J.S. Karlsson, W. Litwin, and T. Risch. LH*lh: A Scalable High Performance Data Structure for Switched Multicomputers. In Proc. of Intl. Conf. on Extending Database Technology (EDBT'96), pages 573-591, 1996.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....section. For LH , insertions usually require one message (from client to server) and three messages in the worst case. Data retrieval requires one extra message as the requested data has to be returned. In this paper, we present a performance evaluation of a variant of the LH SDDS, called LH LH [3], for a cluster of workstations (COW) Previous studies already showed good performance of LH LH when targeting multicomputers. In [2] an actual implementation for a Transputer based machine has been evaluated, while in [7] simulation was applied to investigate LH LH s performance for a ....

....the growing of the LH data structure. In this study, we limit our discussion to the splitting within SDDSs. 3. THE LH LH SDDS The LH SDDS is a generalization of Linear Hashing (LH) to a distributed memory parallel system [5] In this paper, we focus on a particular variant of LH , called LH LH [2, 3]. The LH LH data is stored over a number of server processes and can be accessed through dedicated client processes. These clients form the interface between an application and LH LH. For this study, we assume that each server stores one LH LH bucket of data. Globally, the servers apply the LH ....

J. S. Karlsson, W. Litwin, and T. Risch. LH*lh: A scalable high performance data structure for switched multicomputers. In Advances in Database Technology, pages 573--591, March 1996.

....single element operations to the hash table, such as testandset( in order to provide locks and leases to services that may have many service instances competing to write to the same hash table element. 8 Related Work Litwin et al. s scalable, distributed data structures (SDDS) such as RP [22, 26] helped to motivate our own work. RP focuses on algorithmic properties, while we focused on the systems issues of implementing an SDDS that satis es the concurrency, availability, and incremental scalability needs of Internet services. Our work has a great deal in common with database ....

J. S. Karlsson, W. Litwin, and T. Risch. LH*LH: A Scalable High Performance Data Structure for Switched Multicomputers. In Proceedings of the 5th International Conference on Extending Database Technology, pages 573-591, Avignon, France, Mar 1996.

....usually require one message (from client to server) and three messages in the worst case. Data retrieval requires one extra message as the requested data has to be returned. In this paper, we evaluate the performance of a variant of the LH SDDS, called LH LH, which was proposed by Karlsson [4, 5]. For this purpose, we use a simulation model which is based on the architecture of a Parsytec CC multicomputer. With this model, we investigate how scalable the LH LH SDDS actually is and which factors affect its performance. The reason for our interest in the LH LH SDDS (and the Parsytec CC) is ....

....a certain threshold. In this study, we limit our discussion to the splitting within SDDSs. 3 The LH LH SDDS The LH SDDS is a generalisation of Linear Hashing (LH) to a distributed memory parallel system [7] In this study, we focus on one particular implementation variant of LH , called LH LH [4, 5]. The LH LH data is stored over a number of server processes and can be accessed through dedicated client processes. These clients form the interface between the application and LH LH.We assume that each server stores one LH LH bucket of data, which implies that a split always requires the ....

[Article contains additional citation context not shown here]

J. S. Karlsson, W. Litwin, and T. Risch. LH*lh: A scalable high performance data structure for switched m ulticomputers. In Advances in Database Technology, pages 573--591, March 1996.

....of replication. It is also shown how these changes can occur even in presence of of foreign keys on the tables that are split merged. l A new indexing technique, LH3, has been developed that extends earlier work in scalable data structures, especially LH [Litwin80] LH [Litwin93a] and LH LH [KARL96]. LH and LH LH are distributed data structures which have been modified to exist in an environment with replication and transactions. Also a new compressed and distributed B tree has been developed. l The assumption of cheap communication cost achieved by new technological development has ....

....change which previously only have been reported in conjunction with software change. It also extends the possibilities of on line schema changes to handle split and merge of tables. Most results in this section are new. Chapter 11 describes an extension of LH [Litwin80] LH [Litwin93a] and LH LH [KARL96], the LH 3 index. It is shown how this index is integrated with the handling of replicas and also how it provides very few cache misses in accessing the index data. Chapter 12 describes a new compressed and distributed B tree, the HTTP tree. It is useful for storing indexes of file names, ....

[Article contains additional citation context not shown here]

J.S. Karlsson, W. Litwin, T. Risch, LH*LH: A Scalable High Performance Data Structure for Switched Multicomputers, 5th Int'l Conf on Extending Database Technology (EDBT `96).

No context found.

J.S. Karlsson, W. Litwin, and T. Risch. LH*lh: A Scalable High Performance Data Structure for Switched Multicomputers. In Proc. of Intl. Conf. on Extending Database Technology (EDBT'96), pages 573-591, 1996.

No context found.

J.S. Karlsson, W. Litwin, T. Risch: LH*lh: A Scalable High Performance Data Structure for Switched Multicomputers, Intl. Conf. on Extending Database Technology (EDBT'96), 1996.

....[U94] One aspect of this research is the design of data structures for multicomputer files. A new class of data structures, termed Scalable Distributed Data Structures (SDDSs) has been proposed in [LNS93] Several other SDDSs have been proposed since [LNS93a] D93] LNS94] LN94] VBWY94] [KLR96]. Their common trait is to allow for scalable distributed files that are orders of magnitude larger and faster than the traditional ones. In particular, one can create multi GByte RAM files that reside entirely in the distributed RAM of the multicomputer, and that provide very low access times ....

....cost is two messages, regardless of the file size. The worst access performance of an insert or search corresponds to the case of two hops. These costs for LH are of four messages, also regardless of the number of nodes of the file. The basic principles of LH led to many variants [LNS93a] [KLR96]. There are several ways to perform the load control to increase the average load factor over the basic value of about 70 . There are also a number of ways to perform a split and to organize the interior of the bucket. One may also design LH schemes without the coordinator. Instead, a ....

[Article contains additional citation context not shown here]

Karlson, J., Litwin, W., Risch, T. LH*LH: A scalable High Performance Data Structure for Switched Multicomputers. Extending Database Technology, EDBT96. Springer Verlag, to app.

....RAM storage 3 . The first SDDS proposed and becoming well known, was LH [LNS93] LNS96] K98] R98] Today, several SDDSs are known. They provide hash based, ordered, or multi attribute distributed access orders of magnitude faster than to traditional files [D93] KW94] LNS94] VBWY94] [KLR96], LN96] TZK96] SDDS files may also be much larger than centralized files. It is well known that some applications require the high availability (fault tolerant) schemes, delivering data despite site failure. Many static high availability schemes are known, some for multicomputers [T95a] ....

Karlsson, J. Litwin, W. Risch, T. LH*lh: A Scalable High Performance Data Structure for Switched Multicomputers. Intl. Conf. on Extending Database Technology, EDBT-96, Avignon, March 1996.

....records, if any. The client terminates when it has received, in any order, m = 0 and m = 1 etc. up to m = 2 i n, where i = min (j m ) and n = min (m) with j m = i. This guarantees that all the buckets replied. The principles of LH lead to many variants with performance tradeoffs [LNS96] [KLR96]. There are several ways to perform load control, increasing the average load factor over the basic value of 70 . There are also many ways to perform a split and to organize the bucket interior. One may also design LH schemes without a coordinator [LNS93a] 2.2 The LH sa file structure An ....

....few providing statically the n availability with n 1. There were also numerous schemes targeting the scalability without high availability. These are most of the SDDS schemes referred to in this paper, as well as in [SDDS] aimed at horizontal scalability, at network and switched multicomputers, [KLR96], as well and some others, e.g. SPW90] and [JK93] among the earliest, targeting multiprocessor supercomputers. Other proposals seek vertical scalability, over 39 a fixed and rather small size multicomputer, e,g, 16 nodes, B al95] FBW97] M97a] and [P97] Some of the implementations ....

Karlsson, J. Litwin, W., Risch, T. LH*lh: A Scalable High Performance Data Structure for Switched Multicomputers. Intl. Conf. on Extending Database Technology, EDBT-96, Avignon, March 1996.

....were a minute for the CPU, then every disk access would make the application idle for eight days. More prosaically, a simple aggregate function over a typical 4 GByte disk file with the current I O speed of a few Mb s easily takes hours. 2 other followed, D93] KW94] LNS94] VBWY94] [KLR96], LN96] TZK96] K98] R98] They allow for hash based, ordered, or multi attribute distributed files, much faster and larger than could be the more traditional ones. They also allow for efficient parallel scans of these files. Some applications require high availability (fault tolerant) ....

Karlsson, J. Litwin, W. Risch, T. LH*lh: A Scalable High Performance Data Structure for Switched Multicomputers. Intl. Conf. on Extending Database Technology, EDBT-96, Avignon, March 1996.

....and parallel distributed scans with function (query) shipping. They can be hash partitioned, or ordered with respect to the primary key or support multikey access. Among the SDDS studied [SDDS] probably the best known is the distributed version of Linear Hashing [L80] called LH , LNS96] [KLR96], B99a] K98] R98] SDDS] An SDDS file is manipulated by the SDDS client sites. Each client has its own addressing schema called image that it uses to access the correct server where the record should be. As the existing buckets fill up, the SDDS splits them into new buckets. The clients ....

J. Karlson, W. Litwin, T. Risch. LH*LH: A Scalable High Performance Data Structure for Switched Multicomputers. Extending Database Technology, EDBT96, Springer Verlag.

....give a short introduction to the key concepts of Scalable Distributed Datastructures and the Monet database system. 2.1 SDDSs Scalable Distributed Data Structures (SDDSs) LNS93] can be classified as an general access path mechanism. Several SDDSs has been defined [Dev93] WBW94] KW94] LNS94] KLR96] SDDSs allow storage of a very large number of tuples distributed over any number of nodes. The tuples are distributed to different nodes according to their key value and the state of the SDDS. The primary means for retrieval is again the key value. The objective of SDDSs is to minimize the ....

Jonas S Karlsson, Witold Litwin, and Tore Risch. LH*lh: A Scalable High Performance Data Structure for Switched Multicomputers. In Advances in Database Technology --- EDBT'96, pages 573--591, Avignon, France, March 1996. Springer. References 13

No context found.

Karlsson, J. Litwin, W., Risch, T. LH*lh: A Scalable High Performance Data Structure for Switched Multicomputers. Intl. Conf. on Extending Database Technology, EDBT-96, Avignon, March 1996.

....would back trace the path the client s request was forwarded, updating these servers. The benefit of updating these servers is their decreased load by fewer mal addressed client requests. 3 A sibling of a bucket is another bucket with the same parent. 4 Server Splitting Unlike LH [13] LH lh [7] [2] and other hashbased data structures, our server splitting performs well also for splitting skewed data. LH decides the split by a linearization of which buckets to create, which easily creates problems with unskewed data. Among the non hashed based SDDSs; RP [14] for instance, uses a ....

Jonas S Karlsson, Witold Litwin, and Tore Risch. LH*lh: A Scalable High Performance Data Structure for Switched Multicomputers. In Advances in Database Technology --- EDBT'96, pages 573--591, Avignon, France, March 1996. Springer.

....to estimate the worst case access time of an unpopulated hash table, as the access time of a hash table depends on the hash function, on the overflow management policies used, and on reorganization policies. We are investigating how the use of linear hashing [12] 13]and parallel execution [9] can alleviate some of these problems. 4 Performance Polymorphism In time critical situations it is sometimes possible to define simplified algorithms that are to be executed when the deadline approaches. These simplified algorithms have timeliness trade offs regarding data consistency, ....

Karlsson J., Litwin W., Risch T.: LH*lh : A Scalable High Performance Data Structure for Switched Multicomputers. The 5th International Conference on Extending Database Technology (EDBT'96), Avignon, France, March 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