## Predictive Application-Performance Modeling in a Computational Grid Environment (1999)

Citations: | 64 - 12 self |

### BibTeX

@MISC{Kapadia99predictiveapplication-performance,

author = {Nirav H. Kapadia and José A.B. Fortes and Carla E. Brodley},

title = {Predictive Application-Performance Modeling in a Computational Grid Environment},

year = {1999}

}

### Years of Citing Articles

### OpenURL

### Abstract

This paper describes and evaluates the application of three local learning algorithms --- nearest-neighbor, weighted-average, and locally-weighted polynomial regression --- for the prediction of run-specific resourceusage on the basis of run-time input parameters supplied to tools. A two-level knowledge base allows the learning algorithms to track short-term fluctuations in the performance of computing systems, and the use of instance editing techniques improves the scalability of the performance-modeling system. The learning algorithms assist PUNCH, a network-computing system at Purdue University, in emulating an ideal user in terms of its resource management and usage policies. 1. Introduction It is now recognized that the heterogeneous nature of the network-computing environment cannot be effectively exploited without some form of adaptive or demand-driven resource management (e.g., [10, 11, 12, 14, 18, 27]). A demand-driven resource management system can be characterized by its a...

### Citations

668 |
Robust Locally Weighted Regression and Smoothing Scatterplots
- Cleveland
- 1979
(Show Context)
Citation Context ...red to the 1-NN algorithm), the experiments in this paper use a three-point weighted average algorithm. 5.3. Locally Weighted Polynomial Regression Locally weighted regression (LWR) algorithms (e.g., =-=[2, 3, 9]-=-) t a surface to nearby points, typically via a locally linear or quadratic model. 2 With a linear (quadratic) model, the target concept is locally approximated by a linear (quadratic) surface. In ord... |

554 |
Local Polynomial Modelling and Its Applications
- Fan, Gijbels
- 1996
(Show Context)
Citation Context ...ons to short-term variations without being unduly a ected by them in the longer term. The application-performance modeling system for PUNCH [18, 19]: 1) employs locally weighted polynomial regression =-=[2, 9]-=-, allowing it to work with unknown feature weights and incomplete/noisy information, 2) addresses scalability issues by way ofacache that allows it to exploit the locality of runs [18] and by selectiv... |

480 | Locally weighted learning
- Atkeson, Schaal, et al.
- 1997
(Show Context)
Citation Context ...n of three local learning techniques for run-speci c resource-usage prediction. Local learning algorithms use available data in the region of a query to build a local model for the appropriate output =-=[2]-=-. For this domain, available data consists of measured resource-usage associated with preSchool of Electrical and Computer Engineering Purdue University West Lafayette, IN 47907-1285 fkapadia, fortes,... |

362 |
On the approximate realization of continuous mappings by neural networks
- Funahashi
- 1989
(Show Context)
Citation Context ...ttempt to establish an input-output mapping via a single function y = f(x; ), where is a nite-length parameter vector. While these methods can theoretically approximate any continuous function (e.g., =-=[13]-=-), they may not be appropriate for all tools. For example, semiconductor device simulation tools typically allow users to simulate a device in one, two, or three dimensions. In general, di erent solut... |

313 | editors, The Grid: Blueprint for a New Computing Infrastructure - Foster, Kesselman |

102 | Predicting application run times using historical information
- Smith, Foster, et al.
- 1998
(Show Context)
Citation Context .... Analytical expressions are also tool-speci c, and are generally provided by the administrators or the users of the tool. Examples of systems that utilize cumulative statistical data can be found in =-=[1, 8, 14, 15, 26, 27, 28]-=-. Such systems implicitly assume that a particular tool will exhibit similar resource-usage characteristics across runs | regardless of the input to the tool. 1 Although this assumption has been found... |

82 | Efficient locally weighted polynomial regression predictions
- Moore, Schneider, et al.
- 1997
(Show Context)
Citation Context ...nction yi = fi(x; i); the functions fi are kept as simple as possible. The problem now shifts to the selection of appropriate partitions for the learning system [25]. Non-parametric algorithms (e.g., =-=[2, 6, 22]-=-) address this issue by allowing the number of partitions (and consequently the number of parameters) to change dynamically. Instance-based learning (IBL) algorithms achieve this by recomputing a xed ... |

68 | Predicting Queue Times on Space-Sharing Parallel Computers
- Downey
- 1997
(Show Context)
Citation Context .... Analytical expressions are also tool-speci c, and are generally provided by the administrators or the users of the tool. Examples of systems that utilize cumulative statistical data can be found in =-=[1, 8, 14, 15, 26, 27, 28]-=-. Such systems implicitly assume that a particular tool will exhibit similar resource-usage characteristics across runs | regardless of the input to the tool. 1 Although this assumption has been found... |

63 |
Multiresolution instance-based learning
- Deng, Moore
- 1995
(Show Context)
Citation Context ...nction yi = fi(x; i); the functions fi are kept as simple as possible. The problem now shifts to the selection of appropriate partitions for the learning system [25]. Non-parametric algorithms (e.g., =-=[2, 6, 22]-=-) address this issue by allowing the number of partitions (and consequently the number of parameters) to change dynamically. Instance-based learning (IBL) algorithms achieve this by recomputing a xed ... |

60 | The relative performance of various mapping algorithms is independent of sizable variances in run-time predictions
- Armstrong, Hensgen, et al.
- 1998
(Show Context)
Citation Context .... Analytical expressions are also tool-speci c, and are generally provided by the administrators or the users of the tool. Examples of systems that utilize cumulative statistical data can be found in =-=[1, 8, 14, 15, 26, 27, 28]-=-. Such systems implicitly assume that a particular tool will exhibit similar resource-usage characteristics across runs | regardless of the input to the tool. 1 Although this assumption has been found... |

54 |
Predictability of process resource usage: A measurement-based study on UNIX
- Devarakonda, Iyer
- 1989
(Show Context)
Citation Context ... will exhibit similar resource-usage characteristics across runs | regardless of the input to the tool. 1 Although this assumption has been found to be valid for speci c tools and environments (e.g., =-=[7]-=-), it is not true in general. For instance, multiple users concurrently executing a given tool in a grid environment with radically di erent input data may cause the observed resource-usage to vary ra... |

50 | Scheduling resources in multi-user, heterogeneous, computing environments with SmartNet
- Freund, Gherrity, et al.
- 1998
(Show Context)
Citation Context ...ction It is now recognized that the heterogeneous nature of the network-computing environment cannot be effectively exploited without some form of adaptive or demand-driven resource management (e.g., =-=[10, 11, 12, 14, 18, 27]-=-). A demand-driven resource management system can be characterized by its ability to make automatic cost/performance tradeo decisions at run-time. Such decisions require that the infrastructure be abl... |

46 | Regression by local fitting: Methods, properties, and computational algorithms - Cleveland, Devlin, et al. - 1988 |

44 | Assessing the quality of learned local models
- Schaal, Atkeson
- 1994
(Show Context)
Citation Context ... i is approximated by an independent function yi = fi(x; i); the functions fi are kept as simple as possible. The problem now shifts to the selection of appropriate partitions for the learning system =-=[25]-=-. Non-parametric algorithms (e.g., [2, 6, 22]) address this issue by allowing the number of partitions (and consequently the number of parameters) to change dynamically. Instance-based learning (IBL) ... |

30 |
Numerical Recipes in C: The Art of Scienti c Computing. Cambridge Univ
- Press, Flannery, et al.
- 1992
(Show Context)
Citation Context ...sulted in lower prediction errors and faster learning, and 2) it required less time than higher order models to make a prediction. The regression equations were solved by singular value decomposition =-=[23]-=- to ensure numerical stability. 5.4. Parameter Selection The distance metric is common to all three algorithms. Given that the range and distribution of the values of the features are not known, a dis... |

29 | Mars: A Framework for Minimizing the Job Execution Time in a Metacomputing Environment, Future Generation Computer Systems
- Gehring, Reinefeld
- 1996
(Show Context)
Citation Context ...ction It is now recognized that the heterogeneous nature of the network-computing environment cannot be effectively exploited without some form of adaptive or demand-driven resource management (e.g., =-=[10, 11, 12, 14, 18, 27]-=-). A demand-driven resource management system can be characterized by its ability to make automatic cost/performance tradeo decisions at run-time. Such decisions require that the infrastructure be abl... |

29 |
Prediction-Based Dynamic Load-Sharing Heuristics
- Goswami, Devarakonda, et al.
- 1993
(Show Context)
Citation Context |

27 | SmartNet: A scheduling framework for heterogeneous computing
- Freund, Kidd, et al.
- 1996
(Show Context)
Citation Context ...ction It is now recognized that the heterogeneous nature of the network-computing environment cannot be effectively exploited without some form of adaptive or demand-driven resource management (e.g., =-=[10, 11, 12, 14, 18, 27]-=-). A demand-driven resource management system can be characterized by its ability to make automatic cost/performance tradeo decisions at run-time. Such decisions require that the infrastructure be abl... |

24 | On the design of a demand-based network-computing system: The Purdue University Network-Computing Hubs
- Kapadia, Fortes
- 1998
(Show Context)
Citation Context ...egression), and locally weighted regression. The issues involved in the selection of values for the parameters that de ne these methods are outlined in Section 5.4. Additional details can be found in =-=[17]-=-. 5.1. Nearest-Neighbor K-nearest neighbor (k-NN) algorithms [2] predict the output value(s) for a given query point by using an unweighted average of the output values of the k nearest instances as d... |

19 | An Intelligent Load Sharing Filter - Svensson, History - 1990 |

7 | Resource-usage prediction for demand-based network-computing
- Kapadia, Brodley, et al.
- 1998
(Show Context)
Citation Context |

4 |
The Computational Electronics Hub: A NetworkBased Simulation Laboratory
- Kapadia, Fortes, et al.
- 1996
(Show Context)
Citation Context ...hundred users, and provides access to forty research and commercial tools in computer architecture, parallel programming, and computational electronics. The described learning algorithms assist PUNCH =-=[16, 20]-=- in emulating an ideal user in terms of its resourcemanagement and usage policies. An ideal user is de ned as one who: 1) can predict the resourcerequirements of each run that he/she initiates, 2) pre... |

4 | Nonparametric regression for learning
- Schaal
- 1994
(Show Context)
Citation Context ...sitive to the structural complexity of the function to be learned [9] | consequently, they can be used for a wide range of tools. 4. Learning Algorithm Selection Global parametric learning algorithms =-=[24]-=- such as neural networks attempt to establish an input-output mapping via a single function y = f(x; ), where is a nite-length parameter vector. While these methods can theoretically approximate any c... |

4 |
Intelligent Job Selection for Distributed Scheduling
- Wang, Krueger, et al.
- 1993
(Show Context)
Citation Context |

3 |
Regression by local tting: Methods, properties, and computational algorithms
- Cleveland, Devlin, et al.
- 1988
(Show Context)
Citation Context ...weighted-average algorithms to the distribution of the observed instances. In general, the nearest-neighbor and weightedaverage algorithms cannot track (even linear) polynomial surfaces without error =-=[4]-=-. This is illustrated in Figure 3, which shows the prediction errors for the onenearest neighbor (1-NN), three-point weighted-average (3-Avg), and locally linear LWR (LLWR) algorithms on a synthetic d... |

3 |
Resource usage prediction for demand-based network-computing
- Kapadia, Brodley, et al.
- 1998
(Show Context)
Citation Context ...ted: 1) the order of the local polynomial, 2) the distance metric, 3) the kernel width, and 4) the kernel function. A locally linear polynomial was used for this research because empirical evaluation =-=[19]-=- showed that: 1) it resulted in lower prediction errors and faster learning, and 2) it required less time than higher order models to make a prediction. The regression equations were solved by singula... |

2 |
The PUNCH network desktop
- Kapadia, Fortes
- 1999
(Show Context)
Citation Context ...e corresponding run-speci c resource requirements. The run-speci c values of administratorspeci ed input parameters are automatically extracted by PUNCH from arguments and/or les supplied to the tool =-=[21]-=-. A two-level knowledge base [18, 19] allows the system to tailor its predictions to short-term variations in the performance of the tool and/or the computing environment. The instance-based learning ... |

2 |
anintelligent load sharing lter
- History
- 1990
(Show Context)
Citation Context |

1 |
Rejoinder to discussion of \smoothing by local regression: Principles and methods
- Cleveland, Loader
- 1996
(Show Context)
Citation Context ...ithmetic). 3 The kernel width (i.e., bandwidth) can be xed or variable. A xed kernel width is of limited use because it can lead to inaccurate or unde ned predictions in regions with low data-density =-=[5]. Giv-=-en the absence of a \complete" dataset in this domain, the kernel width was locally optimized [9] by recomputing it for each query. The kernel widths for the nearest-neighbor and weighted average... |

1 | Rejoinder to discussion of "smoothing by local regression: Principles and methods - Cleveland, Loader - 1996 |