#### DMCA

## Workload characteristics of a multi-cluster supercomputer (2004)

### Cached

### Download Links

- [www.liacs.nl]
- [www.cs.huji.ac.il]
- [www.cs.huji.ac.il]
- DBLP

### Other Repositories/Bibliography

Citations: | 78 - 8 self |

### Citations

161 |
Queuing Theory with Computer Science Applications
- Probability
- 1990
(Show Context)
Citation Context ...ed several statistical models to fit the interarrival times of representative and high load period, including hyperexponential, gamma, Weibull, and heavytailed distributions like lognormal and Pareto =-=[17]-=-. We fit the above mentioned distributions (except hyperexponential) using Maximum Likelihood Estimation (MLE) method, and a two-phase hyperexponential distribution using Expectation Maximization (EM)... |

141 | The Workload on Parallel Supercomputers: Modeling the Characteristics of Rigid Jobs
- Lublin, Feitelson
(Show Context)
Citation Context ...stribution and the sample’s empirical distribution). lation rate and cancellation lags are analyzed and modeled, since it may also affect the scheduling process. 4.1 Job Arrival Rate As is studied in =-=[7]-=-, job arrivals are expected to have cycles at three levels: daily, weekly, and yearly. In a yearly cycle, we find that workloads are not distributed evenly throughout the year. Instead, workloads conc... |

104 | Job Characteristics of a Production Parallel - Feitelson, Nitzberg - 1995 |

82 | Workload characterization: A survey
- Calzarossa, Serazzi
- 1993
(Show Context)
Citation Context ...on Workload characterization of parallel supercomputers is important to understand the system performance and develop workload models for evaluating different system designs and scheduling strategies =-=[1, 2]-=-. During the past several years, lots of workload data has been collected [3], analyzed [4–6], and modeled [7–9]. Benchmarks and standards are also proposed for job scheduling on parallel computers [1... |

77 | A comprehensive model of the supercomputer workload, in
- Cirne, Berman
- 2001
(Show Context)
Citation Context ...le). Some characteristics, such as distributions and correlations, vary across different workloads [4, 5, 11]. Other characteristics are studied and reported separately, such as job cancellation rate =-=[9]-=- and conditional distributions (e.g. actual runtime distributions conditioned on requested runtime [4]). In this paper we compare our workload with previous reported ones on a per characteristics basi... |

71 | Using queue time predictions for processor allocation
- Downey
- 1997
(Show Context)
Citation Context ...fs4 2427 6.4 lognormal (µ = 5.3, σ = 2.5) 0.13 Table 6. Job actual runtimes on DAS-2 clusters. Different kinds of distributions have been used to model the actual runtime, for instance, loguniform in =-=[22]-=-, hypergamma in [7] and Weibull in [4]. We evaluate gamma, lognormal and Weibull distributions for actual runtimes on DAS-2 clusters. Figure 7 (b) shows the distribution fitting on fs0. Weibull and lo... |

57 | Benchmarks and standards for the evaluation of parallel job schedulers - Chapin, Cirne, et al. - 1999 |

56 | Workload Modeling for Performance Evaluation
- Feitelson
- 2002
(Show Context)
Citation Context ...on Workload characterization of parallel supercomputers is important to understand the system performance and develop workload models for evaluating different system designs and scheduling strategies =-=[1, 2]-=-. During the past several years, lots of workload data has been collected [3], analyzed [4–6], and modeled [7–9]. Benchmarks and standards are also proposed for job scheduling on parallel computers [1... |

51 | The elusive goal of workload characterization
- Downey, Feitelson
- 1999
(Show Context)
Citation Context ...nd have different arrival patterns in different periods (e.g. peak and none-peak hours in a daily cycle). Some characteristics, such as distributions and correlations, vary across different workloads =-=[4, 5, 11]-=-. Other characteristics are studied and reported separately, such as job cancellation rate [9] and conditional distributions (e.g. actual runtime distributions conditioned on requested runtime [4]). I... |

46 | Modeling of Workload in MPPs - Jann, Pattnaik, et al. - 1997 |

37 | Fitting world-wide web request traces with the EM-algorithm. Performance Evalutation, 52:175
- Khayari, Sadre, et al.
- 2003
(Show Context)
Citation Context ...above mentioned distributions (except hyperexponential) using Maximum Likelihood Estimation (MLE) method, and a two-phase hyperexponential distribution using Expectation Maximization (EM) algorithm 6 =-=[18]-=-. The goodness of fit is assessed using the Kolmogorov-Smirnov test. Results of distribution fitting are shown in Table 2 and 3. Figure 3 and 4 further illustrate how well the different distributions ... |

30 | Characteristics of a Large Shared Memory Production Workload
- Chiang, Vernon
- 2001
(Show Context)
Citation Context ...nd have different arrival patterns in different periods (e.g. peak and none-peak hours in a daily cycle). Some characteristics, such as distributions and correlations, vary across different workloads =-=[4, 5, 11]-=-. Other characteristics are studied and reported separately, such as job cancellation rate [9] and conditional distributions (e.g. actual runtime distributions conditioned on requested runtime [4]). I... |

24 | A comparison of workload traces from two production parallel machines
- Windisch, Lo, et al.
- 1996
(Show Context)
Citation Context ...aries from 1.3 to 3.0 on DAS-2 clusters. The mean and CV are considerably larger in the representative period (see Table 3). Both small (1-2) and large CVs (3-6) have been reported in other workloads =-=[4, 6]-=-. We have selected several statistical models to fit the interarrival times of representative and high load period, including hyperexponential, gamma, Weibull, and heavytailed distributions like logno... |

21 | A measurement-based simulation study of processor co-allocation in multicluster systems
- Banen, Bucur, et al.
(Show Context)
Citation Context ...eration. This research serves as a realistic basis in modeling cluster workloads, which contributes as input for evaluations of different scheduling strategies in a multi-cluster research environment =-=[13]-=-. The rest of the paper is organized as follows. Section 2 provides an overview of the DAS-2 system and workload traces used in our study. Section 3 analyzes the overall system utilization. Section 4 ... |

19 | Construction and use of multiclass workload models
- Calzarossa, Serazzi
- 1994
(Show Context)
Citation Context ...ted on other workloads [11]. Since the user/group structure have an significant impact on the workload modeling, techniques and models have been proposed to capture the user behavior in the workloads =-=[24]-=-. We are also investigating multi-class models on other cluster workloads which are strongly group/VO (Virtual Organization) oriented. 8 The common group and user accounts are mapped onto all five clu... |

1 |
Feitelson Memory usage in the LANL CM-5 Workload
- G
- 1997
(Show Context)
Citation Context ...e Spearman’s r coefficients show positive correlations. This indicates that larger size jobs (using more processors) tend to use more memory than smaller jobs. Similar characteristics are reported in =-=[23]-=-. Correlations between memory usage per processor and job size have two folds on DAS-2 clusters. On fs1-3 small positive correlations are observed, while on fs0 and fs4, weak inverse correlations are ... |