## Dynamic Dependence Analysis: A Novel Method for Data Dependence Evaluation (1992)

Venue: | Proc. Fifth Workshop on Languages and Compilers for Parallel Computing, LNCS 757 |

Citations: | 10 - 1 self |

### BibTeX

@INPROCEEDINGS{Petersen92dynamicdependence,

author = {Paul Petersen and David A. Padua},

title = {Dynamic Dependence Analysis: A Novel Method for Data Dependence Evaluation},

booktitle = {Proc. Fifth Workshop on Languages and Compilers for Parallel Computing, LNCS 757},

year = {1992},

pages = {64--81},

publisher = {Springer Verlag}

}

### OpenURL

### Abstract

This paper analyzes the results of applying data dependence calculations to the Perfect Benchmarks suite. We show that it is possible to measure the optimal parallelism, as defined by our model, and to compare the obtained parallelism for various data dependence tests with the optimal parallelism. Using these techniques, we find that a variation of Banerjee's inequalities is sufficient in all cases to obtain more than half of the available parallelism, and that a full integer-programming based dependence test does not contribute significantly to the measured parallelism. 1 Introduction A frequent topic of discussion concerning parallelizing compilers is the effectiveness of data dependence tests. Several recent studies have been published on this topic, all of which count the number of times that a test determines independence as a measure of the success of the test. It is commonly accepted that static evaluation is not sufficient, and additional measurements should be taken to accura...

### Citations

448 | The omega test: a fast and practical integer programming algorithm for dependence analysis
- Pugh
- 1993
(Show Context)
Citation Context ...roximation to integer programming, or some special case of integer programming. Therefore, it makes sense to use an integerprogramming based method in our evaluation. 6 The Omega test as described in =-=[Pug91]-=- is an extension of the Fourier-Motzkin linearprogramming algorithm to require integer solutions. The ability to simultaneously consider all subscripts of an array reference allows this dependence tes... |

347 |
Dependence Analysis for Supercomputing
- Banerjee
(Show Context)
Citation Context ...on of the coefficients of each equation, and ~ b are the righthand sides of the equations. A can always be factored into a unimodular matrix U and an echelon matrix D (with d 11 ? 0) such that UA = D =-=[Ban88]-=-. The system of diophantine equations described by ~xA = ~ b has a solution iff there exists an integer vector ~ t such what ~ tD = ~ b. Since D is an echelon matrix, the solution to ~ tD = ~ b is muc... |

120 |
Efficient and exact data dependence analysis
- Maydan, Hennessy, et al.
- 1991
(Show Context)
Citation Context ...is commonly accepted that static evaluation is not sufficient, and additional measurements should be taken to accurately assess the performance of data dependence calculations. To quote Maydan et al. =-=[MHL91]-=-: In a loop with a thousand independent pairs, being inexact in just one test could have a devastating effect on the amount of parallelism discovered. Ideally, one would like a standard model to measu... |

87 |
Dependence analysis for subscripted variables and its application to program transformations
- ALLEN
(Show Context)
Citation Context ...he loop limits. However, we are using an extended version of Banerjee's inequalities that assumes that when L k and U k are not compile-time constants, they take the values \Gamma1 and 1 respectively =-=[All83]-=-. The interpretation of 1 in this context is an unknown but arbitrarily large integer. 3.4 Omega Test Data dependence analysis is equivalent to deciding if an integer solution exists to a set of linea... |

62 | Measuring parallelism in computation-intensive scientific/engineering applications - Kumar - 1988 |

17 | MAXPAR: An execution driven simulator for studying parallel systems - Chen - 1989 |

15 | Machine-Independent Evaluation of Parallelizing Compilers
- Petersen, Padua
- 1992
(Show Context)
Citation Context ...s that the program may take. Tables 4 and 5 list the results of the dynamic analysis. The abbreviations in these tables are explained in Table 3. The first two columns in each table are obtained from =-=[PP92]-=-. The first column is the optimal parallelism present in each program. Optimal parallelism is defined in terms of the critical path length when assuming loop-level parallelism, as discussed in Section... |

9 |
The Perfect Club Benchmarks: Effective Performance Evaluation of Supercomputers
- Club
- 1989
(Show Context)
Citation Context ...Delta is a collection of SETL functions that implement a toolkit of the functions required to perform analysis, instrumentation, and parallelization of Fortran programs. We use the Perfect Benchmarks =-=[Per89]-=- as the source for the Fortran code used in this experiment. We determine the number of times each data dependence test is used statically, as well as the impact of each data dependence test on the ex... |

8 |
The Delta Program Manipulation system | Preliminary design
- Padua
- 1989
(Show Context)
Citation Context ...and parallelism, our ideal parallel machine serves as a generic model of available parallel architectures. The experiments described in this paper were done with the Delta program manipulation system =-=[Pad89]-=-. Delta is a collection of SETL functions that implement a toolkit of the functions required to perform analysis, instrumentation, and parallelization of Fortran programs. We use the Perfect Benchmark... |

7 | Experimental Evaluation of Some Data Dependence Tests
- Petersen, Padua
(Show Context)
Citation Context ...o tests uses the loop bounds or discriminates between direction vectors. We use two complex tests that include the effects of loop bounds and direction vectors: a variation on Banerjee's inequalities =-=[PP91]-=-, which has been extended to handle unknown lower and upper bounds, and the Omega test, which represents the integer-programming based dependence tests that are currently the most accurate data depend... |