Scheuermann P., Weikum, G., Zabback, P. "Data Partitioning and Load Balancing in Parallel Disk Systems", VLDB Journal 7(1):48-66, 1998.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....algorithms that determine object placement, and thus the performance, are crucial to the success of a policymanaged storage system. Object placement techniques for large storage systems have been extensively studied in the last decade, most notably in the context of disks arrays such as RAID [4, 8, 9, 10, 21]. Most of these approaches are based on striping a technique that interleaves the placement of objects onto disks and can be classified into two fundamentally different categories. Techniques in the first category require a priori knowledge of the workload and use either analytical or ....

....of objects onto disks and can be classified into two fundamentally different categories. Techniques in the first category require a priori knowledge of the workload and use either analytical or empirically derived models to determine an optimal placement of objects onto the storage system [4, 8, 21]. An optimal placement is one that balances the load across disks, minimizes the response time of individual requests and maximizes the throughput of the system. Since requests accessing independent stores can interfere with one another, these placement techniques often employ narrow ....

[Article contains additional citation context not shown here]

P. Scheuermann, G. Weikum, and P. Zabback. Data Partitioning and Load Balancing in Parallel Disk Systems. VLDB Journal, 7(1):48--66, 1998.

....loop in section 5.2.3 in its management of compute and network resources. Existing solutions to the file assignment problem [DF82, Wol89] use heuristic optimization models to assign files to disks to get improvements in I O response times. The work described on file allocation in [GWS91, SWZ98] will automatically determine an optimal stripe width for files, and stripe those files over a set of homogeneous disks. They then balance the load on those files based on a form of hotspot analysis, and swapping file blocks between hot and cold disks. Hippodrome can expand or contract the ....

Peter Scheuermann, Gerhard Weikum, and Peter Zabback. Data partitioning and load balancing in parallel disk systems. VLDB Journal: Very Large Data Bases, 7(1):48--66, 1998.

....Chen, Rotem, and Seshadri investigated in [6] the problem of adapting existing declustering methods to work in heterogeneous environments. In [7] Christodoulakis and Zioga investigated design principle of placing striped delay sensitive data on a number of disks in a distributed environment. In [22] Scheuermann, Weikum and Zabback studied striping and load balancing techniques in parallel disk systems and showed their relationship to response time and throughput. Also the 20 40 60 80 100 120 140 Flat Global Indices Partitioned (min) Join where mard.matnr = marc.matnr and mard.werks ....

P. Scheuermann, G. Weikum, and P. Zabback. Data partitioning and load balancing in parallel disk systems. VLDB Journal, 7(1):48--66, 1998.

....of the system to implement the desired policies. In addition to an implementation of adaptive caching, they also include a system for adaptive 28 disk striping. This scheme alters striping based on previous work indicating that optimal striping units are dependent on resource contention [7, 46]. While the work performed by this group shares many characteristics with the work presented here, there are many significant differences. First, the approach used in PPFS II is a centralized one, where sensor data is gathered at one point and used to make global decisions. Our system, on the ....

Peter Scheuermann, Gerhard Weikum, and Peter Zabback. Data partitioning and load balancing in parallel disk systems. Technical Report 209, ETH Zurich, January 1994.

....4.1 where we show that Ergastulum is faster and generates as good or better designs. Existing solutions to the file assignment problem [11, 35] use heuristic optimization models to assign files to disks to get improvements in I O response times. The file allocation schemes described 23 in [12, 24] automatically determine an optimal stripe width for files, and stripe those files over a set of homogeneous disks. They then balance the load on those files based on a form of hotspot analysis, and swapping file blocks between hot and cold disks. Ergastulum can select the appropriate number ....

P. Scheuermann, G. Weikum, and P. Zabback. Data partitioning and load balancing in parallel disk systems. VLDB Journal: Very Large Data Bases, 7(1):48--66, 1998.

....on an economic model that factors in the tradeoffs between the service quality and the cost. Existing solutions to the file assignment problem [13, 34] use heuristic optimization models to assign files to disks to get improvements in I O response times. The file allocation schemes described in [16, 28] will automatically determine an optimal stripe width for files, and stripe those files over a set of homogeneous disks. They then balance the load on those files based on a form of hotspot analysis, and swapping file blocks between hot and cold disks. Hippodrome can expand or contract the ....

P. Scheuermann, G. Weikum, and P. Zabback. Data partitioning and load balancing in parallel disk systems. VLDB Journal: Very Large Data Bases, 7(1):48--66, 1998.

....clustering indexes in IBM s DB2. 7] described restartable algorithms for online construction of an index. In parallel database systems, new challenges in online reoganization arise as data and indexes are partitioned across multiple disks, and load imbalance occurs when access patterns change. [12, 13] presented various file striping heuristics for data allocation, data redistribution and load balancing in a shared memory multiprocessing environment. 16] showed 1 how records can be distributed into variable sized fragments and migrated when load imbalance occurs in a shared nothing system. ....

P. Scheuermann, G. Weikum, and P. Zabback. Data partitioning and load balancing in parallel disk systems. VLDB Journal, 7(1), 1998.

....it assumed that requests are distributed equally across the disks. High arrival rates, access hot spots, shared le access on parallel systems, and variable size requests all a ect performance, which makes choosing an appropriate data distribution dependent on a host of interrelated factors. In [18], Scheuermann, Weikum, and Zabback present an analytic model for striping on parallel disk systems which is very similar to the models that we discuss in #4. Our striping models di er by making more simplifying assumptions on service time distributions and considering the case of networks of ....

....and Zabback present an analytic model for striping on parallel disk systems which is very similar to the models that we discuss in #4. Our striping models di er by making more simplifying assumptions on service time distributions and considering the case of networks of servers and clients. In [18], only shared memory multiprocessors are considered and network latencies are ignored. This work points out the importance of le speci c striping tuning even in a shared memory architecture. Redundant storage. Several researchers have observed that disk areal densities are increasing much faster ....

P. Scheuermann, G. Weikum, and P. Zabback, \Data Partitioning and Load Balancing in Parallel Disk Systems," The VLDB Journal, vol. 7, pp. 48-66, Feb. 1998.

....of all the items in the cache are reduced at regular intervals of time, which in turn will #cool off# an object that is no longer required by any MU and hence remove it from the cache. This approach has been very widely used in solving data allocation problems for caches, disks etc. 12] and [9]) 3.3. Design of the server The challenge in designing a broadcast strategy for the server is to allocate data items in such a way that hot items are broadcast more frequently than the less hot data items. Let n be the number of channels in the MLMC Air Cache. We de Thetane a Channel Premium ....

P. Scheuermann, G. Weikum, and P. Zabback. Data partitioning and load balancing in parallel disk systems. The VLDB Journal, 7(1):48#66, 1998.

.... representing the core memory (at first level) the secondary storage (e.g. magnetic disks) and the tertiary storage (e.g. optical jukeboxes, tape drive services) The boxes above the Multimedia Storage Modules representkey functions of the HERMES system architecture like data placement (cf [25], 8] 9] access methods (cf [3] 10] caching [20] delay sensitivedata scheduling [24] Quality of Service (QofS) and metadata management [2] Although these functions may not be clearly separated from other functions inside the Multimedia Storage they are represented as separate boxes to ....

P.Scheuermann, G. Weikum, and P. Zabback. Data partitioning and load balancing in parallel disk systems. VLDB Journal, 7(1):48--66, 1998.

....servers; i.e. the situations in which the Reference Counting policy performs poorly. Data placement has of course also been studied in the context of parallel database systems #Copeland et al. 1988#. A dynamic approach to adjust the data placement in a parallel disk system was proposed in #Scheuermann et al. 1998#. Again, none of these approaches integrate data placement and query optimization nor do they consider the execution of sub optimal plans in order to improve the data placement. Furthermore, the performance tradeo#s of parallel and distributed systems are di#erent #consider, for example, ....

Scheuermann, P., Weikum, G., and Zabback, P. 1998. Data partitioning and load balancing in parallel disk systems. The VLDB Journal 7,1#Feb.#, 48#66.

....of all the items in the cache are reduced at regular intervals of time, which in turn will cool off an object that is no longer required by any MU and hence remove it from the cache. This approach has been very widely used in solving data allocation problems for caches, disks etc. 12] and [9]) 3.3. Design of the server The challenge in designing a broadcast strategy for the server is to allocate data items in such a way that hot items are broadcast more frequently than the less hot data items. Let n be the number of channels in the MLMC Air Cache. We de fine a Channel Premium ....

P. Scheuermann, G. Weikum, and P. Zabback. Data partitioning and load balancing in parallel disk systems. The VLDB Journal, 7(1):48--66, 1998.

....allows estimating the caching benefit of an object, or and holding new objects in in memory buffer for selection [RS98b] Disk load balancing has been studied mostly in the context of database and file system research. Several papers investigated file partitioning between disks in parallel systems [LKB87, SWZ96] and in video servers [GVKR95] These load balancing methods were designed to balance read requests between disks when reading one or a few large files (declustering striping methods) or to balance the use of disks over the long time period (heat methods [CABK88, KH93] However, even if the disk ....

P. Scheuermann, G. Weikum, and P. Zabback. Data Partitioning and Load Balancing in Parallel Disk Systems. Technical Report A/02/96, Department of Computer Science, University of the Saarland, April 1996.

....in data management systems. To achieve higher efficiency of data access from disks, data is often clustered according to its expected use. Recent research on data allocation and placement based on application access patterns in the context of parallel disks and RAID technology is reported in [15]. However, the problem is much more acute Access of Multi Dimensional Datasets on Tertiary Storage Systems 25 when dealing with robotic tertiary systems, and the solutions are different. We have chosen to work closely with a specific application area (climate modeling) where this problem is ....

Scheuermann, P., Weikum, G., and Zabback, P. Data Partitioning and Load Balancing in Parallel Disk Systems, Department Informatik, ETH Zurich, January 1994, Technical Report number 209.

No context found.

P. Scheuermann, G. Weikum, and P. Zabback. Data partitioning and load balancing in parallel disk systems. VLDB Journal, 1998.

.... striping unit is chosen globally [4] This approach is suitable for scientific applications or pure on line transaction processing, in which all files have approximately the same sort of access characteristics (i.e. only large requests or only single block requests) However, as we have shown in [17] for many applications which exhibit highly diverse file access characteristics (e.g. VLSI design, desktop publishing, etc. it is desirable to tune the striping unit individually for each file. We have developed an analytic model for an open queueing system in order to determine heuristically ....

....(e.g. VLSI design, desktop publishing, etc. it is desirable to tune the striping unit individually for each file. We have developed an analytic model for an open queueing system in order to determine heuristically the optimal striping unit on an individual file basis or on a global basis [17, 20, 22]. We observe here that an open queueing model is much more realistic for an environment where a large number of users issue requests to a parallel disk system, as compared to the closed queueing model used in [3, 12] where the number of concurrent I O requests in the system is fixed. In our ....

[Article contains additional citation context not shown here]

Scheuermann, P., Weikum, G., and Zabback, P., "Data Partitioning and Load Balancing in Parallel Disk Systems," Technical Report 209, Department of Computer Science, ETH Zurich, January 1994.

....is evaluated off line and can efficiently drive the system s run time decisions. We are not aware of any previous work that pursues an ap proach to mixed workloads along these lines. A prototype system based on this new approach is being built by extend ing the FIVE experimental file system [29, 30, 32]. The rest of the paper is organized as follows. Section 2 in troduces our system architecture. Section 3 develops an ap proach towards a performance prediction model, for both continuous data and discrete data requests. Finally, Section 4 discusses the admission control and the adaptation ....

....modified after their initial insertion. Discrete objects are allocated to disk such that the expected I O load on behalf of this data is balanced across the disks; this involves coarse grained striping for large objects and simple but effective load balancing heuristics along the lines of [30]. In what follows we do, however, not rely on any specific assumptions on the storage layout for discrete objects. C data and D data both reside on the same shared disk pool, as this provides a much better resource utilization than a partitioned scheme with dedicated disks, from both a disk ....

[Article contains additional citation context not shown here]

Peter Scheuermann, Gerhard Weikum, Peter Zabback, Data Partitioning and Load Balancing in Parallel Disk Systems, Technical Report A/02/96, Department of Computer Science, University of the Saarland, 1996, submitted for publication.

....patterns, horizontal data migration between devices of the same storage level may be needed, for example, for dynamic and incremental load balancing among disks. As Markov chain models are substantially richer than the stationary probability models that have been explored for these purposes [SWZ98], further performance improvements may be possible. Prefetching and caching for Web servers Within the Internet and large intranets, proxy servers are often used for the caching of documents that are likely to be accessed by the local clients (see, e.g. Be96] One reason for the caching of ....

Scheuermann P, Weikum G, Zabback P (1998) Data Partitioning and Load Balancing in Parallel-Disk Systems. VLDB J 7(1): 48--60

....patterns, horizontal data migra tion between devices of the same storage level may be needed, for example, for dynamic and incre mental load balancing among disks. As Markov chain models are substantially richer than the stationary probability models that have been explored for these purposes [SWZ98], further perfor mance improvements may be possible. # Prefetching and Caching for Web Servers: Within the Internet and large intranets, proxy servers are often used for the caching of documents that are likely to be accessed by the local clients (see, e.g. Be96] One reason for the ....

Scheuermann, P., Weikum, G., Zabback, P.: Data Partitioning and Load Balancing in Paral- lel Disk Systems, to appear in: VLDB Journal, 47

....4, the approach is validated by comparing the analytic results to simulation studies. Section 5 discusses some practical system issues, and Section 6 gives an outlook on future work. A prototype system based on this new approach is being built by extending the FIVE experimental file system [SWZ94, SWZ96]. 2 System Architecture We assume that clients submit requests for continuous data to the server. Continuous data objects like videos, audios, or animations are composed of sequences of fragments and constitute data streams that are consumed by the client in a time constrained way according to ....

....[Her96] The prototype is able to handle multimedia data of different types, with different bandwidth requirements, and also variable bandwidth within an object, as required by MPEG 2. The architecture of our server in terms of data placement and load balancing is based on the FIVE prototype [SWZ94, SWZ96], an experimental file system for parallel disk systems. We are currently extending FIVE to support the presented admission control. 6 Future Work Except for requiring a minimal buffer on the client site for incoming fragments, we have disregarded buffering issues so far. In the advanced ....

Peter Scheuermann, Gerhard Weikum, Peter Zabback, Data Partitioning and Load Balancing in Parallel Disk Systems, Technical Report A/02/96, Department of Computer Science, University of the Saarland, 1996, submitted for publication.

No context found.

Scheuermann P., Weikum, G., Zabback, P. "Data Partitioning and Load Balancing in Parallel Disk Systems", VLDB Journal 7(1):48-66, 1998.

No context found.

P. Scheuermann, G. Weikum, and P. Zabback. Data partitioning and load balancing in parallel disk systems. VLDB Journal, 7(1):48--66, 1998.

No context found.

P. Scheuermann, G. Weikum, and P. Zabback. Data partitioning and load balancing in parallel disk systems. VLDB Journal, 7(1), 1998.

No context found.

Scheuermann, P., Weikum, G., Zabback, P.: Data partitioning and load balancing in parallel disk systems. The VLDB Journal (1998) 48--66

No context found.

P. Scheuermann, G. Weikum, and P. Zabback. Data partitioning and load balancing in parallel disk systems. VLDB Journal: Very Large Data Bases, 7(1):48-66, 1998.

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