A. Shatdal and J. F. Naughton. Using Shared Virtual Memory for Parallel Join Processing. In SIGMOD, 1993.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....Early work concentrated on parallelizing individual, traditional content sensitive operators like hybrid hash join [25] and sort (e.g. 10, 20, 1] The abstractions which inspired Flux, Exchange [11] and RiverDQ [23] were proposed to compose such operators into a dataflow. Shatdal and Naughton [27] describe how to leverage shared virtual memory across a shared nothing cluster to implement hybrid hash join. DeWitt et al. present practical techniques for handling data skew for a hash join and external sort [9, 10] These techniques rely on sampling a static data set, which is infeasible in ....

A. Shatdal and J. F. Naughton. Using Shared Virtual Memory for Parallel Join Processing. In SIGMOD, 1993.

....of machines: shared nothing and shared everything parallel machines. The current trend in technology is that the share nothing architecture is wining up hand among different parallel architectures because the easiness in scalability and the high bandwidth (up to 200MB s) in interprocessor network [51]. In analyzing parallel join, we distinguish between I O cost and time cost. I O cost is the amount of I Os that needs to be performed, and time cost is the time to perform that amount of I Os, possibly parallelly by many processors. 8.1 Shared Nothing Parallel Machine It can also be called ....

....of SM and SN , but also comes from the fact the in [55] fewer processors are used in the parallel machine. 8.3 Shared Virtual Memory Shared virtual memory [36, 42] SV M provides a single virtual address space in a sharednothing parallel machine. The main reasons that SV M is useful are [51]: 1. ease of coding 2. ease of conceptualization of shared data structures. Knowing the sensitivity of parallel hash join to data skew, Shatdal et al. proposed a parallel join technique using shared virtual memory [51] to alleviate the effect of data skew in S, the probing relation. The novice ....

[Article contains additional citation context not shown here]

A. Shatdal and J. F. Naughton. Using Shared Virtual Memory for Parallel Join Processing. Proc. of 1993 ACM SIGMOD Int'l Conf. on Management of Data, 1993, pp. 119-128.

....number of variables results in the minimisation algorithms becoming too slow, simulated annealing can be used to find a minimal buffer allocation. 5. 8 Parallelism In recent years, a large amount of research has taken place into parallel join algorithms, particularly parallel hash join algorithms [68, 71, 73, 81]. Many of these algorithms are based on existing hash join algorithms, often the hybrid hash join algorithm. We believe that our technique will be just as important in this domain as in the sequential case. Parallel join algorithms incur network (or shared memory) costs, in addition to the costs ....

....to both design algorithms which are guaranteed to produce optimal results, and to design algorithms which produce similar, or superior, results in less time. In recent years, much research has been conducted into the designing of parallel algorithms to implement relational databases operations [5, 43, 47, 68, 71, 73, 81]. We believe that the work in this thesis can be applied in this environment. For example, parallel join algorithms can be analysed with the cost model described in Chapter 5 to determine whether their use of memory can be improved. Similarly, the clustering of relations on separate machines is a ....

A. Shatdal and J. F. Naughton. Using shared virtual memory for parallel join processing. In Proceedings of the 1993 ACM SIGMOD International Conference on the Management of Data, pages 119--128, Washington, DC, USA, May 1993.

....based architecture that provides the flexibility of a SE system and the scalability of a SN system (see Fig. 1) In [BHT90] the authors present a DSM based database architecture. However, the thrust of their work was efficient coherence maintenance and concurrency control. DSM was also used in [SN93] to enhance join strategies to efficiently handle data skew. However, these studies did not examine the impact of the architectural assumptions on the scalability of the database system. The DSM architecture we propose consists of interconnected processors, each with a large amount of main memory ....

A. Shatdal and J. F. Naughton. Using shared virtual memory for parallel join processing. In Proceedings of the 1993 ACM SIGMOD, pages 119--128, May 1993.

....have focused on handling attribute value skew in the build relation [KO90, HL91, DNSS92] As we show later, attribute value skew in the probe relation and join product skew can have significant impact on the performance of parallel join execution. The algorithm due to Shatdal and Naughton [SN93] handles join product skew as well as attriubute value skew in the build and probe relations in the context of a shared virtual memory architecture by dynamically distributing tuples to idle nodes during join processing. From the above discussion it appears that, query result distribution plays ....

....is modeled based on the Fujitsu Model M2266 (1 GB, 5.25 ) disk drive. The CPU is scheduled using a roundrobin policy and the page replacement in the buffer pool is based on LRU with love hate [H 90] hints. The instruction counts for various operations in the validated simulator are taken from [SN93] and are listed in Table 3. 7.2 Experiment Testbed Algorithms: The followingalgorithms are used to compute the split table in PHHJOIN . All the experiments compute the time taken by PHHJOIN using the split table Table 3: Simulation Parameter Settings CPU Cost Parameter No. Instr. ....

Ambuj Shatdal and Jeffrey F. Naughton. Using shared virtual memory for parallel join processing. Proc. of ACM SIGMOD Conf, May 1993.

....If successful, the model could yield a potentially important analytical evaluation technique for such programs. For example, the model might be useful for application areas such as parallel query processing in database systems where, again, significant partitioning and scheduling issues arise [ShN93]. A common workload characteristic complicating the problem of partitioning is data skew. In such cases, skewed task execution times would have to be represented as a set of (unequal) deterministic quantities, with the partitioning of work represented by the scheduling function. These aspects are ....

A. SHATDAL and J. F. NAUGHTON, Using Shared Virtual Memory for Parallel Join Processing, Computer Sciences Technical Report #1139, Univ. of Wisconsin-Madison, March 1993.

....in using shared virtual memory (SVM) for database systems and global memory management. Bellew et al. BHT90] have used SVM to share data between clients in a client server database system. The thrust of this work is efficient coherence maintenance and concurrency control. Shatdal and Naughton [SN93] assume a multiprocessor system with a software SVM layer to enhance join strategies that efficiently handle data skew. However, neither of these studies have examined the use of SVM for better memory utilization. Franklin et al. FCL92] have proposed a global memory management scheme in the ....

A. Shatdal and J. F. Naughton. Using Shared Virtual Memory for Parallel Join Processing. In Proceedings of the 1993 ACM SIGMOD Conference, pages 119--128, May 1993.

....this study was derived from a simulation model of the Gamma parallel database machine which has been validated against the actual Gamma implementation. This simulator has been used in a number of ongoing studies in scheduling and resource allocation for centralized and parallel database systems [Brow92, Brow93, Meht93a, Meht93b, Shat93]. The simulator is written in the CSIM C process oriented simulation language [Schw90] For this study, we use a centralized configuration that consists of one processing node with a single CPU, memory, and two disks. The remainder of this section provides a more detailed description of the ....

Shatdal, A., and Naughton, J., "Using Shared Virtual Memory for Parallel Join Processing ", Proceedings of the 1993 SIGMOD Conference on the Management of Data, Washington, 1993.

....In shared nothing, intra operator parallelism is based on relation partitioning [Bor90, DeW90, Ape92] Skewed data distributions [Wal91] can yield poor intra operator load balancing. This problem has been addressed by developing specific join algorithms that handle different kinds of skew [Kit90, DeW92, Ber92, Sha93] based on dynamic data redistribution. With inter operator parallelism, distributing the query s operators among all processors can also yield poor load balancing. Much research has been dedicated to inter operator load balancing in sharednothing [Meh95, Rah95, Gar96] which is done statiPage ....

....queues of the SM node. Therefore, a thread gets idle only when there is no more activation of any operator, which means that there is no more work to do on its SMnode which is starving. When an SM node gets starving, we can apply load sharing with another SM node by acquiring some of its workload [Sha93]. However, acquiring activations (through message passing) incurs communication overhead. Furthermore, activation acquisition is not enough since associated data, i.e. hash tables, must also be acquired. Thus, we need a mechanism that can dynamically estimate the benefit of acquiring activations ....

A. Shatdal, J. F. Naughton, "Using Shared Virtual Memory for Parallel Join Processing". ACM-SIGMOD Int. Conf., Washington, May 1993.

....skew. Redistribution skew is also not considered, since techniques like the ones presented in [DeWi92a] can easily be used to eliminate it. On the other hand, remote data access capabilities are required for reducing join product skew. For example, distributed virtual shared memory was used in [Shat93] for handling join product skew. Since no remote data access is assumed in this paper, join product skew may be present. Section 5.6 presents an experiment that explores the effect of join product skew on data placement. 3.5 Result collection The output of a parallel operation sometimes needs to ....

Shatdal, A. and Naughton, J., "Using Shared Virtual Memory for Parallel Join Processing,",<F3.17e+05> Proc. ACM SIGMOD<F3.733e+05> Conf., Washington, DC, May 1993.

....(DSM) based architecture that provides the flexibility of a SE system and the scalability of a SN system (see Fig. 1) In [2] the authors present a DSM based database architecture. However, the thrust of their work was efficient coherence maintenance and concurrency control. DSM was also used in [25] to enhance join strategies to efficiently handle data skew. However, these studies did not examine the impact of the architectural assumptions on the scalability of the database system. PROCESSOR MEMORY DISK RING INTERCONNECT RING INTERCONNECT RING INTERCONNECT DISTRIBUTED SHARED MEMORY Figure ....

....explicitly model joins in the workloads. The reason for this is two fold. First, the major performance differences between joins in SN and SS would show up in the build phase of join processing. This aspect of join performance is captured by the various select queries in the workload. Second, [25] shows how DSM can be exploited to eliminate load imbalances that could arise in the join phase due to skewed distribution of data by the build phase. 4 Results In this section we present the results of our experiments to determine the effects of declustering and skew, the scalability of our ....

A. Shatdal and J. F. Naughton. Using shared virtual memory for parallel join processing. In Proceedings of the 1993 ACM SIGMOD, pages 119--128, May 1993.

....used in this study models a centralized system composed of one CPU and its memory buffer pool, four SCSI disks and a set of external terminals from which scans are submitted. Our simulator is based on the simulator of a parallel database machine that was used in a number of studies, e.g. in [Mehta93a, Shatdal93, Brown93, Mehta93b, Brown94], and is written in the CSIM C process oriented simulation language ( Schwetman90] The terminals model the external workload source for the system. For closed end experiments, each terminal submits a stream of scans, one at a time, waiting for a response to each before sending the next one. ....

A. Shatdal and J. Naughton. Using Shared Virtual Memory for Parallel Join Processing. In Proc. 1993 ACM-SIGMOD Conf. Management of Data, Washington, DC (1993), pp. 119-128.

....positional indexing, which has been proposed in [16, 3] As stated earlier, the physical design strategy underlying PSJ exploits many of these approaches in the context of the Star Schema. A large body of work exists in applying parallel processing techniques to relational database systems (e.g. [4, 20, 22, 19]) From this work has emerged the notion that highly parallel, shared nothing architectures can yield much better performance than equivalent closely coupled systems [18, 13, 5] Indeed, may commercial database vendors have capitalized on this fact [6] Our focus on shared nothing systems is also ....

A. Shatdal and J. Naughton. Using shared virtual memory for parallel join processing. In Proc ACM SIGMOD, pages 119--128, June 1993.

....the 47 next chapter we attempt to answer similar questions for the hash based aggregation algorithm. One argument against shared nothing algorithm has been their poor performance when the data is skewed [WDJ91a] However, techniques that have proven effective for shared nothing algorithms, e.g. SN93] would trivially apply to SMPs. 48 Chapter 4 Hash Aggregation on Shared Nothing Hardware The traditional parallel aggregation algorithms for shared nothing architectures perform poorly when the number of groups is large, as expected in decision support and online analytical processing (OLAP) ....

Ambuj Shatdal and Jeffrey F. Naughton. Using Shared Virtual Memory for Parallel Join Processing. In Proc. of the 1993 ACM-SIGMOD Conference, pages 119--128, Washington, D.C., May 1993.

....be enhanced by optimizing the algorithms by making them aware of the cache and the SMP architecture. One argument against shared nothing algorithm has been their poor performance when the data is skewed [WDJ91] However, techniques that have proven effective for shared nothing algorithms, e.g. SN93] would trivially apply to SMPs. It would be interesting to compare the two approaches for skew handling. Also, cluster of SMPs seems to be gaining popularity for building large scalable database systems, instead of the traditional shared nothing hardware. Investigating the algorithms for a ....

Ambuj Shatdal and Jeffrey F. Naughton. Using Shared Virtual Memory for Parallel Join Processing. In Proc. of the 1993 ACM-SIGMOD Conference, pages 119--128, Washington, D.C., May 1993.

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