D. G. Feitelson and L. Rudolph, Mapping and scheduling in a shared parallel environment using distributed hierarchical control, in: Intl. Conf. Parallel Processing (1990) I-1{I-8.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....Intel Paragon [25] the SGI multiprocessors running IRIX [1] the Meiko CS 2 [14] the Alliant FX 8 [41] and the MasPar and DAP SIMD arrays. Gang scheduling has also been used in a production system on a BBN Butterfly at LLNL [20] which Parts of this research have been presented at conferences [18, 19]. Part of this work was done while at the IBM T. J. Watson Research Center, Yorktown Heights, NY 10598. Supported in part by grants from the Alexander Silberman Hebrew University Foundation for Applied Science and the Ernst David Bergman Fund for Science. ments for processors. The ....

Feitelson, D. G., and Rudolph, L. Mapping and scheduling in a shared parallel environment using distributed hierarchical control. Intl. Conf. Parallel Processing, Aug. 1990, Vol. I, pp. 1--8.

....and at the individual job level. Gang scheduling encompasses a very broad range of schedulers depending on the particular schemes used for partitioning resources and for sharing resources within each partition. One particular approach is based on the distributed hier archical control structure [4, 5, 6]. Within the context of the above description, this scheme can be logically viewed as having a scheduling matrix with log P 1 rows, where the i row contains 2 i partitions each of size P 2 i, 0 i logP, and P denotes the number of system processors. A somewhat different approach, which can ....

....viewed as a generalization of Ousterhout s original global scheduling matrix, has also been considered [13, 14] Due to its promising characteristics, gang scheduling has attracted considerable attention in recent years. Gang schedulers based on the distributed hierar chical control structure [4, 5, 6] have been implemented for the IBM PS 6000 SP2 [8, 30] and for clusters of workstations [9, 29] Similarly, another form of gang scheduling has been implemented on both the IBM SP2 and a cluster of worksta tions [13, 14] The performance of gang scheduling schemes that use distributed ....

D. G. Feitelson and L. Rudolph. Mapping and scheduling in a shared parallel environment using distributed hierarchical control. In Proc. International Conf. Parallel Processing, volume I, pages 1 8, August 1990.

....of various scheduling algorithms. 1 Introduction In recent years massively parallel processors (MPP) computers have made a significant presence. With this growth in MPPs, a number of researchers have developed and are continuing to develop various job scheduling subsystems for these MPPs [1, 2, 3, 4, 5, 6]. During the development of these schedulers and their related algorithms, it is important to have an accurate characterization of the workloads experienced by the MPPs. It is extremely advantageous to have these workloads characterized by a compact model that is representable by a few parameters, ....

D. Feitelson and L. Rudolph, "Mapping and scheduling in a shared parallel environment using distributed hierarchical control," in Intl. Conf. Parallel Processing, pp. 1--8, 1990.

....show that DQT exhibits very good linearity in processor utilization from low load situations to high load situations, independent of job size distribution, and also good stability even with a 99 workload. 2. 2 Related Works Feitelson and Rudolph proposed a Distributed Hierarchical Control (DHC)[7, 6] that is also an instance of TSSS. DQT and DHC are very similar in the nature of distribution and scalability. The major difference between DQT and DHC, however, lies in the assumptions made and the target computing environment. In DQT, the scheduling unit is a process , an entity of parallel ....

D. G. Feitelson and L. Rudolph. Mapping and Scheduling in a Shared Parallel Environment Using Distributed Hierarchical Control. In International Conference on Parallel Processing, volume I, pages 1--8, 1990.

....in a workstation cluster that allows the time sharing of a 64 processor machine interconnected by Myrinet. Feitelson and Rudolph proposed the distributed hierarchical control, which is a scalable implementation of Gang scheduling based on buddy systems. The distributed hierarchical control (DHC) [12, 15, 13] defines a control structure over the parallel machine and combines time slicing with a buddy system partitioning scheme. In [12] a DHC scheme for supporting Gang scheduling was proposed. Gang schedulers based on the distributed hierarchical control structure has been implemented for the IBM ....

D Feitelson and L. Rudolph. Mapping and Scheduling in a Shared Parallel Environment Using Distributed Hiearchical Control. In Proceedings of the 1990 International Conference on Parallel Processing, 1990.

....of gang service have been described in the literature. Hori et al. 16] describe a gang scheduler implementation in a workstation cluster that allows the time sharing of a 64 processor machine interconnected by Myrinet. Feitelson and Rudolph proposed the distributed hierarchical control (DHC) [11, 14, 12], which defines a control structure over the parallel machine and combines time slicing with a buddy system partitioning scheme. In [11] a DHC scheme for supporting gang scheduling was proposed. Gang schedulers based on the distributed hierarchical control structure has been implemented for the ....

D Feitelson and L. Rudolph. Mapping and Scheduling in a Shared Parallel Environment Using Distributed Hiearchical Control. In Proceedings of the 1990 International Conference on Parallel Processing, 1990. 9

....Our emphasis here is to improve throughput and utilization of gang schedulers. Observe that the strategies described in this section can be applied to a large number of gang scheduler implementations, including traditional gang schedulers[3, 15] and distributed hierarchical control schedulers [11, 12]. We may consider two types of parallel tasks in a gang scheduler: Those that should be scheduled as a gang with other tasks in other processors and those for which gang scheduling is not mandatory. Examples of the rst class are tasks that compose a job with ne grain synchronization ....

D Feitelson and L. Rudolph. Mapping and Scheduling in a Shared Parallel Environment Using Distributed Hiearchical Control. In Proceedings of the 1990 International Conference on Parallel Processing, 1990.

....Our emphasis here is to improve throughput and utilization of gang schedulers. Observe that the strategies described in this section can be applied to a large number of gang scheduler implementations, including traditional gang schedulers[3, 15] and distributed hierarchical control schedulers [11, 12]. We may consider two types of parallel tasks in a gang scheduler: Those that should be scheduled as a gang with other tasks in other processors and those for which gang scheduling is not mandatory. Examples of the first class are tasks that compose a job with fine grain synchronization ....

D Feitelson and L. Rudolph. Mapping and Scheduling in a Shared Parallel Environment Using Distributed Hiearchical Control. In Proceedings of the 1990 International Conference on Parallel Processing, 1990.

....Space sharing helps decrease resource fragmentation and increase throughput. Studies have also shown that space sharing can yield higher throughput than time sharing in some environments [24, 28] Gang scheduling is a mixed scheduling scheme that combines time sharing with space sharing [6, 7, 8]. Under gang scheduling, jobs with the same resource requirements are collected in groups. Each group is assigned to a partition of the system for a specific amount of time. When the time allocated to a group is up, a context switch occurs, resources are reallocated, and the jobs in new groups are ....

D. G. Feitelson and L. Rudolph. Mapping and scheduling in a shared parallel environment using distributed hierarchical control. In Proceedings of the International Conference on Parallel Processing, volume I, pages 1--8, August 1990.

....and at the individual job level. Gang scheduling encompasses a very broad range of schedulers depending on the particular schemes used for partitioning resources and for sharing resources within each partition. One particular approach is based on the distributed hierarchical control structure [4, 5, 6]. Within the context of the above description, this scheme can be logically viewed as having a scheduling matrix with log P 1 rows, where the i th row contains 2 i partitions each of size P=2 i , 0 i log P , and P denotes the number of system processors. A somewhat different approach, ....

....viewed as a generalization of Ousterhout s original global scheduling matrix, has also been considered [13, 14] Due to its promising characteristics, gang scheduling has attracted considerable attention in recent years. Gang schedulers based on the distributed hierarchical control structure [4, 5, 6] have been implemented for the IBM RS 6000 SP2 [8, 30] and for clusters of workstations [9, 29] Similarly, another form of gang scheduling has been implemented on both the IBM SP2 and a cluster of workstations [13, 14] The performance of gang scheduling schemes that use distributed hierarchical ....

D. G. Feitelson and L. Rudolph. Mapping and scheduling in a shared parallel environment using distributed hierarchical control. In Proc. International Conf. Parallel Processing, volume I, pages 1--8, August 1990.

.... problem, and suggested methods that allow a processor to dequeue multiple threads or schedule threads onto other processors [15] In a recent study, Feitelson and Rudolph study a scalable distributed hierarchical control structure for gang scheduling in a large scale multiprogrammed environment [9]. 6 Conclusion In this paper, we proposed a class of scheduling algorithms based on processor pools for largescale NUMA multiprocessors. Our simulation experiments use a set of simple workloads and a simple system model to show that parallel application scheduling based on processor pools may ....

D. G. Feitelson and L. Rudolph. Mapping and scheduling in a shared parallel environment using distributed hierarchical control. In 1990 International Conference on Parallel Processing, pages I1--I8, 1990.

....levels in the hierarchy. Besides being a means for achieving scalability, clusters also localize memory accesses, which is the key to good performance in NUMA systems [Unrau1992] Similar structuring methods have been proposed for scheduling subsystems in scalable architectures [Feitelson1990] Feitelson1990a] Ahmad1991] 18 2.4.3. Localizing Data Accesses As processor technology continues to improve at a faster rate than memory or interconnection network technology, the relative increase in communication costs in multiprocessors has become a topic of growing importance. A number of recent studies ....

....is limited) Because each pool of processors may be managed independently, the likelihood of bottlenecks inherent in traditional single ready queue systems is reduced. The notion of grouping processors to enhance scalability has also been proposed by other researchers [Black1990] Feitelson1990] Feitelson1990a] Ahmad1991] Feitelson and Rudolph s distributed hierarchical technique is designed to also gang schedule and load balance multiple applications in large multiprocessor systems [Feitelson1990] Feitelson1990a] Their evaluation of this technique does not take into account NUMA multiprocessors. ....

[Article contains additional citation context not shown here]

D. G. Feitelson and L. Rudolph, "Mapping and Scheduling in a Shared Parallel Environment Using Distributed Hierarchical Control", 1990 International Conference on Parallel Processing, pp. I1-I8, 1990. 126

....tradeoffs among these parallel system design choices, and it is this set of important tradeoffs that motivate our present study. In this paper, we consider a particular gang scheduling strategy that has been proposed in the research literature and uses a distributed hierarchical control structure [9,10,52]. This approach features flexible partitions and flexible quantum lengths for every partition, where each job is allocated to a certain configuration of the resources so that there is a thread of that job on each processor. Several jobs can be executing on different processors at the same time, ....

....2.1 System Model We consider a parallel processing system that consists of P identical processors and L different classes of jobs. Each job class p, 0 p L, divides the P processors into equal size disjoint partitions, and each set can accommodate a job that requires g p processors, 0 g p P [9,10,52]. An infinite capacity queue is associated with each job class, from which the processors select work for execution under a particular gang scheduling policy that we define in the next section. A first come first served (FCFS) queueing discipline is employed at each of these L queues. Variables ....

[Article contains additional citation context not shown here]

D. G. Feitelson and L. Rudolph. Mapping and scheduling in a shared parallel environment using distributed hierarchical control. In Proceedings of the International Conference on Parallel Processing, volume I, pages 1--8, August 1990.

....more complicated. To simplify the presentation, it is also assumed that processes are mapped to processors upon creation and do not migrate at run time. This assumption is reasonable in the context of parallel processing, where load balancing is less bene cial than in distributed systems [5]. However, this assumption is not required and the algorithm is easily extended to deal with migration. In addition it is assumed that only one application is running at a time, and all the processes belong to it. This assumption is also easily removed by performing the necessary bookkeeping for ....

D. G. Feitelson and L. Rudolph, Mapping and scheduling in a shared parallel environment using distributed hierarchical control, in: Intl. Conf. Parallel Processing (1990) I-1{I-8.

....sense that asymptotically they achieve performance commensurate with off line algorithms. Keywords: Distributed Hierarchical Control, fragmentation, gang scheduling, load balancing, mapping, processor utilization, variable partitioning. Parts of this research have been presented at conferences [18, 19]. Part of this work was done while at the IBM T. J. Watson Research Center, Yorktown Heights, NY 10598. 1 1 Introduction While there is great interest in multiuser, general purpose parallel computing, it is not obvious how to partition the computing resources of the parallel computer among ....

D. G. Feitelson and L. Rudolph, "Mapping and scheduling in a shared parallel environment using distributed hierarchical control". In Intl. Conf. Parallel Processing, vol. I, pp. 1--8, Aug 1990.

....partitions of power of two PEs 3.2 [307] Cray T3E partitioninig, gang scheduling, and fair share 5 [346] Crystal partitioning by users 3.3 [151] DASH global queue with locking 4.2 [357] DHC design from Jerusalem gang scheduling with flexible grouping based on buddy system 5.4, 4.1. 1 [197, 199] DQT for RWC 1 gang scheduling with flexible grouping based on buddy system 5.4, 4.1.1 [273] Dynix (Sequent Balance) global queue with locking 4.2 [570] EMBOS (Elxsi) local queues 4.1 [498, 436] EMMA2 local queues 4.1 [26] Flagship local queues with hardware mapping confined to ....

D. G. Feitelson and L. Rudolph, "Mapping and scheduling in a shared parallel environment using distributed hierarchical control". In Intl. Conf. Parallel Processing, vol. I, pp. 1--8, Aug 1990.

....is not always optimal (although it does favor individual jobs) There is some interference in the cache, and overhead for the context switching. In addition, there may be some processor fragmentation. However, due to the time slicing, its effect is less severe than in variable partitioning [14, 15]. Indeed, most studies find that gang scheduling is nearly as efficient as dynamic partitioning [9] 4 The Requirements Literally hundreds of papers have been written about job scheduling in parallel systems (see [11] However, in many respects, we are at the beginning of the beginning. The ....

D. G. Feitelson and L. Rudolph, "Mapping and scheduling in a shared parallel environment using distributed hierarchical control". In Intl. Conf. Parallel Processing, vol. I, pp. 1--8, Aug 1990.

....typically defined to include all the activities (or processes in their terminology) in the application. Off line algorithms to find an optimal gang schedule were presented by B lazewicz et al. 2] Feitelson and Rudolph proposed a scalable implementation of gang scheduling based on buddy systems [9, 11, 12]. In these works gangs are also predefined, even though they are not necessarily identical to jobs (i.e. a job can have multiple independent gangs) This paper is the first to consider the possibility of automatically grouping the activities in a job into gangs at run time. There are many ....

D. G. Feitelson and L. Rudolph, "Mapping and scheduling in a shared parallel environment using distributed hierarchical control". In Intl. Conf. Parallel Processing, vol. I, pp. 1--8, Aug 1990.

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