D. G. Feitelson and L. Rudolph. Toward convergence in job schedulers for parallel supercomputers. In D. G. Feitelson and L. Rudolph, editors, Job Scheduling Strategies for Parallel Processing, volume 1162 of LNCS, pages 1-26. Springer, 1996.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....overall system performance gets better (both response time and utilization) but it still does not prevent from fragmentation and free slots in the schedule. The draw back of gang scheduling is that changes to the OS are often necessary. This was not suitable in our case. In Feitelson and Rudolph [6] a classification of jobs in rigid, evolving, moldable and malleable was done. In Feitelson et al. 7] malleable jobs are defined as jobs which number of assigned processors may change during execution as a result of the system assigning additional resources or requiring to release some. Whereas ....

D. G. Feitelson and L. Rudolph. Towards convergence in job schedulers for parallel supercomputers. Lecture Notes in Computer Science, 1162:1--26, 1996.

....jobs. The application catches these signals and adjusts the number of threads accordingly. Applications capable of changing the number of processes (or threads) during runtime are called malleable, whereas the number of processes of moldable applications can only be chosen at the program start [10]. Applications which need a xed number of processes are called rigid. To extend a moldable application into a malleable one it is necessary to nd a point in the algorithm where the number of threads can be changed easily. It should be reached with a reasonable frequency, e.g. within the ....

Feitelson, D.G., Rudolph, L.: Toward convergence in job schedulers for parallel supercomputers. In Feitelson, D.G., Rudolph, L., eds.: Job Scheduling Strategies for Parallel Processing. Volume 1162 of LNCS. Springer (1996) 1-26

....component size i is p i = q i =Q if i is not a power of 2 and p i = 3q i =Q if i is a power of 2, with Q such that the sum of the probabilities equals 1, and with q = 0:95. This distribution favours small sizes, and sizes that are powers of two, which has been found to be a realistic choice [8]. We will consider four cases for the structure of jobs, which are differentiated by the flexibility of their requests: 1. An ordered request is represented by a tuple of C values (r 1 ; r 2 ; r C ) each generated from distribution D. The positions of the request components in the ....

D.G.Feitelson and L.Rudolph. Toward Convergence in Job Schedulers for Parallel Supercomputers. In Job Scheduling Strategies for Parallel Processing, Lecture Notes in Computer Science 1162, pages 1-26. Springer-Verlag, 1996.

....the job requests is specified by a user or a compiler. A job scheduler dispatches the job to the requested number of processors. Each job is exclusively executed until its completion on these processors. These jobs are called rigid jobs and this kind of scheduling is called pure space sharing [3]. The pure space sharing among rigid jobs has several advantages such that implementation is simple and any user can have optimal performance by executing the job on the requested number of processors exclusively. Although more complex scheduling schemes such as gang scheduling and adaptive space ....

D. G. Feitelson and L. Rudolph. Toward Convergence in Job Schedulers for Parallel Supercomputers. In Job Scheduling Strategies for Parallel Processing, Lecture Notes in Computer Science 1162, pages 1--26. Springer-Verlag, 1996.

....being processed. For the case of sequential jobs, i.e. jobs that require exactly one processor for execution, the involved scheduling problems have been studied intensively for decades [BEP 96] But in many situations the problem arises to find a schedule for a set of parallel jobs [FR95, FR96, BEP 96] Such a set could, for example, be a parallel query execution plan generated by the query optimizer of a parallel database management system [Rah96, GI97] The model studied in this paper assumes that each parallel job demands a fixed number of processors or a specified sub system ....

Dror G. Feitelson and Larry Rudolph. Toward Convergence in Job Schedulers for Parallel Supercomputers. In Dror G. Feitelson and Larry Rudolph, editors, Job Scheduling Strategies for Parallel Processing (IPPS' 96 Workshop, Honolulu, HI), LNCS 1162, pages 1--26, Berlin, 1996. Springer.

....granularities, context switch times, and time slice granularities. Keywords: distributed resource management, parallel job scheduling, distributed operating systems, co scheduling, gang scheduling. 1. Introduction The scheduling of parallel jobs has long been an active area of research [7, 8]. It is a challenging problem because This work was supported by the U.S. Dept. of Energy through Los Alamos National Laboratory contract W 7405 ENG 36. the performance and applicability of parallel scheduling algorithms is highly dependent upon factors at different levels: the workload, the ....

D. G. Feitelson and L. Rudolph. Toward Convergence in Job Schedulers for Parallel Supercomputers. In D. G. Feitelson

....in increasing the overall performance in the presence of load imbalance and communication intensive workloads. Keywords: Paralllel Job Scheduling, Distributed Operating Systems, Communication Protocols. 1 Introduction The scheduling of parallel jobs has long been an active area of research [10, 11]. It is a challenging problem because the performance and applicability of parallel scheduling algorithms is highly dependent upon factors at di erent levels: the workload, the parallel programming language, the operating system (OS) and the machine architecture. Time sharing scheduling ....

Dror G. Feitelson and Larry Rudolph. Toward Convergence in Job Schedulers for Parallel Supercomputers. In D. G. Feitelson and L. Rudolph, editors, Job Scheduling Strategies for Parallel Processing, volume 1162 of Lecture Notes in Computer Science. Springer-Verlag, 1996.

....and is relatively insensitive to the local process scheduling strategy. Keywords: Parallel Job Scheduling, Distributed Operating Systems, Communication Protocols. 1 Introduction The scheduling of parallel jobs across a parallel or distributed system has long been an active area of research [8, 9]. It is a challenging problem because the performance and applicability of parallel scheduling algorithms is highly dependent upon factors at di erent levels: workload, parallel programming language, operating system (OS) and machine architecture. Time sharing scheduling algorithms are ....

Dror G. Feitelson and Larry Rudolph. Toward Convergence in Job Schedulers for Parallel Supercomputers. In Dror G. Feitelson and Larry Rudolph, editors, Job Scheduling Strategies for Parallel Processing, volume 1162 of Lecture Notes in Computer Science. Springer-Verlag, 1996.

....in increasing the overall performance in the presence of load imbalance and communication intensive workloads. Keywords: Parallel Job Scheduling, Distributed Operating Systems, Communication Protocols. 1 Introduction The scheduling of parallel jobs has long been an active area of research [6, 7]. It is a challenging problem because the performance and applicability of parallel scheduling algorithms is highly dependent upon factors at different levels: workload, parallel programming language, operating system (OS) and machine architecture. Time sharing scheduling algorithms are ....

Dror G. Feitelson and Larry Rudolph. Toward Convergence in Job Schedulers for Parallel Supercomputers. In D. G. Feitelson and L. Rudolph, editors, Job Scheduling Strategies for Parallel Processing, volume 1162 of Lecture Notes in Computer Science. Springer-Verlag, 1996.

....Here, S denotes job size. In the second policy, a job scheduler does not guarantee the job to be executed on S processors, that is, the number of processors to execute the job depends on a congestion level of jobs on the parallel computer. A job executed in the first policy is called a rigid job [10]. In this paper, all jobs that arrive at the shared job queue will be executed as rigid jobs. 2.3 Workloads A job assumed in this paper has three parameters: 1) the number of processors that a job requests, or job size, 2) execution time and (3) arrival time. 2.3.1 Job Size Recent analyses ....

D. G. Feitelson and L. Rudolph. Toward Convergence in Job Schedulers for Parallel Supercomputers. In Job Scheduling Strategies for Parallel Processing, Lecture Notes in Computer Science 1162, pages 1--26. Springer-Verlag, 1996.

....with respect to response time, wait time, run time slowdown, and processor utilization. Keywords: Parallel Job Scheduling, Distributed Operating Systems, Communication Protocols, Performance Evaluation. 1 Introduction The scheduling of parallel jobs has long been an active area of research [8, 9]. It is a challenging problem because the performance and applicability of parallel scheduling algorithms is highly dependent upon factors at di erent levels: the workload, the parallel programming language, the operating system (OS) and the machine architecture. The importance of job scheduling ....

Dror G. Feitelson and Larry Rudolph. Toward Convergence in Job Schedulers for Parallel Supercomputers. In Dror G. Feitelson and Larry Rudolph, editors, Job Scheduling Strategies for Parallel Processing, volume 1162 of Lecture Notes in Computer Science. Springer-Verlag, 1996.

....job by assigning a higher weight to it and consequently accepting higher costs. On the other hand makespan scheduling is closely related to the overall use of a multiprocessor. There, individual job weights are ignored. Thus, it generally reflects the goal of the multiprocessor s owner, see also [5]. In this paper we address the problem of off line scheduling of parallel and independent jobs with invariant resource requirements. Makespan scheduling of parallel and independent jobs is similar to bin packing and has been addressed frequently in the past. For instance Garey and Graham [6] have ....

D.G. Feitelson and L. Rudolph. Towards convergence in job schedulers for parallel supercomputers. In D.G. Feitelson and L. Rudolph, editors, IPPS'96 Workshop: Job Scheduling Strategies for Parallel Processing, pages 1--26. Springer--Verlag, Lecture Notes in Computer Science 1162, 1996.

....for predicting queue times on space sharing parallel computers, and a model for using historical information to predict the run times of parallel jobs. The goal of this paper is to evaluate the usefulness of these predictions for processor allocation. 1. 1 Adaptive Jobs Feitelson and Rudolph [8] propose the following classification of adaptive parallel jobs: rigid jobs can run only on a fixed cluster size; moldable jobs can be configured to run on a range of cluster sizes, but once they begin execution, they cannot change cluster size. Evolving jobs change cluster size as they execute; ....

Dror G. Feitelson and Larry Rudolph. Towards convergence in job schedulers for parallel supercomputers. In D. G. Feitelson and L. Rudolph, editors, Job Scheduling Strategies for Parallel Processing, Springer-Verlag LNCS Vol 1162, pages 1--26, April 1996.

....average slowdown of all jobs. 1 Introduction A parallel job is composed of a number of concurrently executing processes, which collectively perform a certain computation. A rigid parallel job has a xed number of processes (referred to as the job s size) which does not change during execution [2]. To execute such a parallel job, the job s processes are mapped to a set of processors using a one to one mapping. In a non preemptive regime, these processors are then dedicated to running this job until such time that it terminates [3] The set of processors dedicated to a certain job is called ....

D. G. Feitelson and L. Rudolph, "Toward Convergence in Job Schedulers for Parallel Supercomputers". In Job Scheduling Strategies for Parallel Processing, D. G. Feitelson and L. Rudolph (eds.), Springer-Verlag, Lect. Notes Comput. Sci. Vol. 1162, pp. 1-26, 1996.

....to find its functional dependence on system load, as in Fig. 2. This means that many different load conditions should be checked. With rigid jobs, load and utilization are completely determined by the arrival rate, so it is easy and justifiable to present the results as a function of utilization [11]. But with adaptive or dynamic partitioning schemes, changing the partition size may change the efficiency of job execution and thus change the utilization. The correct load variable is therefore the arrival rate, not the resulting utilization. This has the unfortunate consequence that it becomes ....

D. G. Feitelson and L. Rudolph, "Toward convergence in job schedulers for parallel supercomputers". In Job Scheduling Strategies for Parallel Processing, D. G. Feitelson and L. Rudolph (eds.), pp. 1--26, Springer-Verlag, 1996. Lect. Notes Comput. Sci. vol. 1162.

....for swapping, the general trend is to retain the same framework, and moreover, to cast it into a standard. Many studies, however, show that more flexibility in both the scheduler actions and the way programs make use of parallelism result in better performance. But there is hope for convergence [25]. For example, theoretical analysis underscores the effectiveness of preemption in achieving low average response times, and also shows that considerable benefits are possible if the scheduler is allowed to tailor the partition sizes in accordance with the current system load. Notably, much of ....

....The following roughly classifies these models according to five criteria: 1. Partition Specification Each parallel job is executed in a partition that consists of a number of processors. The size of such a partition may depend on the multiprocessor, the application, and the load of multiprocessor [25]. Moreover, the size of the partition of a specific job may change during the lifetime of this job in some models: Fixed. The partition size is defined by the system administrator and can be modified only by reboot. Variable. The partition size is determined at submission time of the job ....

[Article contains additional citation context not shown here]

D. G. Feitelson and L. Rudolph, "Toward convergence in job schedulers for parallel supercomputers". In Job Scheduling Strategies for Parallel Processing, D. G. Feitelson and L. Rudolph (eds.), pp. 1--26, Springer-Verlag, 1996. Lecture Notes in Computer Science Vol. 1162.

....to intricate workloads, where each job s parallelism and runtime may interact; thus a more extensive modeling e ort is needed than in trac models in which packets have a prede ned size, and only the arrival process is not known. Rigid jobs are jobs that do not change their parallelism at runtime [8]. We chose not to address malleable or dynamic jobs at this stage due to the complexities involved, and the lack of data, as described below. 1.1 Why Model There are two common ways to use a recorded workload to analyze or evaluate a system design: 1) use the logged workload directly to drive ....

....but not with the large jobs in the third group. 14 0 5 10 15 0 0.2 0.4 0.6 0.8 1 SDSC95 log runtime [1 , 4] 5 , 16] 17 , 400] 0 5 10 15 0 0.2 0.4 0.6 0.8 1 SDSC96 log runtime [1 , 4] 5 , 16] 17 , 400] 0 5 10 15 0 0.2 0.4 0.6 0. 8 1 KTH log runtime [1 , 3] [4 , 8] [9 , 400] 0 5 10 15 0 0.2 0.4 0.6 0.8 1 LANL log runtime [32 , 32] 33 , 128] 129,1024] Figure 4: Distributions of runtimes for di erent job sizes in the four systems. 6.2 Modeling Approaches Given that there is a connection between the parallelism and the runtime, the question ....

[Article contains additional citation context not shown here]

D. G. Feitelson and L. Rudolph, \Toward convergence in job schedulers for parallel supercomputers". In Job Scheduling Strategies for Parallel Processing, D. G. Feitelson and L. Rudolph (eds.), pp. 1-26, Springer-Verlag, 1996. Lect. Notes Comput. Sci. vol. 1162.

....rather wide gap between theoretical studies and schemes proposed by academia on the one hand, and what is done in practice in large installations on the other hand. The reasons for such divergence are that the assumptions leading to and justifying the different schemes are usually quite different [205, 206]. Thus, while all the schemes are designed to solve the general problem of how to schedule parallel jobs on a parallel machine , the detailed assumptions make each instance of this problem distinct from the others. Nat9 p p t p t (a) b) c) p p p t t 1 2 3 1 2 4 t 3 t 4 ....

....by the system software (preemption, migration, swapping) and the amount of information available to the scheduler [206] Another source of differences is assumptions about the jobs and their capabilities in relation to processor allocation. The following classification has been proposed (Fig. 2) [205]: Rigid: jobs that require a certain predefined number of processors. They will not run on less, and will not utilize more. The system has no choice but to grant the requested number. 10 Moldable: jobs that allow the number of processors to be set at the outset, but it cannot change thereafter. ....

D. G. Feitelson and L. Rudolph, "Toward convergence in job schedulers for parallel supercomputers ". In Job Scheduling Strategies for Parallel Processing, D. G. Feitelson and L. Rudolph (eds.), pp. 1--26, Springer-Verlag, 1996. Lecture Notes in Computer Science Vol. 1162.

....rather wide gap between theoretical studies and schemes proposed by academia on the one hand, and what is done in practice in large installations on the other hand. The reasons for such divergence are that the assumptions leading to and justifying the different schemes are usually quite different [123, 124]. Thus, while all the schemes are designed to solve the general problem of how to schedule parallel jobs on a parallel machine , the detailed assumptions make each instance of this problem distinct from the others. Naturally, this leads to different solutions. Regrettably, it also means that it ....

....by the system software (preemption, migration, swapping) and the amount of information available to the scheduler [124] Another source of differences is assumptions about the jobs and their capabilities in relation to processor allocation. The following classification has been proposed (Fig. 2) [123]: Rigid: jobs that require a certain predefined number of processors. They will not run on less, and will not utilize more. The system has no choice but to grant the requested number. Moldable: jobs that allow the number of processors to be set at the outset, but it cannot change thereafter. ....

D. G. Feitelson and L. Rudolph, "Toward convergence in job schedulers for parallel supercomputers ". In Job Scheduling Strategies for Parallel Processing, D. G. Feitelson and L. Rudolph (eds.), pp. 1--26, Springer-Verlag, 1996. Lecture Notes in Computer Science Vol. 1162.

....of the most important resources are computing cycles and memory locations. The allocation of computing cycles allows for some tradeoff between the degree of parallelism and time moldable and malleable jobs may use less processors for more time to accumulate the same overall number of cycles [10]. With memory, such a tradeoff is only possible if paging is used. As paging is typically considered to be too expensive due to its overhead and adverse effect on communication and synchronization, parallel jobs typically have to be memory resident throughout their execution. Memory requirements ....

D. G. Feitelson and L. Rudolph, "Toward convergence in job schedulers for parallel supercomputers". In Job Scheduling Strategies for Parallel Processing, D. G. Feitelson and L. Rudolph (eds.), pp. 1--26, Springer-Verlag, 1996. Lecture Notes in Computer Science Vol. 1162.

No context found.

D. G. Feitelson and L. Rudolph. Toward convergence in job schedulers for parallel supercomputers. In D. G. Feitelson and L. Rudolph, editors, Job Scheduling Strategies for Parallel Processing, volume 1162 of LNCS, pages 1-26. Springer, 1996.

No context found.

D. G. Feitelson, L. Rudolph, "Toward Convergence in Job Schedulers for Parallel Supercomputers", Job Scheduling Strategies for Parallel Processing, pp. 1-26. Springer-Verlag 1996. Lectures Notes in Computer Science vol. 1162.

No context found.

D.G. Feitelson and L. Rudolph. Toward Convergence in Job Schedulers for Parallel Supercomputers. In D.G. Feitelson and L. Rudolph, editors, 2nd Workshop on Job Scheduling Strategies for Parallel Processing, volume 1162 of LNCS, pages 1--26. SpringerVerlag, 1996.

No context found.

D. G. Feitelson, L. Rudolph, "Toward Convergence in Job Schedulers for Parallel Supercomputers", Job Scheduling Strategies for Parallel Processing, Lectures Notes in Computer Science vol. 1162, pp. 1-26, Springer-Verlag 1996.

No context found.

D.G.Feitelson and L.Rudolph. Toward Convergence in Job Schedulers for Parallel Supercomputers. In Job Scheduling Strategies for Parallel Processing, Lecture Notes in Computer Science 1162, pages 1-26. SpringerVerlag, 1996.

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