D. DeWitt and J. Gray. Parallel database systems: The future of database processing or a passing fad? ACM-SIGMOD record, 19(4), 1990.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....KNOWLEDGE BASE In a shared nothing architecture for a multiprocessor database [21] neither main memory nor secondary storage is shared amongst the processors. This approach is a special case of a database machine, whose aim is to speed up data processing in a single site DBMS by using parallelism [4]. Thus, at the physical DBMS level, we can view a shared nothing architecture as a homogeneous distributed database, where each site includes a processor together with its own main memory and secondary storage devices on which each local database resides. In such an environment the only shared ....

....processors, and such that EDB = wW V(w) Herein we do not deal with the difficult data placement problem of distributing the facts of EDB among the sites w W in such a way so that the performance of query processing is maximised. This activity, which is known as declustering, is discussed in [4, 17]. We also associate with IDB a logic program, LP, such that for each Datalog rule, say R, in IDB we create a corresponding Datalog K rule, say R, in LP. Let A be an atom; each literal in R of the form A is mapped to A, KA or PA in R, and each literal in R of the form A is mapped to A, KA or PA ....

D.J. DeWitt and J. Gray, Parallel database systems: The future of database processing or a passing fad?, ACM SIGMOD Record 19 (1990), 104-112.

....as queries become more complex and data sets grow larger. At the same time it is becoming more difficult to extract additional performance from uniprocessor machines needed to meet this demand. Multi processor computers have already proven to be cost effective for solving many classes of problems [DEWI90], and 2 appear to be an attractive way to increase query processing performance. On these parallel systems, relations are typically declustered across the disks. Such declustering allows selection operators to be distributed to the processors that contain fragments of the relation to be queried. ....

D. J. DeWitt and J. Gray, Parallel Database Systems: The Future of Database Processing or a Passing Fad?, SIGMOD Record, Vol. 19, No. 4, December 1990.

....high throughput while the latter requires good response times. Two important decisions have guided the design of DBS3. First, DBS3 implements a parallel dataflow execution model based on fragmented data placement similar to distributed memory (shared nothing) systems like BUBBA [9] and GAMMA [7]. This allows us to take advantage of automatic load balancing while reducing access conflicts to the shared memory. Second, the ESQL compiler translates a query into an optimized parallel program that exploits both inter and intra operation parallelism and yields decentralized execution ....

....search space, rather than relying on restrictive heuristics, is important because a large variety of parallel execution plans can better provide a good trade off between response time minimization and throughput maximization. Optimizing queries for parallel execution is considered an open problem [7]. 16] drastically restricts the optimization search space by the adoption of two heuristic assumptions: the buffer size independent hypothesis and the two phase hypothesis. However, these assumptions rely on a restricted execution model (left deep trees only, no inter operation parallelism) 11] ....

D.J. DeWitt and J. Gray. Parallel database systems: The future of database processing or a passing fad ? ACM SIGMOD Record, 19(4):104--112, 1990.

....a plan that considers sum and max of operators working sequentially and in parallel, respectively. The results obtained from a prototype optimizer are presented. 1 Introduction The relational data model has been found especially suited for massive parallelization due to its set oriented nature [3, 8]. The focus in query optimization so far has been on finding least work plans, since on a uniprocessor the time taken is proportional to the work done. Since the most efficient uniprocessor solutions often have sequential dependencies, making them difficult to parallelize, an important aim is to ....

....but not pipelining. The Papyrus project [9] has developed a model which considers inter operator dependent and independent parallelism, as well as operator cloning, i.e. intra operator parallelism. Cost Model for Query Plans: No good cost models currently exist for the parallel environment [3]. Cost models for the sequential environment consider the cost of a query plan to be the sum of the costs of its components [7] This observation does not hold in the parallel environment, where the cost of the plan is the sum of the costs of the tasks only on the critical path (of the tree ....

[Article contains additional citation context not shown here]

D. J. DeWitt and J. Gray. Parallel database systems: The future of database processing or a passing fad? CACM, 35(6), June 1992.

....environment [SELI79, MACK86] and continues to be done [LIPT90] no good cost models exist for the parallel environment. To quote DeWitt and Gray, While the necessary optimizer technology exists, accurate cost models have not been developed, let al..one validated. More work is needed in this area [DeWI90]. LU91] proposed a cost model in which response time of a query plan is considered. Not having pipelining between successive levels of the plan tree simplifies its calculation considerably. Query execution is modeled in a data flow manner in [WILS91] and a flow rate based expression for the ....

D. J. DeWitt and J. Gray, "Parallel Database Systems: The Future of Database Processing or a Passing Fad?," in ACM SIGMOD Record, Vol. 19, No. 4, December 1990.

....4 we present the results of our simulation experiments and discuss their implications. Concluding remarks are given in Section 5. 2 Overview Our objective is to build a database system that captures the key advantages of the SN and SE approaches. From the discussions in the literature [Sto86, DG90, DG92] we know that the main drawback to sharing is scalability. We identify the following simple guidelines to building a scalable system: 1. Eliminate central resources that can become bottlenecks. 2. Minimize communication and thus avoid making the interconnect a bottleneck. 3. Utilize ....

D. J. DeWitt and J. Gray. Parallel database systems: The future of database processing or a passing fad? SIGMOD RECORD, 19(4):104--112, December 1990.

.... out in the context of the Starfish project, in which a complex object server is developed for the distributed operating system Amoeba [TRSS89] In the database world there is a general consensus that the disk I O seems to be the bottleneck for such shared nothing systems with a local area network [DeGr90]. This paper is organized as follows. In Section 2 we describe the complex object benchmark we used in our performance evaluation. The benchmark has been based on the Altair Complex Object Benchmark [DFMV90] and involves object retrieval, navigation, and updates. In Section 3 we present several ....

D. J. DeWitt & J. Gray, "Parallel Database Systems: The Future of Database Processing or a Passing Fad?," SIGMOD RECORD 19 (4), 1990, pp. 104--112.

....size. 1.2 Motivation We need parallelism of R tree because of the following reasons: 1. Increase the Database capacity. 2. Increase the throughput. 3. Minimize the response time. A high throughput and short response time DBMS using parallelism of R tree should satisfy the following requirements[2], 5] Scalable No distinguished processor will be a bottleneck. As a corollary, queries with large search rectangle should activate as many processors as possible. Minimum Load Queries with small search rectangle should activate few processors so that other processors can be used to handle ....

....for the multimedia data. Section 6 gives the conclusions and the future direction. 2 Survey We assume the hardware architecture consists of r( 2h; h 1) processors connected by a network and each processor has a local disk attached. There are different approaches in designing parallel R tree[2], 5] 10] 1. Independent R trees 2. Super Nodes 3. Node Distribution(Multiplexed R tree) 2.1 Independent R trees Each processor stores and manipulates a subset of the data set. Independent R tree is built in each processor. Searching is executed in each processor separately. There are two ....

David J. DeWitt and Jim Gray. Parallel database systems: The future of database processing of a passing fad? In Proc. ACM SIGMOD, volume 19, pages 104--12, December 1990.

....placement determines the nodes on which these copies will be stored. ACOS: With ACOS the goal of the allocation step is to minimize the overall response time for queries and updates. We have a shared nothing system with a local area network. For such systems the disk IO seems to be the bottleneck [9]. Therefore, we try to use the allocation to enlarge the effective I O bandwidth by parallel disk I Os. 4.4 Mapping of fragments on storage models A next step is mapping the fragments on storage models. The minimum unit of retrieval, and the clustering of these units into files is determined. ....

D. J. DeWitt and J. Gray, "Parallel Database Systems: The Future of Database Processing or a Passing Fad?," ACM SIGMOD Record 19 (4), pp. 104--112, 1990.

....the system to scale up, i.e. as the size of the database grows, additional disks can be introduced to maintain the desired performance. We shall start by focusing on a single user system. Time permitting, we shall investigate a multi user environment assuming a shared nothing architecture [DG90] Based on how the data is assigned to the disks and the employed memory management policies, a scheduler must assign resources (i.e. disk bandwidth and memory) intelligently in order to ensure a hiccup free display while minimizing the amount of resources required to support this display. The ....

D. J. DeWitt and J. Gray. Parallel database systems: The future of database processing of a passing fad? SIGMOD RECORD, 19(4):104--112, December 1990.

....much attention has been paid to the development of parallel DBMSs. Using special purpose hardware has shown not to be successful; instead, a parallel DBMS running on general purpose, shared nothing hardware appears to be the right choice, both from a scientific and from a commercial viewpoint [DeG90]. Also, various query processing strategies have been implemented: dataflow query processing appears to be superior to control flow scheduling [DGS90,DeW79,TeB91] Therefore, this paper studies query processing in a general purpose shared nothing dataflow architecture. Teradata [Ter83] GAMMA ....

....parallelism for one query, turns query optimization into a difficult problem, that cannot be solved using conventional query optimization techniques. So far, little research has been done in this research area although it is identified to be important for the further development of parallel DBMSs [BAC90,DeG90,PMC90]. The query optimizers for most parallel DBMSs are based on the theory developed in [SAC79] however, which is not particularly fit for parallel dataflow query processing. For example, this type of query optimizers only considers linear query trees, although this class of trees does not ....

[Article contains additional citation context not shown here]

D. J. DeWitt & J. Gray, "Parallel Database Systems: The Future of Database Processing or a Passing Fad?," SIGMOD RECORD 19 (1990), 104 -- 112.

....the tuples from the smaller input relation, and then probing the hash table with the tuples from the corresponding partition of the larger input relation. The tasks involved in a hybrid hash join are frequently implemented as a collection of processes, particularly in parallel database systems [DG92] One benefit of doing so is that the construction (and later probing) of the hash table for a given join operation can proceed in parallel with the reading of the input relations from disk. A more substantial benefit, particularly for complex queries, is that scalable parallel data flow ....

D. J. DeWitt and Jim Gray. "Parallel Database Systems: The Future of Database Processing or a Passing Fad?". Communication of the ACM, June, 1992.

....than discussing 2 In the ADAMS implementation, after the first request phase the servers begin filling the next request prior to receiving it, an even more aggressive pre fetch. 43 pre fetching in more general terms. 5. 5 Simple I O Parallelism If I O is seen as the major database bottleneck [DG90] a simple method of improving performance is to parallelize the I O operations by off loading them to parallel page server processors. Several recent object oriented systems are based on page server architectures, such as EXODUS, Gemstone, O2, ObjectStore, and SHORE [CFZ94] If several I O ....

David J. Dewitt and Jim Gray. Parallel Database Systems: The Future of Database Processing or a Passing Fad? SIGMOD Record, 19(4):104--112, December 1990.

....to be retrieved from the network and unwrapped by the receiving operating system. So, sending a tuple over the network implies CPU costs on the sending and receiving processor, and actual transmission, which implies a delay. In general, the CPU costs involved, appear to be the limiting factor [DeG90], and, therefore, the rate in which tuples are transported over the network is determined by the capacity of the CPUs that send and receive the tuples and not by the capacity of the network hardware. 3 An analytical model for dataflow query execution In this section, an analytical model for ....

D. J. DeWitt & J. Gray, "Parallel Database Systems: The Future of Database Processing or a Passing Fad?," SIGMOD RECORD 19 (1990), 104 -- 112.

....relative to those needed by the R tree and R tree. 1 Introduction Parallel database systems have been the subject of increasing attention. This is due in part to the advent of highly parallel architectures, adoption of the relational model, and challenges posed by object oriented systems [13, 25]. Much of the parallel database research has focused on multi attribute declustering techniques (such as Bubba s extended range declustering [7] and multi attribute grid declustering [18] data placement [11] and intra operator parallelization [12] Topics such as algorithms for manipulating ....

D. J. DeWitt and J. Gray, Parallel database systems: the future of database processing or a passing fad?, SIGMOD Record, 19, 4 (December 1990), 104--112.

....parallel database server. It can run on either a MPP (massively parallel processing) system or a clusters of stand alone computers that are connected with a high speed network. We refer to each computer within a MPP system or cluster as a node. XPS uses shared nothing architecture [Sto86] [DG90], each node runs a separate coserver that is equivalent to a single INFORMIX OnLine Dynamic Server 7.x instance. An XPS server includes multiple coservers communicating directly to shared data and performing parallel execution. It provides: ffl Near linear scalability. ffl Near linear speedup. ....

D. DeWitt and J. Gray. Parallel database systems: The future of database processing or a passing fad? sigmod, 19(4):104--112, December 1990.

....systems were replaced by the usage of conventional parallel hardware architectures, the research on the design of parallel software architectures gained an important role in database research in the last few years. Generally parallelism is employed in parallel database systems by declustering [5] and or operator parallelization [7] as inter operator and intraoperator parallelism. Inter operator parallelism is realized by executing different operators of a partitioned query execution plan in parallel. In contrast intraoperator parallelism executes the same operation in parallel on a ....

D.J. DeWitt, J. Gray, Parallel Database Systems: The Future of Database Processing or a Passing Fad?, SIGMOD record, 19, 4, Dec. 1990

....parallel SPMD 2 implementations of SEAR and SPEAR on an IBM SP2 messagepassing multiprocessor and the results of our speedup and scale up experiments are presented. We assume the 2 short for Single Program Multiple Data shared nothing paradigm that is commonly used in parallel databases [12, 13, 6] and the nature of the problem suits this assumption well. Our findings confirm our initial expectations that both algorithms parallelize well, require only a comparatively small amount of communication and achieve near linear speed up that is only diminished by the sequential portions of the ....

....expectation that our algorithms size up 1 well, i.e. running times are linear in the problem size. Figure 6.5 confirms this expectation for a range from 500,000 to 40 million transactions. All other parameters were chosen as above. 1 In addition to speed up and scale up, this term is used in [12] as third criteria to evaluate parallel database performance for the case when the number of processors remains constant while the problem size is increased. 0 50 100 150 200 250 300 350 400 450 500 5 M 10 M 20 M 30 M 40 M time ( sec ) number of transactions PEAR 0.3 PEAR 0.75 PPAR 0.3 ....

David DeWitt and J. Gray. Parallel database systems: The future of database processing or a passing fad? sigmod, 19(4):104--112, December 1990.

....2 tuples [12] as examples of complex objects. Our concepts, however, hold for more general objects as well. Also, we mainly focus on disk I O. In the database world there is a general consensus that the disk I O seems to be the bottleneck for such shared nothing systems with a local area network [7]. This paper is organized as follows. In Section 2 we describe the complex object benchmark we used in our performance evaluation. The benchmark has been based on the Altair Complex Object Benchmark [6] and involves object retrieval, navigation, and updates. In Section 3 we present several storage ....

D. J. DeWitt and J. Gray, Parallel Database Systems: The Future of Database Processing or a Passing Fad?, SIGMOD RECORD 19 (1990) 104--112.

....Section 4 we present the results of our simulation experiments and discuss their implications. Concluding remarks are given in Section 5. 2 Overview Our objective is to build a database system that captures the advantages of both the SN and SE approaches. From the discussions in the literature [26, 8, 6] we know that the main drawback to sharing is scalability. We identify the following simple guidelines to building a scalable system: 1. Eliminate central resources that can become bottlenecks. 2. Minimize communication and thus avoid making the interconnect a bottleneck. 3. Utilize modular, ....

D. J. DeWitt and J. Gray. Parallel database systems: The future of database processing or a passing fad? SIGMOD RECORD, 19(4):104--112, December 1990.

....of such 24 Theta7 systems include database management for multinational companies with a global reach, hospital management systems, a round the clock shopping service, etc. In order to service such applications, next generation databases will be required to keep their data on line all of the time [Dewi90, Silb90]. One implication of completely on line operation is that maintenance operations like checkpoint This research was partially supported by the National Science Foundation under grant IRI 8657323, by an IBM Research Initiation Grant, and by a University of Wisconsin Vilas Fellowship. ing, index ....

DeWitt, D. J. and Gray, J., "Parallel Database Systems: The Future of Database Processing or a Passing Fad?", SIGMOD Record, 19(4), Dec. 1990.

....The total cost of the query is calculated by adding these costs. The model does not accountfor overlap of communication and CPU processing, or contention for network resources. 3. SPATIAL DECLUSTERING In a parallel shared nothing system, the main source of parallelism is partitioned parallelism [5].Partitioned parallelism is achieved by declustering the dataacrossmultiple nodesinthesystem, and then running operators at eachof these nodes. There are two main requirements for achieving effective parallelism. First, good data declustering techniques are required to evenly distribute the data ....

D.J.DeWitt andJ.Gray."Parallel Database Systems: The Future of Database Processing or a PassingFad?". Communication of the ACM,June, 1992.

....still works correctly, it will only suffer a performance degradation due to unsuccessful searches of the index for tuples stored on other processors. 6 Results Scaleup and speedup are useful metrics for evaluating the performance of a parallel algorithm on a multiprocessor database machine [DG90] Scaleup is an interesting metric for multiprocessor database machines as it indicates whether a constant response time can be maintained as the workload is increased by adding a proportional number of processors and disks. Speedup is an interesting metric because it indicates whether additional ....

D. DeWitt and J. Gray. Parallel database systems: The future of database processing or a passing fad. ACM SIGMOD Record, 19(4), December 1990.

....paper differs from the study in [ZAT97] as we propose two algorithms for parallel spatial joins, and evaluate them on a real system using a number of large data sets. 3 Spatial Declustering Techniques In a parallel shared nothing system, the main source of parallelism is partitioned parallelism [DG92] Partitioned parallelism is achieved by declustering the data across multiple nodes in the system, and then running operators at each of these nodes. There are two main requirements for achieving effective parallelism. First, good data declustering techniques are required to evenly distribute ....

D. J. DeWitt and Jim Gray. "Parallel Database Systems: The Future of Database Processing or a Passing Fad?". Communication of the ACM, June, 1992.

No context found.

D. DeWitt and J. Gray. Parallel database systems: The future of database processing or a passing fad? ACM-SIGMOD record, 19(4), 1990.

First 50 documents

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