Waqar Hasan and Rajeev Motwani. Optimization algorithms for exploiting the parallelism-communication tradeoff in pipelined parallelism. In Int. Conf. on Very Large Databases, pages 36--47, Santiago, Chile, 1994.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....in [1] where the execution time of a pipeline participant is observed as a function of the number of processors assigned to it. However, for large join queries, it is rather expected that multiple processes compete for the same processor. Pipelined and bushy parallelism [3] is studied in [1, 4, 2, 6], where the communication across logical links between interacting processes is also considered. However, the network configuration affects the selection of the physical channels to connect two communicating processors. A router may even dynamically choose different routes between the same ....

....are operators annotated with execution algorithm and data access information. The children of a node x are its producers ; x is the consumer . A producer is blocking if it must complete execution before its consumer starts, or pipelining if each tuple it outputs is immediately consumed [4]. In a bushy QEP, a node may have one pipeline and one blocking producer, e.g. a hash join, or two pipeline producers, e.g. a merge join on two sorted inputs. A schedule represents a node to processor mapping for the nodes of a QEP. We do not consider node parallelism. Hence, a node is assigned ....

Waqar Hasan and Rajeev Motwani. Optimization algorithms for exploiting the parallelism-communication tradeoff in pipelined parallelism. In Int. Conf. on Very Large Databases, pages 36--47, Santiago, Chile, 1994.

.... of the most important issues that must be considered is the parallelism communication trade off [5, 6] To consider this, the query to be scheduled is represented as a weighted operator tree in which each node represents an operator and each edge represents the timing constraints between operators [13, 11]. A timing constraint is either a precedence or parallel constraint. The parallel constraint requires that the two adjacent nodes behave as a producer consumer system where the producer sends a long stream of communication data to the consumer. These two nodes, therefore, start and terminate their ....

....system where the producer sends a long stream of communication data to the consumer. These two nodes, therefore, start and terminate their works approximately at the same time. A weighted operator tree in which all edges represent parallel constraints is called a Pipelined Operator Tree (POT) [11]. The communication cost in the POT makes its scheduling different from the classical scheduling problems. Since that data is transmitted in long streams, the important aspect of communication cost is the CPU overhead of sending receiving messages and not the delay for signal propagation. ....

[Article contains additional citation context not shown here]

Hasan, W., and Motwani, R., "Optimization Algorithms for Exploiting the Parallelism, Communication Tradeoff in Pipelined Parallelism," In Proceedings of 20th International Conference on VLDB, pp:36-47, Santigo, Chile, September 1994.

....plan is first constructed for a uniprocessor machine and then parallelized. In one step optimization, the parameters of parallelism are already taken into account when establishing the optimal plan, which thus contains scheduling information. The first approach is adopted in studies like [1] [6], 7] An optimal sequential plan is produced at compile time, and an optimal parallelization of this plan is selected according to some heuristics at runtime. As pointed out by Lanzelotte et al. there is no guarantee that the optimal uniprocessor plan will remain optimal when parallelized ....

W. Hasan and R. Motwani, "Optimization Algorithms for Exploiting the Parallelism-Communication Tradeoff in Pipelined Parallelism," Proc. Int'l Conf. Very Large Databases, pp. 36--47, Santiago, Chile, 1994.

....under worst case conditions. But in computational intelligence, the cultural default is to look for the performance that is expected on average. 4.1 The Di erence between the Average and the Homogeneous Case Scenario We will avoid the term average case scenario, because it is confusing. In [13, 10, 4], the term is used to mean that behavior is averaged over all possible scenarios. However, in [15, 8, 16] the same word is used to denote a di erent concept, which we will here call the homogeneous case scenario. It is important to distinguish between these two. Average scenario analysis ....

....the worst case bound of this value over all possible problem instances is taken as a measure for the algorithm s quality. That way, it is essentially a worst case bound on expected behavior. We do not consider this type of analysis here. 4. 2 Comparing the Worst Case and the Average Scenario In [10], both a worst case and an average performance ratio are given for algorithms that optimize parallel plans for SQL queries. In [13] the average worstcase scenario trade o is identi ed and related to window size in an algorithm for data replication. The authors propose to choose a reasonable ....

Waqar Hasan and Rajeev Motwani. Optimization algorithms for exploiting the parallelism-communication tradeo in pipelined parallelism. In Proceedings of the 20th VLDB Conference, Santiago, Chile, 1994.

....Internet and the ensuing increasingly complex socio economic context of computation. 1 Consider the query optimization problem, for example, arguably the most important and complex problem in databases. Trade offs between parallelism and communication in query optimization have been studied in [CHM, HM], and elsewhere. The Mariposa wide area database system [SAP ] was architected to make such trade offs explicit in an advantageous way. Mariposa assumes that a subquery can be executed in many diverse database sites, and each site submits a bid for the query, specifying a delay for delivering ....

W. Hasan, R. Motwani. Optimization Algorithms for Exploiting the Parallelism-Communication Trade-off in Pipelined Parallelism. Proc. of the 20th International Conference on Very Large Data Bases (VLDB), pp. 36-47, 1994.

....Section 8. 2. Related Work Parallel query processing has been studied in a large variety of facets, see e.g. PMC 90, DG92, HS93, WFA95, Gra95] Most of related work in this eld concentrated on possibilities to speedup highly complex queries with long running times. Approaches as taken in [HM94] and [GI96] suggest a decomposition of the query plans into sub plans which are then executed in parallel on di erent nodes of the parallel processing environment. The granularity of this decomposition varies and can be as ne as parallelizing single operator as studied for example in [SD89, ....

W. Hasan and R. Motwani. Optimization Algorithms for Exploiting the ParallelismCommunication Tradeo in Pipelining Parallelism. In Proc. of the Int'l. Conf. on Very Large Data Bases, pages 36-47, Santiago, Chile, September 1994.

....sequential query evaluation plan (QEP) The latter deals with mapping a sequential QEP to a parallel execution environment. The final result is a parallel query execution plan (cf. Fig. 1) Much research has been devoted to achieve the best possible parallelization of a given sequential plan [8, 9, 13, 3, 10]. A common approach is to incorporate many features of the target architecture in the cost model, e.g. communication costs or hardware description. Based on this information a static parallel schedule is derived [8, 4] However, from a validation point of view increasing the number features ....

....achieve the best possible parallelization of a given sequential plan [8, 9, 13, 3, 10] A common approach is to incorporate many features of the target architecture in the cost model, e.g. communication costs or hardware description. Based on this information a static parallel schedule is derived [8, 4]. However, from a validation point of view increasing the number features considered during optimization is dangerous. The prime reason is that small errors in the estimates propagate through a QEP. Such estimation errors turn out to be exponential [12] and lead to suboptimal parallel schedules. ....

[Article contains additional citation context not shown here]

W. Hasan and R. Motwani. Optimization Algorithms for Exploiting the ParallelismCommunication Tradeoff in Pipelining Parallelism. In Proc. Int'l. Conf. on Very Large Data Bases, Santiago, Chile, September 1994.

....[15] In those first studies [18, 2] the query processing was based on a sequential optimizer in charge of defining an execution plan (e.g. processing order of the relational operators [19] which was parallelized at runtime. In order to reduce the query processing time, later works 1 [20, 21, 13, 22] tried then to integrate parallelization strategies in the sequential optimization process. In those approaches, the query optimizer is also in charge of studying the parallelization issues as the degree of inter parallelism (i.e. the number of relational operation to be executed in parallel) and ....

....operation to be executed in parallel) and the degree of intra parallelism (i.e. the number of processors to be attributed to each operation) Parallel query optimization Different approaches have been yet proposed for complex query optimization on a multi processor machine 2 . Some models [24, 22, 11] parallelized an optimal sequential execution plan. Other prototypes [4, 25] explored parallel execution plans which are not optimal in the sequential case, but which may run faster than the parallelized sequential optimum. In practice the search space to be seeked in order to find an optimal ....

[Article contains additional citation context not shown here]

W. Hasan and R. Motwani. Optimization Algorithms for Exploiting the Parallelism-Communication Tradeoff in Pipelined Parallelism. In Proceedings of the International Conference on Very Large Databases, pages 36--47, Santiago, Chile, September 1994.

....relations and to efficiently manage communication buffers. Such operations are very expensive. Consequently, communications should be considered with the same attention as any other operator working on relation partitions. Indeed, as it was first mentioned by Gray in 1988 [15] and later by Hasan [16, 17], optimizers that do not integrate communication costs cannot work efficiently. Two classes of approaches have been proposed in order to integrate communications into PEPs. Some works suggest to model communications as weights associated to the edges used in the D graph representation (e.g. 16, ....

....17] optimizers that do not integrate communication costs cannot work efficiently. Two classes of approaches have been proposed in order to integrate communications into PEPs. Some works suggest to model communications as weights associated to the edges used in the D graph representation (e.g. [16, 10]) Oppositely, some systems explicitly introduce communication operators. To our knowledge, only three systems have adopted this latter approach : the par LERA database language [18] with its constructor node (see section 6.1) the Volcano prototype [19] exchange operator) and the commercial ....

W. Hasan and R. Motwani. Optimization Algorithms for Exploiting the ParallelismCommunication Tradeoff in Pipelined Parallelism. In Proceedings of the International Conference on Very Large Databases, pages 36--47, Santiago, Chile, September 1994.

....to hold in parallel spaces as well. Therefore, we consider bushy QEPs in our model. We consider both forms of interoperator parallelism [Gra93] Sibling nodes subtrees can be executed simultaneously, thus supporting bushy parallelism. Edges between consecutive nodes can be blocking or pipelining [HM94] so that pipelining can be exploited. In Figure 1, we show an example query graph, and in Figure 2, we present two QEPs for this graph. Non leaf nodes are surrounded by a circle. All nodes and leaves are identified by their join, respectively relation, number. The join nodes are labelled with ....

....are kept and recombined, is satisfied. It should be stressed that the crossover does not simply change the tree structure of a QEP. The subtree transferred from one QEP to another carries with it the algorithm assignments and the settings of the edges as blocking or pipelining (terminology in [HM94] In that sense, crossover has a more complex impact on a QEP than the QEP transformations proposed in [IK91, SHC96] for combinatorial optimization techniques. Mutation. We consider two mutation operators: ffl mutate1(T ; newAlg) changes the join algorithm into newAlg for a randomly selected ....

[Article contains additional citation context not shown here]

Waqar Hasan and Rajeev Motwani. Optimization algorithms for exploiting the parallelismcommunication tradeoff in pipelined parallelism. In Int. Conf. on Very Large Databases, pages 36--47, Santiago, Chile, 1994.

....processed at the same time (inter operation parallelism) 3 . 2 See fig. 5 for an example of a bushy tree. 3 More precisely, two classes of linear trees can be distin 3 Problem formulation and previous works Two basic optimization strategies have been studied. In a first class of methods [Hon92, HM94], the query is first processed by a sequential optimizer and parallelized afterwards. In a second class, 4 [SD90, VZ94] the parallelization strategy is included into the optimization process. In particular, the query optimizer is in charge of defining the degree of interparallelism (i.e. the ....

.... to be attributed to each operation) The problem of optimizer estimation errors : Whatever the optimization strategy, it is necessary to estimate unknown relation parameters (e.g. relation size of intermediate relations and query parameters [MW94] Based on these parameters, some systems [HM94, SD90] determine the degree of inter(intra) parallelism at compile time. Unfortunately, a bad parameter estimation can cause big problems. Thus, as showed in [IC91] or [BK95] the relation size is specially sensible to error propagation: even if estimation errors on a join operation are relatively ....

W. Hasan and R. Motwani. Optimization Algorithms for Exploiting the ParallelismCommunication Tradeoff in Pipelined Parallelism. In Proceedings of the International Conference on Very Large Databases, pages 36--47, Santiago, Chile, September 1994.

....this reduction in optimization cost may come at the price of selecting highly suboptimal plans. Nevertheless, even in this case, effective query scheduling algorithms are still necessary for distributing the execution of the plan on the run time environment during the parallelization phase [HM94, CHM95] Hence, resource scheduling techniques form an important component of any approach to query processing and optimization in parallel database systems. As a result, significant research effort has concentrated on the problem of minimizing the response time of a single query through ....

....and optimization in parallel database systems. As a result, significant research effort has concentrated on the problem of minimizing the response time of a single query through parallelization of an execution plan, i.e. scheduling of the plan s operators on the system s sites [CHM95, GW93, HM94, Hon92, HCY94, LCRY93] Most of these efforts, however, are based on simplifying assumptions that limit their applicability. We address the parallel query plan scheduling problem in its most general form, assuming the full variety of bushy plans and schedules that incorporate independent and ....

[Article contains additional citation context not shown here]

Waqar Hasan and Rajeev Motwani. "Optimization Algorithms for Exploiting the ParallelismCommunication Tradeoff in Pipelined Parallelism". In Proceedings of the 20th International Conference on Very Large Data Bases, pages 36--47, Santiago, Chile, August 1994.

....as a directed graph (operator graph) with edges (i,j) meaning that the operator j cannot start execution until i finishes to process its data. From this point, an allocation of the physical resources can be seen as a classical problem of scheduling (second level) This problem is NP complete [16][17], so heuristic approaches must be proposed. Depending on the characteristics of the operator graph, various heuristic solutions have been considered [18] A. Hamurlai et al. in [19] and W. Hasan et al. 17] 20] proposed scheduling strategies for a given query at run time, but both authors do not ....

....be seen as a classical problem of scheduling (second level) This problem is NP complete [16] 17] so heuristic approaches must be proposed. Depending on the characteristics of the operator graph, various heuristic solutions have been considered [18] A. Hamurlai et al. in [19] and W. Hasan et al. [17] [20] proposed scheduling strategies for a given query at run time, but both authors do not allow operators to control query processing. Wei Hong proposed [21] 22] the idea of a runtime system controlling the query processing in order to detect errors in cost estimation models. The degree of ....

Hasan W. and Motwani R. Optimization Algorithms for Exploiting the ParallelismCommunication Tradeoff in Pipelined Parallelism. In Proceedings of the International Conference on Very Large Databases, pages 36--47, Santiago, Chile, September 1994.

....this is acceptable for some algorithms implementing relational algebra operators, it is not true for object oriented methods with multiple inputs. Moreover, the delay between the start time of a node and the start time of its parent in a pipeline is almost uniformly ignored. Hasan and Motwani [HM94, HM95] study the problem of pipelined parallelism from the viewpoint of communication tradeoff. They propose heuristics to minimize communication overhead taking multitasking into account [HM94] and a method of specifying the operator to processors mapping [HM95] However, they do not study the impact ....

....of a node and the start time of its parent in a pipeline is almost uniformly ignored. Hasan and Motwani [HM94, HM95] study the problem of pipelined parallelism from the viewpoint of communication tradeoff. They propose heuristics to minimize communication overhead taking multitasking into account [HM94], and a method of specifying the operator to processors mapping [HM95] However, they do not study the impact of the network configuration. This configuration affects the selection of physical channels to connect two communicating processors, as well as the load on the channels. A router may even ....

[Article contains additional citation context not shown here]

Waqar Hasan and Rajeev Motwani. Optimization algorithms for exploiting the parallelism-communication tradeoff in pipelined parallelism. In Int. Conf. on Very Large Databases, pages 36--47, Santiago, Chile, 1994.

....Allocation Plan costs Transformed Plan Join XX Best implementation Join cost Parallized plan Load Manager Initial plan (best sequential plan) Fig. 4. The modules of the parallel query optimizer Query Optimizer Different optimization strategies have been studied. In a first class of methods [HON92,HAS94], an optimal sequential query plan is build, and parallelized afterwards. In a second class [SCH90,LAN94] the parallelization strategy is included into the optimization process. Query optimization usually takes place at compile time (before query execution) However, many system parameters such ....

....intermediate relations and the number of free processors in a parallel database system remain unknown until run time. Most current database systems avoid the unknown parameter problem by making certain assumptions about the values of the system parameters. Based on these parameters, some systems [HAS94,SCH90] determine the degree of inter(intra) parallelism at compile time. However, at run time these assumptions may be violated because the run time workload is unpredictable. When these assumptions are violated, queries need to be re optimized to avoid performance degradation. Some authors propose to ....

W. Hasan and R. Motwani. Optimization Algorithms for Exploiting the Parallelism-Communication Tradeoff in Pipelined Parallelism. Proceedings of the International Conference on Very Large Databases, September 1994.

.... compile time optimization: minimizing the response time of a single query through parallelization of an execution plan, i.e. scheduling of the plan s operators on the system s sites (the plan is usually the result of an earlier phase of conventional centralized query optimization) CHM95, GW93, HM94, Hon92, HCY94, LCRY93] and 2. run time execution: achieving some system wide performance goals (e.g. maximizing query throughput) by adaptive scheduling of the operators of multiple concurrent queries [MD93, MD95, RM95] We address the first problem, i.e. parallelization of query execution ....

....on parallel query scheduling has typically ignored the multi dimensional nature of database queries. It has simplified the allocation of resources to a mere allocation of processors, hiding the multidimensionality of query operators under a scalar cost metric like work or time [CHM95, GW93, HM94, HCY94, LCRY93] This one dimensional model of scheduling is inadequate for database operations that impose a significant load on multiple system resources. In this paper, we present a framework for multi dimensional resource scheduling in shared nothing parallel database systems. Building on the ....

[Article contains additional citation context not shown here]

Waqar Hasan and Rajeev Motwani. "Optimization Algorithms for Exploiting the ParallelismCommunication Tradeoff in Pipelined Parallelism". In Proceedings of the 20th International Conference on Very Large Data Bases, pages 36--47, Santiago, Chile, August 1994.

.... compile time optimization: minimizing the response time of a single query through parallelization of an execution plan, i.e. scheduling of the plan s operators on the system s sites (the plan is usually the result of an earlier phase of conventional centralized query optimization) CHM95, GW93, HM94, Hon92, HCY94, LCRY93] and 2. run time execution: achieving some system wide performance goals (e.g. maximizing query throughput) by adaptive scheduling of the operators of multiple concurrent queries [MD93, MD95, RM95] We address the first problem, i.e. parallelization of query execution ....

....on parallel query scheduling has typically ignored the multi dimensional nature of database queries. It has simplified the allocation of resources to a mere allocation of processors, hiding the multi dimensionality of query operators under a scalar cost metric like work or time [CHM95, GW93, HM94, HCY94, LCRY93] This onedimensional model of scheduling is inadequate for database operations that impose a significant load on multiple system resources. In this paper, we present a framework for multidimensional resource scheduling in shared nothing parallel database systems. Building on the ....

[Article contains additional citation context not shown here]

W. HasanandR. Motwani. "Optimization Algorithms for Exploiting the Parallelism-Communication Tradeoff in Pipelined Parallelism". In Proc. of the 20th Intl. Conferenceon Very LargeData Bases, Santiago, Chile, August 1994.

....and modelled in the cost function, or not. One approach to this problem is the generation of an optimal query execution plan and the subsequent construction of an optimal schedule for it. In [7] a parallel schedule is generated from an optimal sequential query execution plan using heuristics. In [6], a scheduling algorithm taking multitasking and communication cost into account is applied on an already available query execution plan. The advantage of this approach is that scheduling can be postponed to run time, when information on the available processors is available. The disadvantage is ....

....wait point must complete before their parent can start execution. The processors executing them may be assigned to tasks starting after the parent join nodes. The task to processor mapping is either undertaken by the optimizer, as in [4, 12, 19] or assigned to a run time scheduler, as in [7, 6]. Scheduling heuristics fall beyond the scope of this study. 5.3 Usage of the Cost Model for Parallel Query Optimisation We have implemented the cost model described thus far within the framework of a parallel optimiser, intended for the optimisation of large join queries. We outline this ....

W. Hasan and R. Motwani. Optimization algorithms for exploiting the parallelism-communication tradeoff in pipelined parallelism. In Int. Conf. on Very Large Databases, pages 36--47, Santiago, Chile, 1994.

....modelled in [1] where the execution time of a pipeline participant is observed as a function of the number of processors assigned to it. However, for large join queries, it is rather expected that multiple processes compete for the same processor. Pipelined and bushy parallelism [3] is studied in [1, 4, 2, 6], where the communication across logical links between interacting processes is also considered. However, the network configuration affects the selection of the physical channels to connect two communicating processors. A router may even dynamically choose different routes between the same ....

....are operators annotated with execution algorithm and data access information. The children of a node x are its producers ; x is the consumer . A producer is blocking if it must complete execution before its consumer starts, or pipelining if each tuple it outputs is immediately consumed [4]. In a bushy QEP, a node may have one pipeline and one blocking producer, e.g. a hash join, or two pipeline producers, e.g. a merge join on two sorted inputs. A schedule represents a node to processor mapping for the nodes of a QEP. We do not consider node parallelism. Hence, a node is assigned to ....

Waqar Hasan and Rajeev Motwani. Optimization algorithms for exploiting the parallelism-communication tradeoff in pipelined parallelism. In Int. Conf. on Very Large Databases, pages 36--47, Santiago, Chile, 1994.

....[13] Compared to our model of a schedule for partitioned and independent parallelism (Definition 4. 1) pipelined execution constrains the placement and execution of compatible clone subsets to ensure that all the clones in a pipe run concurrently they all start and terminate at the same time [16]. This means that it is no longer possible to schedule resources at one site independent of the others, as we suggested in the previous section. Compatible subsets containing clones from the same pipeline must run concurrently. Furthermore, given that the scheduler is not allowed to modify the ....

W. Hasan and R. Motwani. "Optimization Algorithms for Exploiting the Parallelism-Communication Tradeoff in Pipelined Parallelism". In Proc. of the 20th Intl. VLDB Conf., August 1994.

....the XPRS approach to be inapplicable to architectures such as shared nothing that have significant communication costs. Other work on parallel query optimization [SE93, LST91, SYT93, CLYY92, HLY93, ZZBS93, GHK92] also ignored modeling communication overheads of parallelism. Our earlier work [HM94, CHM95] focussed on the parallelization phase and has developed scheduling algorithms that account for the trade off between parallelism and communication. Though query processing in parallel and distributed databases [CP84, OV91, YC84] is fundamentally similar, repartitioning intermediate ....

W. Hasan and R. Motwani. Optimization Algorithms for Exploiting the Parallelism-Communication Tradeoff in Pipelined Parallelism. In Proceedings of the Twentieth International Conferenceon VeryLarge Data Bases, pages 36--47, Santiago, Chile, September 1994.

....[DG92, Val93] to speed up database queries presents a parallelism communication trade off. While work is divided among processors, the concomitant communication increases total work itself [Gra88, PMC 90] We can represent the task to be scheduled as a weighted operator tree [Hon92, Sch90, HM94a] in which nodes represent atomic units of execution (operators) and directed edges represent the flow of data as well as timing constraints between operators. E mail: chekuri cs.stanford.edu. Supported by an OTL grant and NSF Young Investigator Award CCR 9357849, with matching funds from IBM, ....

....approximation algorithms that run in polynomial time and guarantee small bounds on the performance ratio [Mot92] This paper develops two approximation algorithms. The faster algorithm has a tight performance ratio of 3.56 while the slower algorithm has a tight ratio of 2.87. Our earlier work [HM94a] developed algorithms for this problem that are known to have bounded performance ratios only when the shape of the tree is a path or a star. In Section 2, we provide an overview of parallel query optimization and develop a model for scheduling problems. In Section 3, we review past work on the ....

[Article contains additional citation context not shown here]

W. Hasan and R. Motwani. Optimization Algorithms for Exploiting the ParallelismCommunication Tradeoff in Pipelined Parallelism. In Proceedings of the Twentieth International Conference on Very Large Data Bases, pages 36--47, Santiago, Chile, September 1994.

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