S. Chakrabarti and S. Muthukrishnan. Resource scheduling for parallel database and scientific applications. In Proceedings of the 8th Annual ACM Symposium on Parallel Algorithms and Architectures, pages 329--335, Padua, Italy, 1996.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....P F period: 3 p optimal use of resources good use of resources wasted resources F: flushing P: processing L: loading Figure 8: Multi threaded execution of index builder phases that will result in minimal overall running time. Our problem di#ers from a typical job scheduling problem [12] in that we can vary the sizes of the incoming jobs, i.e. in every loading phase we can choose the number of pages to load. Consider an index builder that uses N executions of the pipeline to process the entire collection of pages and generate N sorted runs. By an execution of the pipeline, we ....

S. Chakrabarti and S. Muthukrishnan. Resource scheduling for parallel database and scientific applications. In 8th ACM Symposium on Parallel Algorithms and Architectures, pages 329--335, June 1996.

....six sorted runs on disk. The goal of our pipelining technique is to design an execution schedule for the di erent indexing phases that will result in minimal overall running time (also called makespan in the scheduling literature) Our problem di ers from a typical job scheduling problem [4] in that we can vary the sizes of the incoming jobs, i.e. in every loading phase we can choose the number of pages to load. In the 2 The URLs are replaced by numeric identi ers for compactness 5 thread1 thread2 thread3 indexing time 0 L P F L P F L P F L P F L P F L P F ....

....such as global statistics collection and optimization of the indexing process on each individual node. Our technique for structuring the core index engine as a pipeline has much in common with pipelined query execution strategies employed in relational database systems [7] Chakrabarti, et al. [4] present a variety of algorithms for resource scheduling with applications to scheduling pipeline stages. There has been prior work on using relational or object oriented data stores to manage and process inverted les [1, 3, 8] Brown, et al. 3] describe the architecture and performance of an ....

S. Chakrabarti and S. Muthukrishnan. Resource scheduling for parallel database and scientic applications. In 8th ACM Symposium on Parallel Algorithms and Architectures, pages 329-335, June 1996.

....and generates six sorted runs on disk. The key issue in pipelining is to design an execution schedule for the di#erent indexing phases that will result in minimal overall running time (also called makespan in the scheduling literature) Our problem di#ers from a typical job scheduling problem [2] in that we can vary the sizes of the incoming jobs, i.e. in every loading phase we can choose the number of pages to load. In the rest of this section, we describe how we make e#ective use of this flexibility. First, we derive, under certain simplifying assumptions, the characteristics of an ....

....issues such as global statistics collection and optimization of the indexing process on each individual node. Our technique for structuring the index engine as a pipeline has much in common with pipelined query execution strategies employed in relational database systems [5] Chakrabarti, et al. [2] present a variety of algorithms for resource scheduling with applications to scheduling pipeline stages. There has been prior work on using relational or objectoriented data stores to manage and process inverted files [1, 11 A billion pages will contain roughly 310 million distinct terms [13] ....

S. Chakrabarti and S. Muthukrishnan. Resource scheduling for parallel database and scientific applications. In 8th ACM Symp. on Parallel Alg. and Architectures, pages 329--335, June 1996.

....and generates six sorted runs on disk. The key issue in pipelining is to design an execution schedule for the di erent indexing phases that will result in minimal overall running time (also called ######## in the scheduling literature) Our problem di ers from a typical ### ########## problem [2] in that we can vary the sizes of the incoming ####, i.e. in every loading phase we can choose the number of pages to load. In the rest of this section, wedescribe howwemake e ective use of this exibility. First, we derive, under certain simplifying assumptions, the characteristics of an ....

....issues suchas global statistics collection and optimization of the indexing process on each individual node. Our technique for structuring the index engine as a pipeline has much in common with pipelined query execution strategies employed in relational database systems [5] Chakrabarti, et al. [2] presentavariety of algorithms for resource scheduling with applications to scheduling pipeline stages. There has been prior work on using relational or objectoriented data stores to manage and process inverted les [1, ## A billion pages will contain roughly 310 million distinct terms [13] and ....

S. Chakrabarti and S. Muthukrishnan. Resource scheduling for parallel database and scientic applications. In ### ### ##### ## ######## #### ### ###### ########, pages 329-335, June 1996.

....for the proposed on line schedulers in Section 5 demonstrate a smooth, linear transition of the competitive ratio from the case of unit execution times to unrelated execution times that is governed by the runtime ratio. The importance of this parameter has also been demonstrated recently in [CM96] for off line scheduling of jobs with multiple resource demands, both malleable (allow for virtualization with proportional slowdown) and non malleable. Although we are interested in on line scheduling, it might be appropriate to briefly mention some complexity results for the corresponding ....

Soumen Chakrabarti and S. Muthukrishnan. Resource Scheduling for parallel database and scientific applications. In Proceedings of the 8th Annual ACM Symposium on Parallel Algorithms and Architectures, SPAA '96 (Padua, Italy, June 24--26, 1996), pages 329--335, New York, 1996. ACM SIGACT, ACM SIGARCH, ACM Press.

....and size 1, into 3 slots. There exists a placement of these ads such that each time slot contains 2 adds, but for the corresponding 2 D bin packing problem, the optimal placement has height 3. Many authors have considered the problem of scheduling parallel jobs on a set of processors; see, e.g. [CM96, GG75, FKST93, HSW96, LT94, SWW91, TSWY94, TWY92, Sch96]. If we interpret the size of an ad s i , as the number of processor required by a parallel job, and w i Delta T as the time required by the same job, then the ad placement problem can be interpreted as the scheduling of independent, non malleable jobs for identical processors, where both ....

S. Chakrabarti and S. Muthukrishnan. Resource scheduling for parallel database and scientific applications. In Proc. 8th ACM Symp. on Parallel Algorithms and Architectures, pages 329 -- 335, June 1996.

.... time (or, equivalently, the average weighted completion time) has recently achieved a great deal of attention, partly because of its importance as a fundamental problem in scheduling, and also because of new applications, for instance, in compiler optimization [5] or in parallel computing [3]. There has been significant progress in the design of approximation algorithms for this kind of problems which led to the development of the first constant worst case bounds in a number of settings. This progress essentially follows on the one hand from the use of preemptive schedules to ....

S. CHAKRABARTI AND S. MUTHUKRISHNAN, Resource scheduling for parallel database and scientific applications, in Proceedings of the 8th Annual ACM Symposium on Parallel Algorithms and Architectures, June 1996, pp. 329 -- 335.

....algorithm must ensure that the inter media synchronization constraints defined by the temporal relationships among CM components are met. Handling these synchronization constraints requires a task model that is significantly more complex than the models employed in scheduling theory and practice [CM96, GG75, GGJY76] More specifically, composite multimedia objects essentially correspond to resource constrained tasks with timevarying resource demands. Resource constraints come from the limited amount of server resources available to satisfy the requirements of CM streams and time variability ....

....resource requirements of the presentation over time with their enclosing Minimum Bounding Rectangle (MBR) Although this simplification significantly reduces the complexity of the relevant scheduling problems, it suffers from two major deficiencies. ffl The volume (i.e. resource time product [CM96, GI97] in the enclosing MBR can be significantly larger than the actual requirements of the composite object. This can result in wasting large fractions of precious server resources, especially for relatively sparse composite objects. ffl The MBR simplification hides the timing structure of ....

[Article contains additional citation context not shown here]

Soumen Chakrabarti and S. Muthukrishnan. "Resource Scheduling for Parallel Database and Scientific Applications ". In Proceedings of the Eighth Annual ACM Symposium on Parallel Algorithms and Architectures, pages 329--335, Padua, Italy, June 1996.

....operations that impose a significant load on multiple system resources. With respect to ss resources, all previous work has concentrated on simplified models, assuming that the capacity of all such resources is infinite or that they are all globally accessible to all tasks [GG75, ST94, NSHL95, CM96] Clearly, such models do not account for the physical distribution of resource units or the possibilities of ss resource fragmentation. This limits the usefulness of these models to a shared everything [DG92] context. In this paper, we present a general framework for multi dimensional ts and ss ....

....on the suboptimality of the solution. However, scheduling query plans on shared nothing or hierarchical architectures requires a significantly richer model of parallelization than what is assumed in the classical [Gra66, GG75, GLLRK79] or even more recent [BB90, BB91, KM92, TWY92, WC92, ST94, CM96] efforts in that field. To the best of our knowledge, there have been no theoretical results in the literature on parallel task scheduling that consider multiple ts system resources and explore sharing of such resources among concurrent tasks, or study the implications of pipelined parallelism ....

[Article contains additional citation context not shown here]

Soumen Chakrabarti and S. Muthukrishnan. "Resource Scheduling for Parallel Database and Scientific Applications". In Proceedings of the Eighth Annual ACM Symposium on Parallel Algorithms and Architectures, pages 329--335, Padua, Italy, June 1996.

.... (or, equivalently, the average weighted completion time) has recently achieved a great deal of attention, partly because of its importance as a fundamental problem in scheduling, and also because of new applications, for instance, in compiler optimization [CJM 96] or in parallel computing [CM96] In the last two years, there has been significant progress in the design of approximation algorithms for this kind of problems which led to the development of the first constant worst case bounds in a number of settings. This progress essentially follows on the one hand from the use of ....

S. Chakrabarti and S. Muthukrishnan. Resource scheduling for parallel database and scientific applications. June 1996. Proceedings of the 8th ACM Symposium on Parallel Algorithms and Architectures.

....2 and size 1, into 3 slots. There exists a placement of these ads such that each time slot contains 2 adds, but for the corresponding 2 D bin packing problem, the optimal placement has height 3. Many authors have considered the problem of scheduling parallel jobs on a set of processors; see, e.g. [CM96, GG75, FKST93, HSW96, LT94, SWW91, TSWY94, TWY92, Sch96]. If we interpret the size of an ad s i , as the number of processor required by a parallel job, and w i Delta T as the time required by the same job, then the ad placement problem can be interpreted as the scheduling of independent, non malleable jobs for identical processors, where both ....

S. Chakrabarti and S. Muthukrishnan. Resource scheduling for parallel database and scientific applications. In Proc. 8th ACM Symp. on Parallel Algorithms and Architectures, pages 329 -- 335, June 1996.

....as well as the SS allotment for op i and our model for communication costs. Given an operator clone with a (stand alone) execution time of T and a SS demand of V , we define the volume vector of the clone as the product T Delta V , i.e. the resource time product 2 for the clone s execution [3]. S W , S V , and S TV are used to denote the set of work, demand, and volume vectors (respectively) for the set S of all the clones to be scheduled. We use the W , V , and TV superscripts in this manner throughout the paper. The length of a n dimensional vector v is its maximum component. ....

S. Chakrabarti and S. Muthukrishnan. "Resource Scheduling for Parallel Database and Scientific Applications". In Proc. of the 8th ACM Symp. on Parallel Algorithms and Architectures, June 1996.

....in which giving more of a resource (say, more processors) improves performance. Therefore, our problem is a version of a multidimensional malleable resource problem. Prior work involves malleable scheduling of parallelizable jobs [7, 9] or scheduling a mixture of malleable and non malleable jobs [19, 8], but none addresses our problem. For example, in our case, the processing time of a job depends on two variables, the number of codes and the assigned power per slot; both these resources are malleable, and they have a strongly (nonlinearly) coupled effect on data rate through Equations ....

S. Chakrabarti and S. Muthukrishnan, "Resource scheduling for parallel database and scientific applications," in In Proc. of the ACM Symposium on Parallel Algorithms and Architectures, 1996, pp. 329--335.

....25, 8] A version closely related to ours is the one in which there is a bound on the number of jobs that may be in the shop at any time [20] our formulation is more general. Although many results exist on scheduling parallel processors and database queries with additional resource constraints [7, 2, 5], flow shop problems with limited storage and resource constraints have not been thoroughly studied before. Also, our problem may be thought of as a special packet routing problem on a network of three nodes with a bounded queue at each node [14] no relevant results exist for our special ....

....general problem may be thought of as 3 machine flow shop scheduling with limited storage and additional resource constraints. This problem is clearly NP hard. In general, scheduling problems with additional resource constraints do not have small constant factor approximations (see for example, [2, 7]) We propose the simple algorithmic approach of solving the P = D = T = 1 version of the problem and simulating the order of job execution in parallel for the general case. Although this approach of simulating the sequential schedule with multiple machines is a provably good approximation in some ....

Chakrabarti, S., and Muthukrishnan, S. Resource scheduling for parallel database and scientific applications. In Proceedings of the Eighth Annual ACM Symposium on Parallel Algorithms and Architectures (SPAA) (Padua, Italy, June 1996), pp. 329--335.

....constraints, the problem appears to be very different from most 7 precedence constrained scheduling results that are based on basic lower bounds like average work per processor or the critical path length in the precedence graph. We obtain an O(log T ) approximation for both WACT and makespan [28], where T is the longest to shortest job time ratio. Our algorithm uses a technique that deliberately introduces delays to improve on a greedy schedule. We also show that the log T blowup is unavoidable for certain instances. Since our models were carefully abstracted from real life scheduling ....

.... Non malleable Precedence On line Makespan procs type processor processor OE off line WACT Garey et al. [57] N A N A Theta p ; on makespan Feldmann et al. [51] m par p Theta any on makespan Hall et al. [70] 1; m seq Theta Theta any off both 1; m seq Theta Theta ; on both This chapter [29, 28] m par p Theta any on both m par Theta p ; on both m par p p any on makespan m par p p any off WACT m par p p forest SPG on both Table 4.1: Comparison of results. N A=not applicable, seq=sequential, par=parallel, SPG=seriesparallel graph. A SPG is expressed recursively as follows: ....

S. Chakrabarti and S. Muthukrishnan. Resource scheduling for parallel database and scientific applications. In Symposium on Parallel Algorithms and Architectures (SPAA), Italy, June 1996. ACM.

....randomized on line algorithm with expected performance within 5:78 of optimal. 4.3 Perfectly malleable jobs with tree precedence We shall consider perfectly malleable jobs, as in Feldmann et al. [6] and out tree precedence constraints. A study of precedence and non malleability is initiated in [5]. Our DualPack routine is as follows. We remove from J any j with PathToRoot(j) D, set J 0 = Knapsack(J ; mD) and list schedule J 0 as in [6] let OE = p 5 Gamma 1) 2 be the golden ratio; whenever there is a job j with all of its predecessors completed and the number of busy ....

S. Chakrabarti and S. Muthukrishnan. Resource scheduling for parallel database and scientific applications. To appear in SPAA 96, June 1996.

No context found.

S. Chakrabarti and S. Muthukrishnan. Resource scheduling for parallel database and scientific applications. In Proceedings of the 8th Annual ACM Symposium on Parallel Algorithms and Architectures, pages 329--335, Padua, Italy, 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