E. Babb. Implementing a Relational Database by Means of Specialized Hardware. ACM Transactions on Database Systems, 4(1):1--29, 1979.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....the disk resident code should be solved. Acharya et al. 1] propose a stream based programming model for these purposes. For database applications, however, significant amount of research has already been conducted. First, there is literature on database machines, which were studied some time ago [7, 31, 6, 26]. Special purpose hardware, which were employed by the database machines, had high cost and moderate performance, which eventually led the demise of database machines. Smart disk systems, on the other hand, use commodity hardware, lowering the cost of the system. Also the VLSI technology has ....

....able to execute channel programs that perform I O on behalf of their hosts. Database machines employed processors on different levels of the disk architecture. For example, Banerjee et al. 7] proposed putting a processor per disk head. The others offered processor per track and processor per disk [31, 6, 26]. Unfortunately, the special purpose components eventually led to the demise of earlier database architectures. As we have mentioned in Section 2, there are many differences between our work and the database machines. First of all, smart disk systems use commodity hardware, which makes them ....

E. Babb. Implementing a relational database by means of specialized hardware. ACM Transactions on Database Systems, Vol. 4, No 1., March 1979.

....the disk resident code should be solved. Acharya et al. 1] propose a stream based programming model for these purposes. For database applications, however, significant amount of research has already been conducted. First, there is literature on database machines, which were studied some time ago [7, 31, 6, 26]. Special purpose hardware, which were employed by the database machines, had high cost and moderate performance, which eventually led the demise of database machines. Smart disk systems, on the other hand, use commodity hardware, lowering the cost of the system. Also the VLSI technology has ....

....able to execute channel programs that perform I O on behalf of their hosts. Database machines employed processors on different levels of the disk architecture. For example, Banerjee et al. 7] proposed putting a processor per disk head. The others offered processor per track and processor per disk [31, 6, 26]. Unfortunately, the special purpose components eventually led to the demise of earlier database architectures. As we have mentioned in Section 2, there are many differences between our work and the database machines. First of all, smart disk systems use commodity hardware, which makes them ....

E. Babb. Implementing a relational database by means of specialized hardware. ACM Transactions on Database Systems, Vol. 4, No 1., March 1979.

....model in which the smart disk resident code and the host resident code interact using streams. For database applications, however, there is some previous research to build upon. First, there is a vast amount of literature on database machine architectures which were studied long time back [6, 24, 5, 21]. Compared to these architectures, the smart disk system does not use any special purpose hardware and relies mostly on COTS components. Special purpose components with their modest performance and high costs eventually led to the demise of earlier database machines. Today s interconnection ....

....able to execute channel programs that perform I O on behalf of their hosts. Database machines employed processors on different levels of the disk architecture. For example, Banerjee et al. 6] proposed putting a processor per disk head. The others offered processor per track and processor per disk [24, 5, 21]. Unfortunately, the specialpurpose components eventually led to the demise of earlier database architectures. An important difference between these works and ours is that we take efficient query optimization techniques as well as the experience in parallel query optimization into account. We also ....

E. Babb. Implementing a relational database by means of specialized hardware. ACM Transactions on Database Systems, Vol. 4, No 1., March 1979.

....cost of applying a hash function is negligible. ffl The network cost is measured as the number of bytes transmitted. 3 Tuple Bit Vectors In this section we describe several semijoin techniques and extend them using the tuple bit vector idea. 4 3. 1 Hash Filters Hash filters were introduced in [Blo70, Bab79] and promoted later in [Mul83, Mul90, Qad88] A hash filter is a bit vector used to encode the joining relationship. When joining R i with R j , the join attribute values of R i are hashed to some addresses in the bit vector whose corresponding bits are then set to 1. A zero bit after hashing ....

E. Babb. Implementing a relational database by means of specialized hardware. ACM Transactions on Database Systems, 4(1):1--29, 1979.

.... early 1980s the field of hardware database machines was an active area of research [47] 32] The disk processing in the database machines fell into roughly four categories: processor per head (e.g. OSU s DBC [10] SURE [56] processor per track (e.g. CASSM [97] RAP [72] RARES [59] and CAFS [9]) processor per disk (e.g. SURE [56] and multi processor cache (e.g. RAP.2 [88] DIRECT [31] INFOPLEX [62] RDBM [42] and DBMAC [68] In each architecture, a central processor(s) acted as the front end of the system. For the most part, these machines were exceptionally good at pushing ....

E. Babb. "Implementing a relational database by means of specialized hardware," ACM Transactions on Database Systems, Vol. 4, No. 1, March 1979.

....for hash join: If initially tuples are already fully or partially sorted, sort based join techniques may very well outperform hash join since for sorted data, sort based join execution reduces to one linear scan of each relations. Hash join can generally expect to outperform sort merge join [7, 14, 17, 19, 23, 26]. However, mostly their performances differ by percentages rather than factors [54] In [54] Graefe et al. presented a quite detailed comparison between sort based join and hash join. Their conclusion is that these two classes of join techniques are to a large extent complimentary to each other. ....

....sort merge join could outperform hash join. 7 Some Speed Up Techniques for Join Executions This section discusses some general techniques that can be used along with other join techniques to speed up join executions. Bit vector filtering: Join execution can be improved by using bit vectors [7, 16, 21, 25, 38] for small or medium sized joins. Before performing join, an vector of n bits is initialized and each bit is set to 0. During the first phase (sorting phase for sort merge join and hashing phase for hash join) each join attribute value in R is hashed and the result is used to set bits in the bit ....

[Article contains additional citation context not shown here]

Babb E., Implementing a Relational Database by Means of Specialized hardware. ACM TODS, V4, N1, March 1979, pp. 1-29.

....loop join plans flatten out as well and show the same performance as generalized hash teams. 7 Related Work The use of bitmaps is becoming increasingly popular to support decision support queries. In the database context, bitmaps have been used to speed up the execution of joins in distributed [Bab79, VG84] as well as centralized systems [Bra84] In these proposals so called Bloom filters [Blo70] are used to filter out tuples without join partners. HM97] use bitmap signatures for processing joins involving predicates on nested sets. Also, bitmap indexing is a well known concept; see, e.g. ....

E. Babb. Implementing a relational database by means of specialized hardware. ACM Trans. on Database Systems, 4(1):1--29, March 1979.

....technique of hash filtering can be applied in a parallel 2 database environment to reduce the relation cardinalities. Note, however, that prior work on hash filters (or called bit vector filters) only considered their use on the joining attribute due mainly to the focus on linear execution trees [1] [8] 23] 31] 1 , thus not fully taking advantage of the opportunity for utilizing multiple hash filters to reduce a single relation. As can be seen later, such an opportunity is made available by the execution of a bushy tree, and can lead to a very significant reduction effect on relation ....

E. Babb. Implementing a Relational Database by Means of Specialized Hardware. ACM Transactions on Database Systems, 4(1):1--29, March 1979.

....HJ IO Cost C C P C C C HJ CPU Cost C C P C C hash C write C C write C C C comp C ( 1 2 1 2 2 1 2 1 2 1 1 2 2 1 2 1 2 4 2 = l l l Some accelerators are supported in our algorithm. A technique named bit filtering [Babb79, Dewitt85] can be used to reduce the cost of rewriting C2. The idea is to associate each bucket with a bit value in a vector. Those buckets with null value in the corresponding bit of the vector are thrown away directly after the partition so as to save CPU and IO cost. Moreover, for each bucket we manage a ....

Babb E., "Implementing a Relational Database by Means of Specialized Hardware", ACM Transactions on Database Systems, Vol 4 n1, March 1979.

....Active Disk projects at Carnegie Mellon [3] and UC Santa Barbara Maryland [4] Moving the processing of stored data closer to disks is not a novel approach. Many research projects on database machines were directed to the same goal in the late 1970s and early 1980s [5] RAP [6] CASSM [7] CAFS [8], GRACE [9] Delta [10] and so on. However, these early database machine projects did not result in practical systems, mainly because of the high costs and the long time required to develop specialized hardware. The most significant difference between the older approaches and recent work is the ....

E. Babb. Implementing a relational database by means of specialized hardwaer. ACM Trans. on Database Systems, 4(1):1--29, March 1979.

....or to the uncertainty reduction and a reliable execution direction selection. Advancement and switching of processes is controlled by the dynamic optimizer at frequent execution points. Manipulation of record sets is supplemented with heavy usage of record ID lists, bitmap filters (Babb s arrays [Babb79]) and estimation supporting structures. The second lesson has to do with integration of new technologies into the existing product. Query processing and optimization are known to be complex and deeply interconnected. Hence, execution strategies selected with a newly introduced feature tend to be ....

E. Babb, "Implementing a Relational Database by Means of Specialized Hardware," ACM Transactions on Database Systems, Vol. 4, No. 1, (March 1979).

....modeled this way if one assumes that the networks they use have sufficient bandwidth for the number of processing nodes they have. 4 Our work therefore builds on the join algorithms already proposed for those machines. See [BR85, H83] for a modern review of database machines and original papers [B79, BDFW83, BO79, D79, GS81, MH81, S79, Schm79, SG75]. For the purposes of this paper any data distribution strategy that partitions a relation roughly evenly based on some subset of its attributes will be satisfactory. The one we use is hash partitioning. To partition a relation R based on attribute A, we use a function h whose domain is the ....

....set of processing node identifiers. For each tuple t of R we put t in processor h( t. A) Thus, when doing a join based on the clause R.A = S.B, if R is already partitioned based on A using hash function h, we send each S tuple x to h(x. B) The idea of using hash partitioning for joins comes from [B79, KTM83, B84, DG85] and has been analyzed favorably [DG85] In a previous report [SS85] analyzes join processing using hash partitioning, concentrating on the case when the join fields do not constitute a key. That paper explores various data reduction strategies such as having the network filter out duplicate ....

[Article contains additional citation context not shown here]

E. Babb, "Implementing a Relational Database by Means of Specialized Hardware," ACM TODS 4,1 (March, 1979), 1-29.

....used effectively in distributed join ####################################################### DR QUERY hash(name) include DR REFERRAL hash(name) catalog function Org Names x x x Figure 7. A Catalog Function with a Bit Vector Filter. ############################################ processing [1, 3, 6, 21]; Nomenclator is the first to apply them to distributed selection predicate optimization. Indices and bit vector filters are general techniques that can be used by catalog functions for other attributes. It would also be interesting to build catalog functions that use more knowledge about the ....

E. Babb, "Implementing a Relational Database by Means of Specialized Hardware," ACM Transactions on Database Systems 4(1)(March 1979).

....the technique of hash filtering can be applied in a parallel database environment to reduce the relation cardinalities. Note, however, that prior work on hash filters (or called bit vector filters) only considered their use on the joining attribute due mainly to the focus on linear execution trees [1] [7] 25] 1 , thus not fully taking advantage of the opportunity for utilizing multiple hash filters to reduce a single relation. As can be seen later, such an opportunity is made available by the execution of a bushy tree, and can lead to a very significant reduction effect on relation ....

E. Babb. Implementing a Relational Database by Means of Specialized Hardware. ACM Transactions on Database Systems, 4(1):1--29, March 1979.

....technique of hash filtering can be applied in a parallel database environment to reduce the relation cardinalities. Note, however, that previous work on hash filters (or called bit vector filters) only considered their use on the joining attribute due mainly to the focus on linear execution trees [1, 8, 23, 31] 1 , thus not fully taking advantage of the opportunity for utilizing multiple hash filters to reduce a single relation. As can be seen later, such an opportunity is made available by the execution of a bushy tree and can lead to a very significant reduction effect on relation cardinalities, ....

Babb E (1979) Implementing a relational database by means of specialized hardware. ACM Trans Database Syst, 4(1):1--29

....affected as the amount of available memory is reduced. 1 The join algorithms were also analyzed in the presence of non uniform join attribute values. We also considered how effectively the different join algorithms could utilize processors without disks. Finally, bit vector filtering techniques [BABB79, VALD84] were evaluated for each of the parallel join algorithms. ############################# 1 This set of experiments can also be viewed as predicting the relative performance of the various algorithms when the size of memory is constant and the algorithms are required to process relations larger ....

....a query tree in a dataflow fashion. If the result of a query is a new relation, the operators at the root of the query tree distribute the result tuples on a round robin basis to store operators at each disk site. To enhance the performance of certain operations, an array of bit vector filters [BABB79, VALD84] is inserted into the split table. In the case of a join operation, each join process builds a bit vector filter by hashing the join attribute values while building its hash table using the inner relation [BRAT84, DEWI84, DEWI85, VALD84] When the hash table for the inner relation has been ....

[Article contains additional citation context not shown here]

Babb, E., "Implementing a Relational Database by Means of Specialized Hardware" ACM Transactions on Database Systems, Vol. 4, No. 1, March, 1979.

....search space for the query. Some catalog functions use relatively static information to constrain the search, like knowledge about the conditions used to distribute data to data repositories (called the partitioning criteria of a relation. Other catalog functions build indices or hash filters [1] to capture the distribution patterns in changing data, or dynamically search the network for information to speed query processing. Still others use semantic constraints like information about integrity constraints or the domains of attributes to constrain the search. The active catalog uses ....

....used one catalog and one data access function during the experiments. An initial referral to the catalog function was available from the distributed catalog service. After being started by Nomenclator, the catalog function used an internal Nomenclator relation to retrieve bit vector filters[1] that described the hash values of the attribute to be selected at each data repository. The catalog function compared the hash value of the attribute in the query with those in the filter to decide which data repositories to include in the referrals it generated for the query resolver. Data ....

E. Babb, "Implementing a Relational Database by Means of Specialized Hardware," ACM Transactions on Database Systems 4(1), pp. 1-29 (March 1979).

....and advances in hardware technology would lead to the widespread use of commercial database machines. However, while most projects appeared promising initially, it is only very recently that the first of these special purpose computers are becoming commercially available. The ICL CAFS machine [MCGR76, BABB79] has been shipped in small quantities. The Britton Lee IDM (Intelligent Database Machine) appears to be the first database machine to reach the market place in large volumes. 2 Despite these exceptions, the overwhelming evidence is that the majority of the database machine designs proposed will ....

Babb, E. "Implementing a Relational Database by Means of Specialized Hardware," ACM TODS, Vol. 4, No. 1, March 1979.

....range within the relation, etc. and make sure to scan it in the same way. Storing the access method used should use very little space. Note that only one pass of the data is needed to apply the backward semijoin. 2. 2 Improving Hash Filters with Tuple Bit Vectors Hash filters [Blo70, Bab79] can be used in a fashion similar to semijoin projections. Rather than transmitting the values for the join attribute, one hashes the values into a hash table and transmits the table. The hash table is often significantly smaller than the total size of the projection, although they do not give ....

E. Babb. Implementing a relational database by means of specialized hardware. ACM Transactions on Database Systems, 4(1):1--29, 1979.

....transitions of the Sort Merge and nested loops algorithms become more obvious in the environment of this query. Figure 4 reflects the performance of the join algorithms on the same query used for Figure 3. The difference is that in Figure 4 all the algorithms use bit vector filtering techniques [BABB79, BRAT84, VALD84]. The notable performance improvements demonstrated are the result of eliminating, at an earlier stage of processing, those tuples that will not produce any result tuples. The bit vector filtering technique used by the hash partitioning and Sort Merge algorithms are very similar. 4 Prior to the ....

....(phantom) tuples being propagated along to the final joining process. The phantom tuples will, however, be eliminated by the final joining process. The number of phantom tuples can be reduced by increasing the size of the bit vector or by splitting the vector into a number of smaller vectors [BABB79]. A separate hash function would be associated with each of the smaller bit vectors. The costs associated with bit vector filtering are modest. For the given test, a single bit vector of length 4K bytes was used. Since the hash partitioning algorithms already compute the hashed value of each ....

[Article contains additional citation context not shown here]

Babb, E., Implementing a Relational Database by Means of Specialized Hardware, ACM Transactions on Database Systems, Volume 4, No. 1, 1979.

....search space for the query. Some catalog functions use relatively static information to constrain the search, like knowledge about the conditions used to distribute data to data repositories (called the partitioning criteria of a relation) Other catalog functions build indices or hash filters [4] to capture the distribution patterns in changing data, or dynamically search the network for information to speed query processing. Still others use semantic constraints like information about integrity constraints or the domains of attributes to constrain the search. Information in the active ....

....all the values for an attribute in an index data structure, but this is only feasible for a few attributes with a small number of values. More frequently, catalog functions use a variation of an index, called a bit vector filter, that has been used to improve the performance of distributed joins [4, 13, 32, 113]. Catalog functions hash each value of an attribute in a data repository, and then set the corresponding bits in the bit vector filter (see Figure 3.2) The catalog function hashes the value for the same attribute in a query, and then checks to see if the corresponding bit is set in the filter. ....

Babb, E., "Implementing a Relational Database by Means of Specialized Hardware," ACM Transactions on Database Systems 4(1), pp. 1-29 (March 1979).

....their functionality on the multiple slave processors. The master then simply forms the union of the results passed back from the slave processes. Only in the case of joins is a more complex operation required, and some additional communications overhead is created. In this case, semi joins [1] are used to exploit the parallelism provided by the DAFS architecture. Other data manipulation operations which can cause problems are those which generate the history of a particular attribute or are based on sequenced information. As the data is distributed using a hash based function, ....

E. Babb. Implementing a relational database by means of specialised hardware. ACM Transactions on Database Systems, March 1979.

No context found.

E. Babb. Implementing a Relational Database by Means of Specialized Hardware. ACM Transactions on Database Systems, 4(1):1--29, 1979.

No context found.

E. Babb. Implementing a Relational Database by Means of Specialized Hardware. ACM Transactions on Database Systems, 4(1):1--29, 1979.

No context found.

Babb, E., "Implementing a Relational Database by Means of Specialized Hardware", ACM Transactions on Database Systems, Vol. 4, No. 1, March, 1979.

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