John B. Carter, John K. Bennett, and Willy Zwaenepoel. Techniques for Reducing Consistency-Related Communication in Distributed Shared-Memory Systems. ACM Transactions on Computer Systems, 13(3):205--243, August 1995.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....implementations of distribution policies may be dynamically loaded into the system without the need for a shutdown and subsequent restart. Differentiating policies is not new, but has mainly been a subject of interest in distributed shared memory systems (see, e.g. the work on Orca [2] Munin [5], TreadMarks [1] and more recently, InterWeave [6] However, only recently, some work has been done in relation to shared data spaces [4] but is still directed towards computation intensive applications and without the flexibility of runtime adaptations. The paper is structured as follows. ....

J. Carter, J. Bennett, and W. Zwaenepoel. Techniques for Reducing Consistency-Related Communication in Distributed Shared Memory Systems. ACM Trans. Comp. Syst., 13(3):205--244, Aug. 1995.

....memory consumption and can eliminate diff accumulation. The home in a home based protocol can be either fixed [19] or mobile [21] There is also variation for the coherence operations, such as a multiple writer protocol, or a single writer protocol. The multiple writer protocol introduced in Munin [17] supports concurrent writes on di#erent copies using the di# technique. It may however incur heavy di# overhead compared with conventional single writer protocols. Another choice is between the update protocol (e.g. Orca [15] and the invalidate protocol used in many pagebased DSM systems such as ....

....Java memory model. We believe that adaptive protocols 3 are superior to non adaptive ones due to their adaptability to object access patterns in applications. An adaptive cache coherence protocol is able to detect the current access pattern and adjusts itself accordingly. Some DSMs (e.g. Munin [17]) support multiple cache coherence protocols and allow the programmer to explicitly associate a specific protocol with the shared data. This is not transparent to the programmer and it is di#cult to dynamically switch between di#erent protocols in response to changes in the access pattern. Several ....

[Article contains additional citation context not shown here]

J. B. Carter, J. K. Bennett, W. Zwaenepoel, Techniques for Reducing Consistency-Related Communication in Distributed Shared-Memory Systems, ACM Transactions on Computer Systems 13 (3) (1995) 205--243.

....and eliminates di# accumulation. The home in a home based protocol can be either fixed [18] or migratable [9] There are also choices for the coherence operations, such as that between a multiple writer protocol or a single writer protocol. The multiple writer protocol introduced in Munin [8] supports concurrent writes on di#erent copies using the di# techniques. It may however incur heavy di# overhead compared with conventional single writer protocol. Another choice is between the update protocol (e.g. Orca [5] and the invalidate protocol in many page based DSM systems (e.g. ....

....memory access patterns in a application determine which protocol is more e#cient and therefore suitable. That motivated us to go after an adaptive protocol. An adaptive cache coherence protocol is able to detect the current access pattern and adjusts itself accordingly. Some DSMs (e.g. Munin [8]) support multiple cache coherence protocols and allow the programmer to explicitly associate a specific protocol with the shared data. This is not transparent to the programmer and cannot dynamically switch between di#erent protocols in response to changes in access pattern. Several page based ....

[Article contains additional citation context not shown here]

John B. Carter, John K. Bennett, and Willy Zwaenepoel. Techniques for Reducing Consistency-Related Communication in Distributed Shared-Memory Systems. ACM Transactions on Computer Systems, 13(3):205--243, 1995.

....diffs, quickly applies diffs, and minimizes the space used to store and transmit diffs. We show that the algorithm exhibits significantly reduced memory consumption and network costs when compared to existing differencing mechanisms. The basic method employed by all existing diff mechanisms, [5, 13, 18, 19] can be summarized as follows. Initially write access is disabled for the shared data region (we will refer to them as shared blocks) When a process attempts to modify the block, a write access fault occurs and invokes an interrupt handler. The handler creates an immutable copy of the block, ....

J. B. Carter, J. K. Bennett, and W. Zwaenepoel. Techniques for reducing consistency-related communication in distributed shared memory systems. In TOCS, 1995.

....[1, 6] It provides exibility to cache coherence mechanisms on when to make writes to be globally visible. The traditional eager coherence protocol enforces the cache coherence on every memory write [10] The lazy coherence protocol, on the other hand, permits temporary inconsistent cache copies [3, 8, 9, 2, 7]. Those copies of data in di erent caches need to be reconciled only after the execution of a synchroniza tion instruction. Although this lazy coherence protocol eliminates unnecessary cache coherence activities, it incurs signi cant overhead on posting writes and reconciling cache lines at each ....

J. Carter, J. Bennett, and W. Zwaenepoel, \Techniques for Reducing ConsistencyRelated Communication in Distributed Shared-Memory Systems," ACM Trans. on Computer Systems, Vol 13(3), Aug. 1995, pp. 205-243.

....the operation completes. The only dioeerence is that the relation ; is augmented with synchronization operations. 5. 1 Locks The same ownership mechanism used to totally order write operations over a page can be used to obtain a distributed implementation of locks, similar to the one in Munin [10]. For each lock l, there is always exactly one owner, which is the manager that last acquired the lock on behalf of its local process. Each manager keeps a hint of l s probable owner, to whom requests to acquire the lock are sent. Requests received by P are forwarded to the probable owner, except ....

....centralized solutions, betting on locality of reference: the fewer the number of processes that contend for the same lock, the less the communication costs. An extreme but not unusual situation occurs when the last process that acquired the lock reacquires it. In such a case, no messages are sent [10]. 5.2 Barriers A barrier establishes a synchronization point for a set of processes. Informally, no process in the set can go beyond the barrier until it is aware of what every other process has done before getting to the barrier. Therefore, after a barrier, every manager must end up with a ....

[Article contains additional citation context not shown here]

J. Carter, J. Bennett, and W. Zwaenepoel. Techniques for reducing consistency-related communication in distributed shared memory. ACM Trans. Comput. Syst., 13(3):

....information. As a result, LRC based SDSM generally involves fewer messages and less data exchanged than other SDSM implementations. As SDSM relies on virtual memory traps, its memory coherence is maintained at the OS page granularity. The multiple writer (MW) protocol and lazy diff creation [6] are designed to alleviate false sharing and to reduce the data transfer amount of SDSM when fine grain application programs are run. MW allows multiple processing nodes to modify different portions of the same shared memory page at the same time, whereas lazy diff creation permits each processor ....

....software systems, and application benchmarks. 4. Experimental Setup Coordinated checkpointing techniques are built on top of TreadMarks [2] for our experiments. TreadMarks implements the LRC protocol [14] with several optimization techniques, including multiple writer and lazy diff creation [6], to lower the amount of data movement and interprocess communication. It adopts the UDP IP protocol for message exchange, uses the SIGIO signal for delivering request messages, and relies on a virtual memory trap (SIGSEGV) for invoking the memory coherence enforcement mechanism. All checkpoints ....

J. B. Carter, J. K. Bennett, and W. Zwaenepoel. Techniques for Reducing Consistency-Related Communication in Distributed Shared Memory Systems. ACM Trans. on Computer Systems, 13(3):205--243, August 1995.

.... PCR uses the logged data to identify whether the accesses to the shared memory page involve any write operation: if not, the page status is set to read only; otherwise, a twin, pristine copy of that page is created and the page status is set to read write, satisfying multiple writer protocol [6] requirements. As a result, PCR eliminates both page faults and idle time resulting from shared memory access misses during the recovery process. Figure 2 shows the differences of the coherence enforcement actions involved during failure free execution and the recovery processes under SCR and PCR ....

J. B. Carter, J. K. Bennett, and W. Zwaenepoel. Techniques for Reducing Consistency-Related Communication in Distributed Shared Memory Systems. ACM Trans. on Computer Systems, 13(3):205--243, August 1995.

....of Li and Hudak [26] a large body of knowledge has been created focusing, primarily, on improving the performance of such systems by reducing the number and size of messages sent. This has been accomplished using weak consistency protocols (e.g. 14, 24, 7] as well as other optimizations (e.g. [15, 2, 3, 27, 30, 8]) Work has also been done on decreasing the overhead of the transmission protocol used to send the required consistency maintenance messages [34, 33, 10] 2.3. Active Networks Active networks[13, 29, 32, 36] allow customized (possibly application specific) programs to execute in the network. By ....

J. B. Carter, J. K. Bennett, and W. Zwaenepoel. Techniques for Reducing Consistency-Related Communication in Distributed Shared Memory Systems. ACM Trans. on Computer Systems, 13(3):205--243, Aug. 1995.

....pioneering work of Li and Hudak [15] a large body of knowledge has been created focusing, primarily, on improving the performance of such systems by reducing the number and size of messages sent. This has been accomplished using weak consistency protocols [7, 13, 4] as well as other optimizations [8, 1, 2, 21]. Work has also been done on decreasing the overhead of the transmission protocols used to send consistency maintenance messages [23, 22, 6] and on the underlying networks themselves (e.g. Gigabit Ethernet, Myrinet [5] 2.2. LOTEC Sui and Graham [20] described a DSM based object programming ....

J. B. Carter, J. K. Bennett, and W. Zwaenepoel. Techniques for Reducing Consistency-Related Communication in Distributed Shared Memory Systems. ACM Trans. on Computer Systems, 13(3), 1995.

....of Thisresearchwas supportedinpartby the Natural Sciencesand Engineering ResearchCouncilofCanadaunder grantOGP 0194227. such systemsby reducing the numberandsizeofmessages sent. This has been accomplished using weak consistency protocols [CBZ91, KCZ92, BZS93] as well as other optimizations [CBZ95, ACDZ97, ACRZ97, Lu97, SB98, BPA98] Work has also been done on decreasing the overhead of the transmission protocol used to send the required consistency maintenance messages [vECGS92, vEBBV95, Buo98] and on the underlying networksthemselves (e.g. Gigabit Ethernet, Myrinet [BCF 95] Despite ....

J. B. Carter, J. K. Bennett, and W. Zwaenepoel. Techniques for Reducing Consistency-Related Communication in Distributed Shared Memory Systems. ACM Trans. on Computer Systems, 13(3):205--243, August 1995.

....by the computation and the amount of messages 2 required. The interest in the potential of software DSM based systems is evidenced by the vigorous program of research and by several prototypes that have been developed in recent years, including Midway, Munin, TreadMarks, and Quarks [BZS93, CBZ95, KCDZ94, CKK95] DSM systems, however, have not yet reached the maturity of the message passing systems. In fact, researchers have identified bottlenecks in the performance of software DSM systems and have proposed techniques for removing them [CBZ91, CK96, KCDZ94, LDCZ97] In this chapter we ....

....written value of a write operation to be visible to other processes until the writer executes a release operation, it simulates another type of mechanism, called multiple writer protocol, which allows multiple writers write to the same page simultaneously. Write shared Data Protocol Munin [Car93, CBZ95, CBZ91, Car95] proposed a write shared data protocol to implement the conventional release consistency. The write shared data protocol uses the page based mechanism. However, it allows more than one process to write to the same page simultaneously. Inconsistent copies of a page may exist in the ....

[Article contains additional citation context not shown here]

J. B. Carter, J. K. Bennett, and W. Zwaenepoel. Techniques for reducing consistency-related communication in distributed shared memory systems. ACM Transactions on Computer Systems, 13(3):205--243, August 1995.

....computer scientist would tell you, That s just how memory works. Then two things happened. The first is that memory systems in multiprocessors got more and more complicated [17, 18, 22, 36] The second is the 2 invention of distributed shared memory (DSM) for the message passing multicomputer [6, 9, 10, 12, 13, 37, 38]. DSM provides the illusion of a shared address space on top of hardware that only supports message passing. Caching and out of order instruction dispatching can pose a problem for multiprocessors. The hardware of each processor enforces the restriction that the processor sees its own memory ....

J. K. Bennett, J. K. Carter, and W. Zwaenepoel. Techniques for reducing consistencyrelated communication in distributed shared memory systems. ACM Transactions on Computer Systems, 13(3):205--243, August 1995.

....synchronisation are provided in most modern architectures. The answer is that in a distributed system these instructions are unusable for synchronisation unless strict memory coherence is enforced. Strict coherence has been shown to be expensive, and unnecessary for many distributed applications [20]. In order to maintain sufficient flexibility until we have gathered enough experience with distribution in Mungi, we decided to support synchronisation in the OS. In a later version this can be moved to a library should that turn out to be sufficient. The remainder of this section describes in ....

J. B. Carter, J. K. Bennett, and W. Zwaenepoel. Techniques for reducing consistency-related communication in distributed shared memory systems. ACM Transactions on Computer Systems, 13:205--243, 1995.

....consistency model. The sequential model is not suitable for a COW due its large overhead. For this reason, release consistency models tolerate some degree of inconsistency among the copies of shared data because the updates are notified only upon a process synchronization. The eager release model [3] propagates the updates when a process either releases a critical section or it synchronizes with other processes. This paper presents Hive, a Java library implementing a distributed shared memory consisting of a set of areas. An area is a sequence of consecutive positions whose size is chosen by ....

Bennett J.K., Carter J.B., Zwaenepoel W., "Techniques for Reducing Consistency-Related Communication in Distributed Shared-Memory Systems ", ACM TOCS, Vol.1, No. 3, Aug. 1995.

....directives allowed indicate data parallelism. In certain Distributed Shared Memory (DSM) systems concurrent task ID s are available, either through intrinsic function calls, or through parameters passed by the user when spawning a task. Examples are Treadmarks [2] Cilk [52] Cashmere [46] Munin [12], Ivy [37] and Shasta [42] The recently announced OPEN MP standard, endorsed by a number of vendors, tries to unify existing DSM application programmer interfaces. The main difference between the DSMs is in the type of memory consistency offered, and in the support for hardware shared memory ....

J.B. Carter, J.K. Bennett, W. Zwaenepoel, "Techniques for reducing consistency-related communication in distributed shared memory systems," ACM Trans. Computer Syst., Vol. 13, pp. 205--243, August 1995 26

....modify their caches simultaneously, the implementation must choose between the values written. Various solutions to these two problems differentiate distributed shared memory systems: Kai Li s pioneering system used a single writer per page policy [15] Munin introduced weak consistency models [16, 17], and recent research considers other models and implementation techniques [18, 19, 20, 21, 22] In our work, we have investigated three aspects: architectures for building protocols [23] correctness of protocols [24, 25] and distributed memory models for high latency networks [26] Here we ....

John B. Carter, John K. Bennet, and Willy Zwaenepoel. Techniques for reducing consistency-related communication in distributed shared memory systems. ACM Transactions on Computer Systems, 1995. To appear.

No context found.

J. Carter, J. Bennett, and W. Zwaenepoel. Techniques for reducing consistency-related communication in distributed shared memory systems. ACM Transactions on Computer Systems, 13(3):205--243, Aug. 1995.

No context found.

J.B. Carter, J.K. Bennett, and W. Zwaenepoel. Techniques for reducing consistency-related communication in distributed shared memory systems. ACM Transactions on Computer August 1995.

....scalable performance regardless of the speed of the individual processing elements and provides no mechanism for tuning by the compiler or run time system. These observations have led a number of researchers to propose building cache controllers that can execute a variety of caching protocols [6, 21], support multiple communication models [8, 11] or accept guidance from software [12, 17] We are investigating cache designs that will implement a variety of caching protocols, support both shared memory and message passing efficiently, accept guidance from software to tune its behavior, and ....

J.B. Carter, J.K. Bennett, and W. Zwaenepoel. Techniques for reducing consistency-related communication in distributed shared memory systems. ACM Transactions on Computer Systems, 13(3):205--243, August 1995.

....given the illusion of a large global address space encompassing all available memory, eliminating the task of explicitly moving data between processes located on separate machines. Both hardware DSM systems (e.g. Alewife [15] DASH [36] FLASH [31] and software DSM systems (e.g. Ivy [37] Munin [13], TreadMarks [28] have been implemented. The majority of software DSM systems use page based memory protection hard ware and the low level message passing facilities of the host operating system to implement the necessary shared memory abstractions. Programmers write programs using the ....

....is allowed to proceed. This restriction severely limits the performance achievable by these systems because of the large amount of network traffic necessary to maintain coherence on each write to shared data. In order to relieve the network pressure caused by sequential consistency, the Munin [13] system used a multiple writers protocol. This protocol allows more than one process to modify a portion of a shared page of data at the sanhe tinhe, as long as the programmer ensures that a single word is not write shared without an intervening synchronization point. By buffering writes and only ....

[Article contains additional citation context not shown here]

J. Carter, J. Bennett, and W. Zwaenepoel. Techniques for reducing consistency- related communication in distributed shared memory systems. T'asactios o Computer S!lstems, 13(3):205 243, August 1995.

No context found.

John B. Carter, John K. Bennett, and Willy Zwaenepoel. Techniques for Reducing Consistency-Related Communication in Distributed Shared-Memory Systems. ACM Transactions on Computer Systems, 13(3):205--243, August 1995.

No context found.

J. B. Carter, J. K. Bennett, and W. Zwaenepoel. Techniques for reducing consistency-related communication in distributed shared-memory systems. ACM Transactions on Computer Systems, 13(3):205--243, 1995.

No context found.

J. Carter, J. Bennett, and W. Zwaenepoel. "Techniques for Reducing Consistency-Related Communication in Distributed Shared Memory Systems." ACM Trans. Comp. Syst., 13(3):205--244, Aug. 1995.

No context found.

J. B. Carter, J. K. Bennett, and W. Zwaenepoel. Techniques for Reducing Consistency-Related Communication in Distributed Shared-Memory Systems. ACM Transactions on Computer Systems, 13(3):205--243, 1995.

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