extent of j s knowledge about the state of other nodes in the ....

....messages [9, 7] A gossip message from j to i contains log records of updates that j believes are missed by i, and version vector information describing the state of j as well as the extent of j s knowledge about the state of other nodes in the system. The Two phase Gossip protocol [7] improves [15] by sending fewer version vectors in a gossip message. It also describes a more general method for garbage collecting log records. Agrawal and Malpani s protocol [1] decouples sending update logs from sending version vector information. Thus, separate policies can be used to schedule both types of ....

G. T. Wuu and A. J. Bernstein. Efficient solution to the replicated log and dictionary problems. In Proc. of the 3d ACM Symp. on Principles of Distr. Computing, pp. 233-242, 1984. This article was processed using the L A T E X macro package with LLNCS style

.... stability require explicit acknowledgments [1, 2] or transitive acknowledgments [3, 4] An explicit acknowledgment is sent by a receiver in direct response to a message it receives; explicit acknowledgments may be sent separately for each sender or may be grouped to form a vector acknowledgment [5] for all of the senders. A transitive acknowledgment, included in a newly generated message, provides information about the receipt of messages that causally precede [6] the message. The major performance considerations for determining the stability of a message are: 1) the bandwidth consumed ....

....2 N sends a message m, then eventually all nodes in N will determine that m is stable. Proof. The proof is a simple extension of the proof of Theorem 1. 2 4 Timestamp vs. Vector Acknowledgments We now compare our timestamp acknowledgment scheme with the traditional vector acknowledgment scheme [5, 10] based on four different criteria: bandwidth used, cycles used, memory used, and latency to the determination of message stability. Because there is a minimum rate at which each sender sends messages and because each receiver sends acknowledgments at a constant rate, the number of messages ....

G. T. J. Wuu and A. J. Bernstein, "Efficient solutions to the replicated log and dictionary problems," Operating Systems Review, vol. 20, no. 1 (January 1986), pp. 57-66.

....each event corresponds to an update. Each update record, e contains the fields e:op (the operation and its parameters) e:p (the site that first executed the operation) and e:V TS (a vector timestamp attached to the event) Sites distribute information by exchanging their logs. Several authors [8, 15, 1, 10] require that updates satisfy the consistent log property (similar to requiring that multicasts be causally delivered) Given a log L, the first i events of L are denoted by L[i] If event e is in L, then its position is denoted by index(i) We will use L[e] as a shorthand notation for ....

G.T.J. Wuu and A.J. Bernstein. Efficient solutions to the replicated log and dictionary problems. In Proc. 3rd. Symp. on Principles of Distributed Computing, pages 233--242, 1984.

....a node holding the lock and another one requesting it, and (2) ensuring that applications have sufficient synchronization so they are data race free is complex. We explore causal consistency (CC) because it appears that it has two main advantages: 1) CC is meaningful for several applications [2, 9, 16, 30], and (2) protocols that implement CC require only weak synchronization that permits the design of efficient implementations. Thus, in this paper we explore the design of a protocol that can be used to provide causally consistent distributed services. There exist protocols that implement causal ....

....related distributed services. This Section presents a protocol implementing this consistency criterion in this architecture. 4. 1 Tracking Causality The most fundamental data structure to track causality relations and implement causally consistent objects is a two dimensional matrix of integers [9, 30, 5, 23]. Although a version vector is sufficient for tracking causality in the full replication case that was discussed previously, partial replication requires that updates to objects be recorded separately. For example, consider server node s i with copies of objects x a and x b , and server node s j ....

G.T. Wuu and A.J. Berntsein. Efficient Solutions to the Replicated Log and Dictionary Problems. In Proc. 3rd ACM Symposium on Principles of Distributed Computing, Montreal, 1984, pp. 233-242.

....into a logical tree structure. Each of these subtransactions commits locally, with the provision that if the root of the tree commits, all its children commit too. In this approach, persistent transmission of messages is required, for instance, by mean of stable queues [7] or gossip mechanisms [20, 35]. Another approach that avoids two phase commit is given in [27] transactions commit at a primary site and are propagated asynchronously. A partial order among transactions is defined for controlling update propagation; the use of a partial order is also proposed in [25] and in [15] in the ....

G.T.J. Wuu and A. Bernstein, "Efficient solutions to the replicated log and dictionary problems," in Proc. 3rd ACM Symposium on Principles of Distributed Computing, 1984.

....weak consistency are Internet news, air traffic control, airline reservation and stock exchanges. Grapevine [SBN84] and the Global Name Service [Lam86] were among the first systems to use weak consistency. Other weak consistency protocols were presented in [DGH 87, DGP90, Gol92, LLS90, QP93, WB84] These protocols are useful and interesting; however, they require a node to communicate with all other nodes. This would be acceptable for small networks with a few replicas, but it will incur a large communication overhead for wide area networks like the Internet. This paper evaluates the ....

G. Wuu and A. Bernstein. Efficient Solutions to the Replicated Log and Dictionary Problems. In Proceedings of the third ACM Symposium on Principles of Distributed Computing, pages 233--242, August 1984.

....words: distributed systems, causality, logical time, matrix time, fault tolerance. 1 Introduction Matrix clocks have been introduced in the context of asynchronous distributed systems. They have nice properties that can be used to design distributed database protocols and fault tolerant protocols [WB84, KB91]. Unfortunately, they are costly to implement: in a distributed system that consists of n sites, the naive algorithm to compute the matrix clock on the fly requires that an n Theta n matrix of integers be stored at each site and tagged onto each message: if the number of sites is large, it is ....

....in a distributed system that consists of n sites, the naive algorithm to compute the matrix clock on the fly requires that an n Theta n matrix of integers be stored at each site and tagged onto each message: if the number of sites is large, it is necessary to use an optimized algorithm. [WB84] describes several such optimized algorithms. Usually, optimized algorithms do not actually compute the genuine matrix clock, but an approximation thereto. The approximation is less expensive to compute but provides less information than the genuine clock. However, many applications will be happy ....

[Article contains additional citation context not shown here]

G. T. J. Wuu and A. J. Bernstein. Efficient solutions to the replicated log and dictionary problems. In Proc. 3rd ACM Symp. on PODC, pages 232--242, 1984. Public/Draft-16- July 22, Chorus syst`emes Cheaper matrix clocks CS/TR-94-63

No context found.

G.T.J. Wuu and A. Bernstein, "Efficient Solutions to the Replicated Log and Dictionary Problems," in Proc. 3rd ACM Symposium on Principles of Distributed Computing, 1984.

No context found.

Wuu, G. T. J. and Bernstein, A. J. (1984) Efficient solutions to the replicated log and dictionary problems. In Proc. 3rd ACM Ann. Symp. on Principles of Distributed Computing, pp. 233-- 242.

No context found.

G. T. J. Wuu and A. J. Bernstein. Efficient solutions to the replicated log and dictionary problems. In Proceedings of the ACM Symposium on Principles of Distributed Computing (PODC'84), pages 232--242. ACM, 1984. 10 A Proof of Lemma 2 The hypothesis of lemma 2 concern two stamps (say and ) in which we can identify some sort of conflict between

No context found.

G.T.J. Wuu and A.J. Bernstein. Efficient solutions to the replicated log and dictionary problems. In 3nd ACM Symposium on PODC, volume 99, pages 233-242, 1984.

No context found.

Wuu and Bernstein. 1984. Efficient Solutions to the Replicated Log and Dictionary Problems. Proceedings of the Third ACM Symp. on Princ. Distrib. Computing. Vancouver, B.C. August 27-29, 233-242.

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