P. Krueger, T.-H. Lai, and V. Dixit-Radiya. Job scheduling is more important than processor allocation for hypercube computers. IEEE Trans. on Parallel and Distributed Systems, 5(5):488--497, 1994.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....time. There has been considerable prior research into each of the two topics of i) scheduling of parallel jobs [1,2,6,10,14,19,20,23,24,26] and ii) contiguous node allocation strategies [3,4,5,11,13,27] There have also been a few studies that have considered both these issues in combination [12,15,18]. However, only [15] addresses the impact of contiguous node allocation schemes in conjunction with a job scheduling policy that takes fairness into consideration by use of a FCFS (First Come First Served) scheduling policy. In [15] contiguous and non contiguous node allocation schemes for ....

....in [17] 16] presents a strategy that minimizes network contention due to both communication and I O traffic. All the above studies focus exclusively on the topic of contiguous node allocation schemes, but do not address the issue of job scheduling onto the parallel systems. The studies in [12,18] focus on job scheduling techniques for improving the performance of hypercube computers. In [12] the roles of processor allocation and job scheduling in improving the performance of hypercube computers are compared. A new scheduling algorithm is proposed for improving the average response time ....

[Article contains additional citation context not shown here]

P. Krueger, T. Lai, and V.A. Dixit-Radiya, "Job Scheduling Is More Important than Processor Allocation for Hypercube Computers", IEEE Trans. Parallel and Distributed Systems, vol. 5, no. 5, pp. 488-497, May 1994.

....scheduling model and provide several applications for the model. O(n log n) algorithms are developed for some of the problems formulated and some others are shown to be NP hard. 1 Introduction The problem of scheduling a multiprocessor computer system has received considerable attention [3, 4, 5, 6, 7, 8, 9, 10]. In this paper, we develop a model to schedule a parallel computer system in which the parallel computer operates under control of a host processor. The host processor is referred to as the master processor and the processors in the parallel computer are referred to as the slave processors. The ....

P. Krueger, T. Lai, and V. Dixit-Radiya, Job scheduling is more important than processor allocation for hypercube computers, IEEE Trans. on Parallel 4 Distributed Systems, 5, 5, 488-497, 1994.

.... truck that delivers the raw material for the job) need not be the same as the one that postprocesses job i (i.e. the truck that brings back the finished goods corresponding to this job) While the problem of scheduling multiprocessor computer systems has received considerable attention [3] 4] [10], 12] 14] 15] 18] 22] it appears that the master slave model has not been studied prior to the work of Sahni [19] It is interesting to note that the master slave scheduling model may be regarded as a variant of the job shop (see [1] 2] for a definition of a job shop as well as for ....

P. Krueger, T. Lai, and V. Dixit-Radiya, Job scheduling is more important than processor allocation for hypercube computers, IEEE Trans. on Parallel Distributed Systems, 5, 5, 488-497, 1994.

.... truck that delivers the raw material for the job) need not be the same as the one that post processes job i (i.e. the truck that brings back the finished goods corresponding to this job) While the problem of scheduling multiprocessor computer systems has received considerable attention [3] 4] [10], 12] 14] 15] 18] 21] it appears that the master slave model has not been studied prior to the work of Sahni [19] It is interesting to note that the master slave scheduling model may be regarded as a variant of the job shop (see [1] 2] for a definition of a job shop as well as for ....

....makespan schedule can be found in O(n log n) time. Fast polynomial time algorithms to obtain minimum makespan schedules in which the pre and post processing orders are the same (or reverse) and a job may wait between the completion of one task and the start of the next are also developed in [10]. For no wait scheduling, the single master master slave model and the coupled task model of [17] are identical. Orman and Potts [17] show that many versions of this latter problem are strongly NP hard. These results carry over to the no wait master slave model. The outline of the rest of this ....

P. Krueger, T. Lai, and V. Dixit-Radiya, Job scheduling is more important than processor allocation for hypercube computers, IEEE Trans. on Parallel 4 Distributed Systems, 5, 5, 488-497, 1994. 20

.... rst may be expected to lead to better performance [16] But what if there is a correlation between size and running time If this is an inverse correlation, we nd a win win situation: the larger jobs are also shorter, so packing them rst is statistically similar to using SJF (shortest job rst) [47]. But if size and runtime are correlated, and large jobs run longer, scheduling them rst may cause signi cant delays for subsequent smaller jobs, leading to dismal average performance [53] System Correlation CTC SP2 0:029 KTH SP2 0.011 SDSC SP2 0.145 LANL CM 5 0.211 SDSC Paragon 0.305 ....

P. Krueger, T-H. Lai, and V. A. Dixit-Radiya, \Job scheduling is more important than processor allocation for hypercube computers". IEEE Trans. Parallel & Distributed Syst. 5(5), pp. 488-497, May 1994.

....obtained from the IEEE. manner to submitted jobs. But this approach su ers from fragmentation, where free processors cannot meet the requirements of the next job, and therefore remain idle until additional ones become available. As a result system utilization is typically in the range of 50 80 [21, 16, 8, 11, 15]. It is well known that the best solutions for this problem are to use dynamic partitioning [20] or gang scheduling [6] However, these schemes have practical limitations. The only ecient and widely used implementation of gang scheduling was the one on the CM5 Connection Machine; other commercial ....

P. Krueger, T-H. Lai, and V. A. Dixit-Radiya, \Job scheduling is more important than processor allocation for hypercube computers". IEEE Trans. Parallel & Distributed Syst. 5(5), pp. 488-497, May 1994.

....of processors becomes available, the job is executed, and the processors are dedicated to it until it terminates or is killed. This scheme is called variable partitioning [3] Allocating partitions on a FCFS basis results in severe fragmentation, and typical utilization of such systems is 5080 [6, 7, 9, 12]. Two solutions that have proposed to this problem, dynamic partitioning [11, 1] and gang scheduling [4] are difficult to implement and do not enjoy much use. A far simpler approach is to use a non FCFS policy when allocating partitions, for example by allowing small jobs from the back of the ....

P. Krueger, T-H. Lai, and V. A. Dixit-Radiya, "Job scheduling is more important than processor allocation for hypercube computers ". IEEE Trans. Parallel & Distributed Syst. 5(5), pp. 488--497, May 1994.

....(FCFS) manner to submitted jobs. But, this approach suffers from fragmentation, where free processors cannot meet the requirements of the next job and therefore remain idle until additional ones become available. As a result, system utilization is typically in the range of 50 80 percent [21] [16], 8] 11] 15] It is well known that the best solutions for this problem are to use dynamic partitioning [20] or gang scheduling [6] However, these schemes have practical limitations. The only efficient and widely used implementation of gang scheduling was the one on the CM 5 Connection ....

P. Krueger, T.-H. Lai, and V.A. Dixit-Radiya, "Job Scheduling Is More Important than Processor Allocation for Hypercube Computers," IEEE Trans. Parallel and Distributed Systems, vol. 5, no. 5, pp. 488-497, May 1994.

....workloads that conform to the di erent models leads to drastically di erent results. Consider a workload that is composed of jobs the use power of two processors. In this case a reasonable scheduling algorithm is to cycle through the di erent sizes, because the jobs of each size pack well together [16]. This works well for negatively correlated and even uncorrelated workloads, but is bad for positively correlated workloads [16, 17] The reason is that under a positive correlation the largest jobs dominate the machine for a long time, blocking out all others. As a result, the average response ....

....the use power of two processors. In this case a reasonable scheduling algorithm is to cycle through the di erent sizes, because the jobs of each size pack well together [16] This works well for negatively correlated and even uncorrelated workloads, but is bad for positively correlated workloads [16, 17]. The reason is that under a positive correlation the largest jobs dominate the machine for a long time, blocking out all others. As a result, the average response time of all other jobs grows considerably. But which model actually re ects reality Again, evaluation results depend on the selected ....

P. Krueger, T-H. Lai, and V. A. Dixit-Radiya, \Job scheduling is more important than processor allocation for hypercube computers". IEEE Trans. Parallel & Distributed Syst. 5(5), pp. 488-497, May 1994.

....compare many real workload traces and synthetic workload models, both analytically and through simulation, observing their effects on the performance evaluation of several classes of scheduling and allocation strategies. Included were two scheduling algorithms: First Come First Served and ScanUp [16], a multi level queuing algorithm, and three static allocation strategies: First Fit [24] Frame Sliding [3] and Paging [17] The real traces were captured from four production machines in use for scientific computing at research labs and supercomputer sites around the world (two IBM SP 2s, an ....

....of processors that is a function of the number of free processors in the system and of characteristics of the job. algorithms. Of special interest to our study is Krueger s work on scheduling and allocation performance under workloads exhibiting negative correlations between jobsize and runtimes [16]. As we shall see, our work further explains some of the phenomena he observed. The only recent study that we know of that has focused on the experimental methodology itself, i.e. the effect choice of workload has on scheduling performance results, is that of Chiang et al. 2] They compare the ....

[Article contains additional citation context not shown here]

P. Krueger, T. Lai, and V. A. Dixit-Radiya. Job scheduling is more important than processor allocation for hypercube computers. IEEE Transactions on Parallel and Distributed Systems, 5(5):488--497, May 1994.

....job onto it [3, 4, 12, 20] In most cases, however, processor utilization is far from optimal because of fragmentation. Krueger et al. proposed a new job scheduling scheme, called scan , and found that job scheduling order, not mapping, is more important to achieve higher processor utilization [11]. All methods, however, are batch scheduling and an interactive programming environment can not be provided. CM 5[18] and Paragon[10] provide time sharing scheduling. In CM 5, partitioning can only be changed at system bootup time, and in Paragon (OSF 1) partitioning and the partition in which a ....

P. Krueger, T.-H. Lai, and V. A. Dixit-Radiya. Job Scheduling Is More Important than Processor Allocation for Hypercube Computers. IEEE Transactions on Parallel and Distributed Systems, 5(5):488--497, 1994.

....linear programming has been proposed. The reported good experimental results for the second method have been explained by a particular topology of the criterial function. For P j cube j j C max Largest Dimension First (LDF) heuristic with tight performance ratio 2 0 1=m is proposed in [85] In [62] an experimental study is reported for dynamic scheduling (i.e. the set of tasks is not known in advance) for problem P j cube j j P C j . It is observed that even sophisticated processor allocation strategies alone cannot guarantee good performance. A set of Scan strategies is proposed which ....

P.Krueger, T.-H.Lai, V.A.Dixit-Radiya, "Job scheduling is more important than processor allocation for hypercube computers", IEEE Transactions on Parallel and Distributed Systems 5/5 (1994) 488-497. 23

....class of contiguous allocation strategies restricts the nodes allocated to a given job to form a convex shape. Fig. 1 shows an example. Performance suffers significantly due to processors being wasted because of internal and external fragmentation. Utilizations of only 34 to 66 are reported [9, 18, 8, 10]. In contrast, the class of noncontiguous allocation strategies allocates nodes that are dispersed throughout the system. Fig. 3 shows examples. They experience no fragmentation and thus outperform contiguous strategies reaching utilizations of up to 78 [10, 15, 14] To further improve the ....

....1 2 3 0 4 5 6 7 unallocated node node allocated to job A node allocated to job B link affected by job A link affected by job B Figure 2: External contention for the link between (6,0) and (7,0) machine. Using First Come, First Served scheduling, utilizations of only 34 to 66 are reported [9, 18, 8, 10] due to serious fragmentation problems. This is unfavorable and contradicts the goal of high throughput over a stream of jobs. 2.2 Non contiguous allocation strategies and contention To utilize unallocated nodes that are not necessarily contiguous, non contiguous processor allocation strategies ....

P. Krueger, T. Lai, and V. A. Dixit-Radiya. Job scheduling is more important than processor allocation for hypercube computers. IEEE Transactions on Parallel and Distributed Systems, 5(5):488--497, May 1994. 18

....compare many real workload traces and synthetic workload models, both analytically and through simulation, observing their effects on the performance evaluation of several classes of scheduling and allocation strategies. Included were two scheduling algorithms: First Come First Served and ScanUp [15], a multi level queueing algorithm, and three static allocation strategies: First Fit [23] Frame Sliding [3] and Paging [16] The real traces were captured from four production machines in use for scientific computing at research labs and supercomputer sites around the world (two IBM SP 2s, an ....

....and processor allocation algorithms have offered insights into the effects of workload characteristics on those algorithms. Of special interest to our study is Krueger s work on scheduling and allocation performance under workloads exhibiting negative correlations between jobsize and runtimes [15]. As we shall see, our work further explains some of the phenomena he observed. The only recent study that we know of that has focused on the experimental methodology itself, i.e. the effect choice of workload has on scheduling performance results, is that of Chiang et al. 2] They compare the ....

[Article contains additional citation context not shown here]

P. Krueger, T. Lai, and V. A. Dixit-Radiya. Job scheduling is more important than processor allocation for hypercube computers. IEEE Transactions on Parallel and Distributed Systems, 5(5):488--497, May 1994.

....rst come rst serve (FCFS) manner to submitted jobs. But this approach su ers from fragmentation, where free processors cannot meet the requirements of the next job, and therefore remain idle until additional ones become available. As a result system utilization is typically in the range of 50 80 [21, 16, 8, 11, 15]. It is well known that the best solutions for this problem are to use dynamic partitioning [20] or gang scheduling [6] However, these schemes have practical limitations. The only ecient and widely used implementation of gang scheduling was the one on the CM 5 Connection Machine; other ....

P. Krueger, T-H. Lai, and V. A. Dixit-Radiya, \Job scheduling is more important than processor allocation for hypercube computers". IEEE Trans. Parallel & Distributed Syst. 5(5), pp. 488-497, May 1994.

....that execute the job is the same as that requested by the job. A job scheduler needs to select a job to dispatch in appropriate order so as to execute multiple jobs efficiently. Many job scheduling algorithms, e.g. conventional FCFS (First Come First Served) LJF [1] Backfilling [2, 3] Scan [4], etc. have been proposed, and performance of these algorithms have been evaluated. Many previous performance evaluation works assumed that characteristics of parallel jobs, or a parallel workload, followed a simple mathematical model. However, recent analysis of real workload logs, which are ....

....investigated performance of job scheduling algorithms under more realistic workloads, which have above job size characteristics [5, 6, 7, 9] For instance, Lo, Mache and Windisch compared performance of job scheduling algorithms under various workload models. They showed that the ScanUp algorithm [4] performed well, or increased processor utilization, as the proportion of jobs requesting power of two processors in the workload increased. However, mechanisms how the job size characteristics affect performance of job scheduling algorithms have not yet been clear. This paper presents ....

[Article contains additional citation context not shown here]

P. Krueger, T. Lai, and V. A. Dixit-Radiya. Job Scheduling Is More Important than Processor Allocation for Hypercube Computers. IEEE Trans. on Parallel and Distributed Systems, 5(5):488--497, 1994.

....higher resource utilization, than FCFS. However, the fragmentation can still be quite large [363] Despite the intuitive appeal of some of these scheduling policies, studies indicate that they do not necessarily perform better than a straightforward FCFS strategy in a partitioned environment [331, 330] 7 . Moreover, systems using these schemes tend to saturate under lighter loads than FCFS. Using uniform requests for subcubes of different sizes in a hypercube as a concrete example, saturation may occur at loads as low as 40 50 of capacity [330] One simple improvement that has been ....

....FCFS strategy in a partitioned environment [331, 330] 7 . Moreover, systems using these schemes tend to saturate under lighter loads than FCFS. Using uniform requests for subcubes of different sizes in a hypercube as a concrete example, saturation may occur at loads as low as 40 50 of capacity [330]. One simple improvement that has been suggested is to change the selection algorithm on line, depending on the characteristics of the workload. Thus a decreasing order of jobs will be used to improve utilization when large batch jobs are dominant, but small jobs will be scheduled first when the ....

[Article contains additional citation context not shown here]

P. Krueger, T-H. Lai, and V. A. Dixit-Radiya, "Job scheduling is more important than processor allocation for hypercube computers". IEEE Trans. Parallel & Distributed Syst. 5(5), pp. 488--497, May 1994.

....first may be expected to cause less fragmentation, and therefore higher resource utilization, than FCFS. Despite the intuitive appeal of some of these scheduling policies, studies indicate that they do not necessarily perform better than a straightforward FCFS strategy in a partitioned environment [203, 202] 5 . Moreover, systems using these schemes tend to saturate under lighter loads than FCFS. Using uniform requests for subcubes of different sizes in a hypercube as a concrete example, saturation may occur at loads as low as 40 50 of capacity [202] 5 The more optimistic results in [230, 217] ....

....FCFS strategy in a partitioned environment [203, 202] 5 . Moreover, systems using these schemes tend to saturate under lighter loads than FCFS. Using uniform requests for subcubes of different sizes in a hypercube as a concrete example, saturation may occur at loads as low as 40 50 of capacity [202]. 5 The more optimistic results in [230, 217] are due to a model in which threads are independent, kept in a global queue, and PEs are allocated singly. In such a model, there is no loss to fragmentation. 21 now execution time time PEs reservation for large job terminated expected ....

[Article contains additional citation context not shown here]

P. Krueger, T-H. Lai, and V. A. Dixit-Radiya, "Job scheduling is more important than processor allocation for hypercube computers". IEEE Trans. Parallel & Distributed Syst. 5(5), pp. 488--497, May 1994.

....jobs are submitted via a queuing system such as NQS. This approach results in severe fragmentation, because processors which cannot fulfill the demands of the next job in the queue must remain idle until more processors are freed. FCFS based schedulers show a typical system utilization of 50 80 [6, 10, 12, 16]. Two solutions have been proposed to this problem, but both suffer from practical limitations. This first is dynamic partitioning [15] in which jobs may gain or lose some of their processors dynamically during their lifetime. Jobs may also be halted in favor of other jobs, and renewed later, ....

P. Krueger, T-H. Lai, and V. A. Dixit-Radiya, "Job scheduling is more important than processor allocation for hypercube computers". IEEE Trans. Parallel & Distributed Syst. 5(5), pp. 488--497, May 1994.

....and partitioned may be formed using arbitrary subsets of processors. A rigid job is submitted for execution along with a specification of the number of processors that it requires. The scheduler then creates a partition of that size and schedules the job to execute within that partition [53,32,20,1,9,31,33]. With moldable jobs, it is the scheduler that selects the partition size [44] Evolving and malleable jobs require partitions that are not only flexible but can also change dynamically at runtime. This places an added burden both on the programmer, who must write application code that requests ....

....This scheme has been called variable partitioning or pure space sharing in the literature. Despite its simplicity and the resulting drawbacks in terms of responsiveness, fragmentation, and reliability, this scheme is widely used. It is especially common on large distributed memory machines [53,27,20,9,31]. The reason is that it gets the job done, albeit not optimally, but with relatively little investment in system development. In an industry where time to market is a crucial element of success, this is a true virtue [1] As a result, users sometimes have to revert to signup sheets as the actual ....

P. Krueger, T-H. Lai, and V. A. Dixit-Radiya, "Job scheduling is more important than processor allocation for hypercube computers". IEEE Trans. Parallel & Distributed Syst. 5(5), pp. 488--497, May 1994.

....directly, and to batch jobs that are submitted via a queueing system such as NQS. But this approach suffers from fragmentation, where processors cannot meet the requirements of the next queued job and therefore remain idle. As a result system utilization is typically in the range of 50 80 [12, 9, 4, 7]. It is well known that the best solutions for this problem are to use dynamic partitioning [11] or gang scheduling [3] However, these schemes have practical limitations. The only 1 efficient and widely used implementation of gang scheduling was the one on the CM 5 Connection Machine; other ....

P. Krueger, T-H. Lai, and V. A. Dixit-Radiya, "Job scheduling is more important than processor allocation for hypercube computers". IEEE Trans. Parallel & Distributed Syst. 5(5), pp. 488--497, May 1994.

No context found.

P. Krueger, T.-H. Lai, and V. Dixit-Radiya. Job scheduling is more important than processor allocation for hypercube computers. IEEE Trans. on Parallel and Distributed Systems, 5(5):488--497, 1994.

No context found.

P. Krueger, T.-H. Lai, and V. Dixit-Radiya. Job scheduling is more important than processor allocation for hypercube computers. IEEE Trans. on Parallel and Distributed Systems, 5(5):488--497, 1994.

No context found.

P. Krueger, T.-H. Lai, and V. Dixit-Radiya. Job scheduling is more important than processor allocation for hypercube computers. IEEE Trans. on Parallel and Distributed Systems, 5(5):488--497, 1994.

No context found.

P. Krueger, T. H. Lai, and V. A. Radiya, "Job Scheduling is More Important than Processor Allocation for Hypercube Computers," IEEE Transactions on Parallel and Distributed Systems, vol. 5, pp. 488 -- 497, May 1994.

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