Karpjoo Jeong. Fault-tolerant Parallel Processing Combining Linda, Checkpointing, and Transactions. Ph.D. Thesis, Courant Institute of Mathmatical Sciences, New York University, January, 1996.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....good performance, many data mining applications have proved difficult to parallelize. Fortunately, the computation is coarse grain parallel, i.e. it can be parallelized into large, seldom interacting tasks. Coarse grain parallel computations are suitable computations for networks of workstations [48]. The time it takes to create, coordinate, and terminate parallel tasks is critical. Therefore a good parallel programming environment and runtime environment is critical to parallel data mining. In the remainder of this section, we survey four distributed computing systems that run on networks of ....

....be very expensive. The current implementation of Piranha is not fully fault tolerant and it does not run on networks of heterogeneous workstations. Commercial products based on Linda and Piranha are available from Scientific Computing Associates. 2.4. 6 Persistent Linda Persistent Linda (PLinda) [20, 47, 48] is a set of extensions to Linda to support robust parallel computation. It is currently under development at New York University. The three major extensions to Linda are: ffl Lightweight transactions. Transactions are a sequence of operations that are always executed atomically regardless of ....

Karpjoo Jeong. Fault-tolerant Parallel Processing Combining Linda, Checkpointing, and Transactions. Ph.D. Thesis, Courant Institute of Mathmatical Sciences, New York University, January, 1996.

....that have to be independently analyzed, each one requiring considerable computational e#ort. One common idea to solve this kind of problems is based on using the idle times of LAN, WAN (or Web ) interconnected workstations, trying to make use of their potential CPU power, currently wasted. In [10], K. Jeong gives a good introduction to the problem, many references to di#erent solutions, and his own solution based on a fault tolerant approach, using Linda, checkpointing and transactions. What we shall see here is how the problem can be solved using our model. 16 In the first place, the ....

....to serve as the basic model where both existing and new components can be integrated to build up the required applications. With regard to the example used in this paper for showing the expressiveness and adequacy of our model, it constitutes the skeleton of many real and useful applications. In [10] an elementary particle physics problem is cited, while it can also be used in other NP hard problems, like first order logic theorem proving, whose natural complexity can be approached by using distributed solutions based on the master worker approach. With our model we also obtain a simple ....

K. Jeong, Fault-Tolerant Parallel Processing Combining Linda, Checkpointing and Transactions, PhD thesis (Dept. of Computer Science, New York University, 1996).

....grain massive parallel computation with 10 5 to 10 8 large records that have to be independently analyzed, each one requiring considerable computational e#ort. One common idea to solve this type of problems is based on using the idle times of LAN, WAN (or Web ) interconnected workstations. In [7], K. Jeong gives a good introduction to the problem, many references to di#erent solutions, and his own solution based on Linda, checkpointing and transactions. What we shall see here is how the problem can be easily solved using our model. In the first place, the solution to the problem can be ....

K. Jeong. Fault-Tolerant Parallel Processing Combining Linda, Checkpointing, and Transactions. PhD Dissertation, Dept. of Computer Science, New York University, 1996.

....Linda 3[10] and involved the addition of a type ts and a primitive tsc. The idea of adding multiple tuple spaces has led to many different proposals of how multiple tuple spaces could be incorporated within Linda[12,13] and many implementations include multiple tuple spaces in one form or another[8,19,11,17,14,15]. Multiple tuple spaces were introduced as an effective way of hiding information. Information within a tuple space can only be accessed by those processes that know about the tuple space. As the use of Linda has changed to incorporate different styles of distributed computing the need to hide ....

K. Jeong. Fault-tolerant parallel processing combining Linda, checkpoint ing and transactions. PhD thesis, New York University, 1996.

....Examples include seismic computations [1] and materials science [13] Typically the computation is controlled by a single master process and the data manipulated by the computation is located on a single disk with all I O being performed by the master. A fault tolerant version of this structure [6, 2, 10] allows cheap recovery since only the particular task affected by a machine failure needs to be recovered. It is possible to increase capacity and bandwidth of storage at a single machine using RAID techniques [5] but in some computations the data manipulated outstrips the capacity of a single ....

K. Jeong. Fault-Tolerant Parallel Processing Combining Linda, Checkpointing, a nd Transactions. PhD thesis, New York University, Jan. 1996.

....design study [4] considered the use of atomic actions as a mechanism to support fault tolerant parallel programming over a NOW. Fault tolerance for a bag of tasks type structure has been considered before, e.g. 3] 8] but without providing access to large scale data on secondary storage. Plinda [11] which supports access to persistent tuple spaces and a transaction mechanism does have some similarity to this work. The experiments described here attempt to exploit parallelism in a NOW of modest scale to perform large scale computations in a fault tolerant way without altering operating system ....

Karpjoo Jeong. Fault-Tolerant Parallel Processing Combining Linda, Checkpointing, and Transactions. PhD thesis, New York University, Department of Computer Science, January 1996.

....design study [4] considered the use of atomic actions as a mechanism to support fault tolerant parallel programming over a NOW. Fault tolerance for a bag of tasks type structure has been considered before, e.g. 3] 7] but without providing access to large scale data on secondary storage. Plinda [10] which supports access to persistent tuple spaces and a transaction mechanism does have some similarity to this work. The experiments described here attempt to exploit parallelism in a NOW of modest scale to perform large scale computations in a fault tolerant way without altering operating system ....

Karpjoo Jeong. Fault-Tolerant Parallel Processing Combining Linda, Checkpointing, and Transactions. PhD thesis, New York University, Department of Computer Science, January 1996.

....restored tuple space. That is, a single failure leads to massive rollback. This method is fastest in the failure free case. ffl Message Logging and Replay. This method is not supported in the current release of PLinda. All three methods are guaranteed to be correct. The interested reader may see [15] for a proof. The default in PLinda is to use the commit consistent execution method. This method is currently called Process Private snapshot in the user interface (this will be changed to continuation committing in the near future) To change the method, click on the settings button of the ....

Karpjoo Jeong. Fault-Tolerant Parallel Processing Combining Linda, Checkpointing, and Transactions. PhD thesis, Department of Computer Science, New York University, January 1996.

....of restoring tuple space from the latest checkpointed state. Some committed updates may have been lost, but the checkpoint contains a consistent state from which applications can resume execution. If that execution terminates, it will be a possible execution of the corresponding Linda program[18]. The checkpointing frequency is a runtime parameter and can be decided by the user or runtime kernel, according to the size of shared memory and the reliability of processors where tuple space servers are running. 3 Tunable Fault Tolerance Mechanisms Three Fault Tolerance Mechanisms PLinda ....

....implicitly include heterogeneous processing. Results obtained from experiments on the isolated Solaris network are the average of three runs. All other results are the average of ten runs. Tradeoffs Between the Three Fault Tolerance Mechanisms Using a previous prototype version of PLinda [19, 18], we conducted experiments to examine how the three fault tolerance mechanisms work on different applications. For these experiments we designed a simple master worker program where the master spawns eight worker processes and generates task description tuples, and each worker repeatedly executes ....

K. Jeong. Fault-tolerant Parallel Processing Combining Linda, Checkpointing, and Transactions. PhD thesis, New York University, Department of Computer Science, 1996.

....tuple space from the latest checkpointed state. Some committed updates may have been lost, but the checkpoint contains a consistent state from which applications can resume execution. If that execution terminates, it will be a possible failure free execution of the corresponding Linda program[19]. The checkpointing frequency is a runtime parameter and can be decided by the user or runtime kernel, according to the size of shared memory and the reliability of processors where tuple space servers are running. 3 Tunable Fault Tolerance Mechanisms Three Fault Tolerance Mechanisms PLinda ....

....implicitly include heterogeneous processing. Results obtained from experiments on the isolated Solaris network are the average of three runs. All other results are the average of ten runs. Tradeoffs Between the Three Fault Tolerance Mechanisms Using a previous prototype version of PLinda [20, 19], we conducted experiments to examine how the three fault tolerance mechanisms work on different applications. For these experiments we designed a simple master worker program where the master spawns eight worker processes and generates task description tuples, and each worker repeatedly executes ....

K. Jeong. Fault-tolerant Parallel Processing Combining Linda, Checkpointing, and Transactions. PhD thesis, New York University, Department of Computer Science, 1996.

....monitoring of parallel computation is very difficult. In this subsection, we explain how to extend the parallel algorithm in subsection 3.2. Since it is not easy to support fault tolerance through application code alone, we use a fault tolerant computing system called Persistent Linda or PLinda[16, 17, 18] developed at New York University. PLinda is based on the shared memory Linda model[19, 15] and guarantees that shared memory called tuple space survives failure 1 . In the PLinda model, each process executes a series of steps (each step called an atomic action or transaction) that are ....

K. Jeong, Fault-tolerant Parallel Processing Combining Linda, Checkpointing, and Transactions, PhD thesis, New York University, (1996).

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