M. Stumm and S. Zhou. Algorithms Implementing Distributed Shared Memory. IEEE Computer, 23(5):54--64, May 1990.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....a functionality that invokes userdefi. r actions on eachmachine that shares the data. Although in the example above other metaobjects arenotifi. of a method call immediately, some mechanisms for distributed shared data improve performance by using an algorithm in which thenotifi. is delayed [24]. Such an algorithm is also available in Open C if the programmer defi.og a subclass of OcMetaObj to implementit. When implementing such an algorithm, it is necessary to distinguish methods that modify the shared data from methods that simply read the data. Category names of reflect methods are ....

Stumm, M. and S. Zhou, "Algorithms Implementing Distributed Shared Memory," IEEE Computer, vol. 23, no. 5, pp. 54--64, 1990. 18

....receive the data. The final picture is that of a difficult system to program. In part to address the difficulty of programming parallel distributed systems, Shared Memory Models are being considered as an alternative to the traditional communication model in distributed architectural platforms [3, 6, 7, 9]. These models provide a set of shared addresses that the processes can use in the same way they access data locally stored, i.e. using the primitives data=read(address) and write(address,data) In general, memory models give a simpler view than the communication model in applications where the ....

Stumm, S., Zhou, S. Algorithms Implementing Distributed Shared Memory. IEEE Computer, 23(5):54-64, May 1989. 6

....receive the data. The final picture is that of a difficult system to program. In part to address the difficulty of programming parallel distributed systems, Shared Memory Models are being considered as an alternative to the traditional communication model in distributed architectural platforms [6, 8, 12, 13, 15, 17]. These models provide a set of shared addresses that the processes can use in the same way they access data locally stored, i.e. using the primitives data=read(address) and write(address,data) In general, memory models give a simpler view than the communication model in applications where the ....

Stumm, S., Zhou, S. Algorithms Implementing Distributed Shared Memory. IEEE Computer, 23(5):54-64, May 1989.

....memories, however, are accessed asynchronously, implying that they can be read or written any time an application programmer wants to. There is no need to do any special synchronization activity before or after the access. DIPC s shared memory uses the single writer multiple readers protocol [7], and employs the strict consistency model [4] meaning that a read will return the most recently written values. Read or write access rights are assigned to computers, and all the processes on that machine can enjoy the rights. DIPC can manage shared memories in two modes: In the segment mode, ....

Michael Stumm and Songnian Zhou, "Algorithms Implementing Distributed Shared Memory", Computer, May 1990. 5

....the write operation changes the state of DSM. This is the job of coherence protocol and the way of doing it depends on the consistency model assumed for the DSM. Besides the essential problems mentioned above, the problem of false sharing and thrashing is frequently pointed out ( 16] 18] [21]] False sharing occurs while two or more processes on distinct hosts try to access distinct locations on the same page. When one process gets an exclusive access to a location (e.g. for writing) other processes are not allowed to access locations on the same page, although they attempt to ....

Stumm M., and Zhou S.: Algorithms Implementing Distributed Shared Memory, IEEE Computer, vol. 24, no. 5, May 1990, Pages 54--64.

....we are aware of the limits and the drawbacks of the shared memory communication scheme. The major disadvantage of such a mechanism is the problem of ensuring memory consistency [5] 8] and its costs. 2. 1 Various algorithms Distributed shared memory has been well studied for many years [7], 5] Many basic algorithms are provided for implementing DSVM. These algorithms may be compared to each other in terms of communication overheads and cost, data migration and replication (See [7] for more details) 1. The most natural algorithm is based on a centralized client server approach. ....

....its costs. 2.1 Various algorithms Distributed shared memory has been well studied for many years [7] 5] Many basic algorithms are provided for implementing DSVM. These algorithms may be compared to each other in terms of communication overheads and cost, data migration and replication (See [7] for more details) 1. The most natural algorithm is based on a centralized client server approach. A centralized algorithm associates one or more pages to each server which updates these pages. Write and read operations are executed through the server by sending requests and receiving ....

M. Stumm and S. Zhou. Algorithms implementing distributed shared memory. IEEE Computer, pages 54--63, May 1990.

....done by the programmer, but at least partial control from the compiler as well as automatic policy selection by Munin itself is envisioned. The idea of having different policies for different memory objects extends to the NUMA multiprocessor domain as well. A recent paper by Stumm and Zhou [SZ90] considers several different strategies for managing a DSM system. They argue that the best strategy is applicationdependent, and advocate providing multiple mechanisms from which an application programmer can choose. We believe that this idea also extends to the NUMA memory management ....

....results suggest the utility of allowing users to tune memory management policy for their applications. The notion of tuning policies to applications is not a new one. Stumm and Zhou argue for application dependent selection of memory management policy for distributed shared memory systems in [SZ90] and Bennett, Carter, and Zwaenepoel include memory object level selection of policy in their object oriented distributed shared memory system called Munin [BCZ90a, BCZ90b] The design of DUnX version two encourages consideration of this idea of tuning policies to applications by providing ....

M. Stumm and S. Zhou. Algorithms implementing distributed shared memory. IEEE Computer, 23(5):54--64, May 1990.

....shared memory system Munin [2] as in the Dash system, release consistency [14] is used to do this. Munin additionally provides different types of shared variables with strict or weak coherency protocols. An overview of virtual shared memory algorithms and systems can be found in [26] 30] [32], and [33] In Pact, virtual shared memory management is done with a fixed distributed manager algorithm with write invalidation similar to Li s algorithm in [20] This management is done in the runtime system instead of the operating system. This has the advantage that Pact can be implemented ....

Michael Stumm and Songnian Zhou. Algorithms Implementing Distributed Shared Memory. IEEE Computer, 23(5):54--64, May 1990.

....1995] is similar to the SR model, with the same disadvantage of requiring the programmer to explicitly bind memory regions to a region identifier. 2.2. 4 Distributed Shared Memory: TreadMarks A variety of distributed shared memory (DSM) systems have been designed, implemented, and evaluated [Stumm and Zhou, 1990] . Ivy [Li, 1988] was the first DSM system. Since then, other systems, including Midway [Bershad et al. 1993] and Blizzard [Schoinas et al. 1994] have also been described in the literature. The TreadMarks system [Keleher et al. 1994, CHAPTER 2. BACKGROUND AND RELATED WORK 15 Node 0 Node 2 ....

M. Stumm and S. Zhou. Algorithms Implementing Distributed Shared Memory. IEEE Computer, 23(5):54--64, May 1990.

....Some people point out similarities between Planet s basic functions and those of other operatingsystem or middleware level memory management systems. These include distributed shared memory (DSM) and single address space (SAS) systems. Planet s mobile memory segment is related to DSM [29, 22] and SAS systems like Opal [9] in that memory areas for application programs are transferred between computer sites in an automatic, transparent way. The difference is in their address sharing methods. DSM and SAS provide programmers with a virtual view wherein single address space may exist. ....

M. Stumm and S. Zhou. Algorithms implementing distributed shared memory. IEEE Computer, 23(5):54--64, May 1988.

....memory. The best known example is the Shared Virtual Memory of Li [20] The unit of replication is a page; several copies of the same page may be stored on different processors. Stumm and Zhou also provides a good comparison of existing implementation techniques of distributed shared memories [27]. This paper concerns another model, namely, the shared object model. It is a refinement of the shared memorywhere the unit of replication is an object. Several experimental distributed systemsbased on objects exist: Eden [2] Emerald [9] Clouds [19] Amoeba Orca [24] Guide [5] They define ....

....failure. 3 Queueing model and assumptions In order to compare the replication strategies wemust define someperformance metrics and model each strategy. In the context of distributed shared memory algorithms, Stumm and Zhou used a simple model for performance comparisons of different algorithms [27]. The average cost per data access to the entire system has been chosen as the performance measure. We would have used a similar model for our comparison, however, it suffers from several drawbacks. In particular, the cost of multicast communication is exaggerated. In this model, communication ....

[Article contains additional citation context not shown here]

M. Stumm, S. Zhou, "Algorithms Implementing Distributed Shared Memory ", Computer, vol. 23, num. 5, pp. 54-64, May 1990.

....It covers, in more detail, the three categories Tanenbaum discusses in his book [Tanenbaum 95] and the implementations given as examples in his book. 3.1 Basic Schemes for Implementing DSM There are four basic approaches for the implementation of DSM. Stumm and Zhou describe them as [Stumm et al. 1990] follows: Central Server; Migration; Read Replication; and Full Replication Schemes. 3.1.1 Central Server Scheme This is the simplest scheme for implementing DSM, depicted in Figure 2. The central server maintains the only copy of the data and controls all accesses to the ....

M. Stumm S. Zhou, Algorithms Implementing Distributed Shared Memory, IEEE Computer, Vol. 23, No 5, May 1990, pp 54-64.

....shared address space. However, we chose to implement the multiple readers, single writer (MRSW) replication policy in order to provide a consistency model that allows for some caching without an enormous amount of complication. We based this idea on an algorithm provided by Stumm and Zhou [SZ90]. Details about our implementation of the MRSW model can be found in section 3. We also needed to choose a policy for region replacement when there is not enough shared memory to hold all regions resident in memory at once. We chose a simple and effective LRU replacement strategy. Information ....

Michael Stumm and Songnian Zhou. Algorithms Implementing Distributed Shared Memory. From IEEE Computer, pp. 54-64, 1990.

....memory, the algorithms using them will vary in complexity. This is especially true when considering multi process shared memory algorithms. Shared memory models are being considered as an alternative to the traditional message passing model, specially for distributed architectural platforms [7, 8, 15, 13, 18, 20]. Most proposed implementations of the shared memory model require reads to return the value of the most recent write to the location being read [4, 6, 10, 21] However there is a trade off between guarantees provided by a memory type and the attainable performance of the overall algorithm ....

Stumm, S., Zhou, S. Algorithms Implementing Distributed Shared Memory. IEEE Computer, 23(5):54-64, May 1989.

....pages. The computation performed by each process is assumed to be piece wise deterministic; that is, the computational states generated by a process is fully determined by a sequence of data pages provided for a sequence of read operations. For the DSM model, we assume the read replication model [21], in which the system maintains a single writable copy or multiple read only copies for each data page. The memory consistency model we assume is the sequential consistency model, in which the version of a data page a process reads should X Reader Page X Copy Of X X . Request (2) 2) 2) X ....

M. Stumm and S. Zhou. Algorithms implementing distributed shared memory. IEEE Computer, pp. 54--64, May 1990.

....detect write operations to pages that are also shared by other processors, and invalidate or update other copies of the page. Faults to pages currently residing at other processors cause those pages to be copied to the faulting processor. Examples of such systems include Ivy [53] Mermaid [74], Amber [17] Clouds [62] Munin [12] Midway [9] and Treadmarks [44] In Ivy and Mermaid, copies of a page are invalidated as soon as a write to a page is detected by one processor. This makes these systems susceptible to the problem of false sharing at page granularity. The other sys Chapter ....

M. Stumm and S. Zhou. Algorithms implementing distributed shared memory. In IEEE Computer, Vol. 23, No. 5, May 1990.

....be divided into either a replication or a non replication algorithm. In the following paragraphs, we discuss the read and write operations. When a client, needs info about a specific variable, it sends a request to a specific PE; the server, where the data resides, feeds the data back to the client[63]. Central Server Algorithm The simplest strategy for implementing distributed shared memory is to use a central server, which is responsible for serving all accesses to shared data and maintaining the copy of the shared data. Both the read and write operations involve sending a request message to ....

STUMM, M., AND ZHOU, S. Algorithms Implementing Distributed Shared Memory. IEEE Computer Magazine (May 1990), 54--64.

....the DSM mechanism is implemented, the algorithms for implementing DSM, and the memory consistency model of the shared data. At the implementation level there are three main approaches: hardware, software and hybrid. Several basic algorithms for implementing DSM are described in the literature [SZ90], but the most widely used ones employ replication techniques to enhance the performance of language and system implementations. The use of replication introduces the inconsistency problem, which occurs when a process writes a replicated shared data. So, it is necessary to define a memory ....

Stumm, M., Zhou, S.: "Algorithms Implementing Distributed Shared Memory". Computer, Vol. 23, No 5. May 1990.

....platforms. These applications are good candidates for porting to an SCI based platform. Since shared memory model makes programming and debugging easier, and software more portable, efforts have been made to provide shared memory on such a system via software supported virtual memory management [25, 29, 21, 9]. To achieve reasonable performance, shared data must be replicated in various forms such as replicated data in the size of a page or some small units. Data coherency must be dealt with which requires a significant amount of housekeeping on the operating system part. Those systems struggle to ....

M. Stumm and S. Zhou. Algorithms implementing distributed shared memory. Computer, 23(5):54-- 64, May 1990.

....The interface definition fully specifies the interaction between the client and the server by defining which procedures will be remote, the data types of all variables passed and whether the variables are input values or output values. 3.1. 3 Distributed Shared Memory Distributed shared memory [82, 102] is an abstraction for supporting the concept of real shared memory in a network environment without physically shared memory. In contrast to message passing and remote procedure calls, in a distributed shared memory system a process which stores a value does not need to know the existence or ....

M. Stumm and S. Zhou. Algorithms Implementing Distributed Shared Memory. IEEE Computer, 23(5):54--64, 1990.

....at the level of shared segments mapped in the virtual address spaces of different objects. Other systems, such as Munin [Bennett et al. 1990] and Orca [Bal et al. 1992] have supported consistency at the level of structured objects. A classification of consistency protocols for DSM is presented by Stumm and Zhou [1990]. These are classified along two dimensions, namely migrating non migrating, and replicated non replicated data. Most commonly used protocols are based on read replication, which supports the single writer multiple reader based synchronization. A shared item can be migrated to or replicated in the ....

Michael Stumm and Songnian Zhou. Algorithms implementing distributed shared memory. IEEE Computer, pages 55--64, May 1990.

No context found.

M. Stumm and S. Zhou, `Algorithms implementing distributed shared memory', IEEE Computer, 23(5), 54--64 (1990).

No context found.

M. Stumm and S. Zhou. Algorithms Implementing Distributed Shared Memory. IEEE Computer, 23(5):54--64, May 1990.

No context found.

M. Stumm and Songnian Zhou. Algorithms implementing distributed shared memory. IEEE Computing, 23(5), May 1990.

No context found.

Michael Stumm and Songnian Zhou. Algorithms implementing distributed shared memory. IEEE Computer, pages 54-64, May 1990.

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