Scalability Model for the LOFAR Direction Independent Pipeline

A.L. Varbanescu; A. Plaat; A.P. Mechev; H. Intema; H.J.A Rottgering; T.W. Shimwell

arxiv: 1906.11516 · v1 · pith:34P2W6JWnew · submitted 2019-06-27 · 🌌 astro-ph.IM

Scalability Model for the LOFAR Direction Independent Pipeline

A.P. Mechev , T.W. Shimwell , A. Plaat , H. Intema , A.L. Varbanescu , H.J.A Rottgering This is my paper

Pith reviewed 2026-05-25 14:22 UTC · model grok-4.3

classification 🌌 astro-ph.IM

keywords LOFARprefactor pipelinescalability modelprocessing timecalibration model sizeLoTSS surveyradio interferometry

0 comments

The pith

A model built from scaling tests predicts LOFAR prefactor pipeline processing times for varying CPU counts, data sizes, and calibration models.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tests the runtime of the LOFAR prefactor pipeline while changing the number of CPUs, the volume of data, and the size of the calibration sky model. It assembles the measurements into a model that forecasts how long any given combination of those parameters will take to process. Readers would care because the LoTSS survey must handle thousands of large data sets, so knowing processing times in advance helps schedule work and choose efficient settings. The tests also reveal that smaller calibration models reduce calibration time substantially while keeping output quality nearly the same.

Core claim

We present these results as a comprehensive model which will be used to predict processing time for a wide range of processing parameters. We also discover that smaller calibration models lead to significantly faster calibration times, while the calibration results do not significantly degrade in quality. Finally, we validate the model and compare predictions with production runs from the past six months, quantifying the performance penalties incurred by processing on a shared cluster.

What carries the argument

The empirical model derived from measurements of pipeline completion time as a function of CPU number, data size, and calibration sky model size.

Load-bearing premise

The scaling relationships measured in controlled tests continue to hold for full-scale LoTSS production runs on a shared cluster once performance penalties from resource contention are accounted for.

What would settle it

Compare the model's predicted processing time for a new LoTSS data set against the actual time measured in a production run; a large discrepancy beyond the observed shared-cluster overhead would falsify the model.

Figures

Figures reproduced from arXiv: 1906.11516 by A.L. Varbanescu, A. Plaat, A.P. Mechev, H. Intema, H.J.A Rottgering, T.W. Shimwell.

**Figure 1.** Figure 1: The major steps of the prefactor DI pipeline. 3.1. Processing Metrics The goal for our scalability model is to understand the effect of several parameters on the job completion time of LOFAR software. We do this by testing the processing time for various values of data size, number of CPUs used and sky model size. Averaging ratio Time averaging parameter (sec) Channels per Subband Averaged Size (Gb) 64x 8 … view at source ↗

**Figure 2.** Figure 2: The size of the sky model (measured in number of sources) [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 7.** Figure 7: The 0.3Jy model, here shown shaded in gray, is the one [PITH_FULL_IMAGE:figures/full_fig_p005_7.png] view at source ↗

**Figure 3.** Figure 3: Plots of the run time as a function of input data size plots. The agreement between our model and production runs are an encouraging result for future software performance modelling. Finally, we present [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

**Figure 6.** Figure 6: The run time of the gsmcal solve step as a function of the cutoff sensitivity is not linear. As shown in [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗

**Figure 5.** Figure 5: The processing time of the gsmcal solve step is linear with the size of the sky model as measured by the number of sources. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

**Figure 7.** Figure 7: Four images made using the wsclean software (Offringa et al., 2014) from the data set6 . The four images were calibrated with sky models of various flux cutoffs ranging from 0.05Jy (top left) to 1.5Jy (bottom right). Flux statistics for the green regions in the four images are listed in [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗

**Figure 8.** Figure 8: The processing time of the gsmcal solve step decreases exponentially with the number of CPUs requested. The model in Equation 3 is shown in a dashed line. As this is a 1/x model, it shows diminishing returns past 16 CPUs. # [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗

**Figure 10.** Figure 10: Test randomly submitting jobs to the GINA with different number of requested CPUs. The long tail for 8 and 16 CPU jobs shows that some jobs can take several hours to launch. [PITH_FULL_IMAGE:figures/full_fig_p011_10.png] view at source ↗

**Figure 11.** Figure 11: The queuing model built from two linear fits to the queu [PITH_FULL_IMAGE:figures/full_fig_p011_11.png] view at source ↗

**Figure 14.** Figure 14: Fit of an exponential model to the Download and Extrac [PITH_FULL_IMAGE:figures/full_fig_p011_14.png] view at source ↗

**Figure 17.** Figure 17: Processing time for the gsmcal solve step in a production environment. Data from this test ranges from 07/2018-01/2019. The dashed red line shows the prediction for a 1GB run, obtained from section 4.3. We see two distributions, which correspond to data averaged to 1GB and 512 MB. It should be noted that the left peak corresponds to 512MB data, as seen in [PITH_FULL_IMAGE:figures/full_fig_p012_17.png] view at source ↗

**Figure 18.** Figure 18: The scalability model for processing data through the [PITH_FULL_IMAGE:figures/full_fig_p012_18.png] view at source ↗

read the original abstract

LOFAR is a leading aperture synthesis telescope operated in the Netherlands with stations across Europe. The LOFAR Two-meter Sky Survey (LoTSS) will produce more than 3000 14 TB data sets, mapping the entire northern sky at low frequencies. The data produced by this survey is important for understanding the formation and evolution of galaxies, supermassive black holes and other astronomical phenomena. All of the LoTSS data needs to be processed by the LOFAR Direction Independent (DI) pipeline, prefactor. Understanding the performance of this pipeline is important when trying to optimize the throughput for large projects, such as LoTSS and other deep surveys. Making a model of its completion time will enable us to predict the time taken to process large data sets, optimize our parameter choices, help schedule other LOFAR processing services, and predict processing time for future large radio telescopes. We tested the prefactor pipeline by scaling several parameters, notably number of CPUs, data size and size of calibration sky model. We present these results as a comprehensive model which will be used to predict processing time for a wide range of processing parameters. We also discover that smaller calibration models lead to significantly faster calibration times, while the calibration results do not significantly degrade in quality. Finally, we validate the model and compare predictions with production runs from the past six months, quantifying the performance penalties incurred by processing on a shared cluster. We conclude by noting the utility of the results and model for the LoTSS Survey, LOFAR as a whole and for other telescopes.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper delivers a practical empirical runtime model for the LOFAR prefactor pipeline plus the concrete result that smaller sky models cut time without much quality loss, validated on real production data.

read the letter

The main point is that the authors measured how the prefactor pipeline scales with CPU count, data volume, and calibration sky-model size, then fitted a model they claim can predict run times across a range of settings. They also found that smaller sky models give much faster calibration with little quality drop, and they compared the model predictions to six months of actual LoTSS production runs while noting the extra cost of running on a shared cluster. That combination of controlled scaling tests and retrospective production check is the useful part. It gives operators numbers they can use for scheduling the rest of the survey and for choosing model sizes. The smaller-model observation is the kind of result that can be applied immediately. The validation step against real runs is better than pure lab scaling, because it at least quantifies some shared-resource penalty. The soft spot is the load-bearing assumption that the fitted relationships will keep working when cluster contention, I/O interference, or parameter combinations differ from the tested set. The retrospective check only shows consistency with the particular past runs; it does not prove the functional forms will remain predictive under changed load. The abstract gives no equations or error bars, so the full paper needs to show the exact model, how it was fitted, and prediction accuracy for the claim to be reusable by others. This is for people who run or schedule large LOFAR or similar radio pipelines. A reader who needs concrete guidance on LoTSS throughput or who is tuning calibration parameters will get direct value. It deserves peer review because the empirical grounding and production comparison are solid enough to be worth referee time, even if the model stays empirical.

Referee Report

3 major / 1 minor

Summary. The paper claims to derive an empirical scalability model for LOFAR's prefactor Direction Independent pipeline from controlled tests that vary CPU count, data volume, and calibration sky-model size. The model is presented as usable for predicting processing times across a wide parameter range for surveys such as LoTSS. A secondary claim is that smaller sky models yield significantly faster calibration with no significant quality degradation. The model is validated by comparing predictions to six months of actual production runs on a shared cluster, with quantification of contention penalties.

Significance. If the fitted scaling relationships prove robust and generalizable, the work would provide a practical tool for optimizing throughput and scheduling on LOFAR and similar future instruments. The inclusion of retrospective production validation supplies some independent grounding beyond the controlled tests, which is a methodological strength.

major comments (3)

[Abstract / model sections] Abstract and model-construction sections: the manuscript asserts a 'comprehensive model' for predicting processing time but supplies neither the explicit functional forms (how CPU, data-size, and sky-model terms combine) nor the fitting procedure or coefficient values; without these the central predictive claim cannot be evaluated or reproduced.
[Validation section] Production-validation section: the comparison with six months of LoTSS runs is retrospective and demonstrates consistency only for the observed load patterns; it does not test whether the functional forms remain predictive under different contention regimes or new parameter combinations, which is load-bearing for the claim that the model can be used for future production scheduling.
[Calibration-quality results] Calibration-quality claim: the statement that smaller sky models do not significantly degrade results lacks reported quantitative metrics (e.g., residual statistics, source counts, or statistical tests) and therefore does not support the 'no significant degradation' assertion at the level required for the secondary claim.

minor comments (1)

[Figures] Figure captions and axis labels should explicitly state the range of parameters tested and the number of repeated runs per point to allow readers to assess statistical reliability.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive and detailed review. The comments identify important areas for improving clarity, reproducibility, and rigor. We respond to each major comment below and indicate the planned revisions.

read point-by-point responses

Referee: [Abstract / model sections] Abstract and model-construction sections: the manuscript asserts a 'comprehensive model' for predicting processing time but supplies neither the explicit functional forms (how CPU, data-size, and sky-model terms combine) nor the fitting procedure or coefficient values; without these the central predictive claim cannot be evaluated or reproduced.

Authors: We agree that the submitted manuscript did not include explicit functional forms, the fitting procedure, or coefficient values, even though scaling trends were shown via figures and text. This omission limits reproducibility of the central claim. In the revised manuscript we will add a dedicated subsection that states the combined functional form (time as a function of CPU count, data volume, and sky-model size), describes the fitting method applied to the controlled-test data, and reports the resulting coefficients. revision: yes
Referee: [Validation section] Production-validation section: the comparison with six months of LoTSS runs is retrospective and demonstrates consistency only for the observed load patterns; it does not test whether the functional forms remain predictive under different contention regimes or new parameter combinations, which is load-bearing for the claim that the model can be used for future production scheduling.

Authors: The validation is retrospective and reflects only the contention patterns present during the six-month LoTSS production period. We accept that this does not constitute a test of the functional forms under substantially different regimes or new parameter combinations. In revision we will expand the discussion to state these limitations explicitly and qualify the forward-looking claims about scheduling use, while retaining the value of the existing consistency check under observed conditions. revision: partial
Referee: [Calibration-quality results] Calibration-quality claim: the statement that smaller sky models do not significantly degrade results lacks reported quantitative metrics (e.g., residual statistics, source counts, or statistical tests) and therefore does not support the 'no significant degradation' assertion at the level required for the secondary claim.

Authors: The original text relied on qualitative assessment of calibration outputs. We agree that quantitative metrics are needed to support the claim at the required level. In the revised manuscript we will add comparisons of residual-image statistics (RMS noise) and source-detection counts across the tested sky-model sizes, together with any available statistical summary, to substantiate the assertion of no significant degradation. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical model grounded in measurements and externally validated

full rationale

The paper constructs its scalability model directly from controlled experimental runs that measure completion time while varying CPU count, data volume, and calibration sky-model size. These measurements constitute independent input data. The resulting functional forms are then compared against six months of separate production-run logs on the shared cluster, providing an external benchmark that is not part of the fitting process. No equation or claim reduces by construction to a prior fit, self-citation, or renamed input; the validation step supplies falsifiable grounding outside the original test data. The derivation is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The work rests on empirical fitting of runtime data rather than first-principles derivation; the model coefficients are fitted parameters and the generalization from test to production runs is an untested domain assumption.

free parameters (1)

model coefficients for CPU, data size, and sky model scaling
The comprehensive model is built from measured runtimes, implying fitted coefficients whose exact values and functional form are not stated in the abstract.

axioms (1)

domain assumption Pipeline runtime scales predictably and continuously with the tested parameters across the full range needed for LoTSS.
The model is presented as usable for a wide range of parameters; this continuity and predictability is assumed rather than derived.

pith-pipeline@v0.9.0 · 5838 in / 1305 out tokens · 32789 ms · 2026-05-25T14:22:10.200520+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

14 extracted references · 14 canonical work pages · 2 internal anchors

[1]

doi: https://doi.org/10.1016/j.future.2013.07

ISSN 0167-739X. doi: https://doi.org/10.1016/j.future.2013.07

work page doi:10.1016/j.future.2013.07 2013
[2]

URL https://doi.org/10.5281/zenodo.1487962. H. Intema, P. Jagannathan, K. Mooley, and D. Frail. The GMRT 150 MHz all-sky radio survey-ﬁrst alternative data release TGSS ADR1. Astronomy & Astrophysics , 598:A78,

work page doi:10.5281/zenodo.1487962
[3]

Jones, T

E. Jones, T. Oliphant, P. Peterson, et al. SciPy: Open source sci- entiﬁc tools for Python, 2001–. URL http://www.scipy.org/. [Online; accessed June 28, 2019]. S. Kavulya, J. Tan, R. Gandhi, and P. Narasimhan. An anal- ysis of traces from a production mapreduce cluster. In 2010 10th IEEE/ACM International Conference on Cluster, Cloud and Grid Computing , ...

work page 2001
[4]

doi: 10.1109/ CCGRID.2010.112. S. Kazemi, S. Yatawatta, S. Zaroubi, P. Lampropoulos, A. De Bruyn, L. Koopmans, and J. Noordam. Radio interferometric calibration using the sage algorithm. Monthly Notices of the Royal Astro- nomical Society, 414(2):1656–1666,

work page 2010
[5]

1109/MSP.2009.934719. C. Marco, C. Fabio, D. Alvise, G. Antonia, G. Alessio, G. Francesco, M. Alessandro, M. Elisabetta, M. Salvatore, and P. Luca. The glite workload management system. In Journal of Physics: Conference Series, volume 219, page 062039. IOP Publishing,

work page arXiv 2009
[6]

Mechev, J

A. Mechev, J. B. R. Oonk, A. Danezi, T. W. Shimwell, C. Schrijvers, H. Intema, A. Plaat, and H. J. A. Rottgering. An Automated Scalable Framework for Distributing Radio Astronomy Process- ing Across Clusters and Clouds. In Proceedings of the Interna- tional Symposium on Grids and Clouds (ISGC) 2017, held 5-10 March, 2017 at Academia Sinica, Taipei, Taiwan...

work page 2017
[7]

doi: 10.1093/mnras/stu1368. H. Sanjay and S. Vadhiyar. Performance modeling of parallel appli- cations for grid scheduling. Journal of Parallel and Distributed Computing, 68(8):1135 – 1145,

work page doi:10.1093/mnras/stu1368
[8]

doi: https://doi.org/10.1016/j.jpdc.2008.02.006

ISSN 0743-7315. doi: https://doi.org/10.1016/j.jpdc.2008.02.006. URL http://www. sciencedirect.com/science/article/pii/S0743731508000464. T. Shimwell, H. R¨ ottgering, P. N. Best, W. Williams, T. Dijkema, F. De Gasperin, M. Hardcastle, G. Heald, D. Hoang, A. Horneﬀer, et al. The LOFAR Two-metre Sky Survey-I. Survey description and preliminary data release...

work page doi:10.1016/j.jpdc.2008.02.006 2008
[9]

T. W. Shimwell, C. Tasse, M. J. Hardcastle, A. P. Mechev, W. L. Williams, P. N. Best, H. J. A. R¨ ottgering, J. R. Callingham, T. J. Dijkema, F. de Gasperin, D. N. Hoang, B. Hugo, M. Mirmont, J. B. R. Oonk, I. Prandoni, D. Raﬀerty, J. Sabater, O. Smirnov, R. J. van Weeren, G. J. White, M. Atemkeng, L. Bester, E. Bon- nassieux, M. Br¨ uggen, G. Brunetti, K...

work page internal anchor Pith review Pith/arXiv arXiv
[10]

Templon and J

J. Templon and J. Bot. The dutch national e-infrastructure. To ap- pear in Proceedings of Science edition of the International Sym- posium on Grids and Clouds (ISGC) 2016 13-18 March 2016, Academia Sinica, Taipei, Taiwan, Oct

work page 2016
[11]

org/10.5281/zenodo.163537

URL https://doi. org/10.5281/zenodo.163537. G. van Diepen and T. J. Dijkema. DPPP: Default Pre-Processing Pipeline. Astrophysics Source Code Library, Apr

work page doi:10.5281/zenodo.163537
[12]

URL http://lofar.ie/wp-content/ uploads/2018/03/station_data_cookbook_v1.2.pdf. W. Williams, R. Van Weeren, H. R¨ ottgering, P. Best, T. Dijkema, F. de Gasperin, M. Hardcastle, G. Heald, I. Prandoni, J. Sabater, et al. LOFAR 150-MHz observations of the Bo¨ otes ﬁeld: catalogue and source counts. Monthly Notices of the Royal Astronomical Society, 460(3):2385–2412,

work page 2018
[13]

C. Witt, M. Bux, W. Gusew, and U. Leser. Predictive performance modeling for distributed computing using black-box monitoring and machine learning. CoRR, abs/1805.11877,

work page internal anchor Pith review Pith/arXiv arXiv
[14]

doi: https://doi.org/ 10.1006/jpdc.1996.0151

ISSN 0743-7315. doi: https://doi.org/ 10.1006/jpdc.1996.0151. URL http://www.sciencedirect.com/ science/article/pii/S0743731596901513. L. T. Yang, X. Ma, and F. Mueller. Cross-platform performance prediction of parallel applications using partial execution. In Pro- 16 ceedings of the 2005 ACM/IEEE conference on Supercomputing , page

work page doi:10.1006/jpdc.1996.0151 1996

[1] [1]

doi: https://doi.org/10.1016/j.future.2013.07

ISSN 0167-739X. doi: https://doi.org/10.1016/j.future.2013.07

work page doi:10.1016/j.future.2013.07 2013

[2] [2]

URL https://doi.org/10.5281/zenodo.1487962. H. Intema, P. Jagannathan, K. Mooley, and D. Frail. The GMRT 150 MHz all-sky radio survey-ﬁrst alternative data release TGSS ADR1. Astronomy & Astrophysics , 598:A78,

work page doi:10.5281/zenodo.1487962

[3] [3]

Jones, T

E. Jones, T. Oliphant, P. Peterson, et al. SciPy: Open source sci- entiﬁc tools for Python, 2001–. URL http://www.scipy.org/. [Online; accessed June 28, 2019]. S. Kavulya, J. Tan, R. Gandhi, and P. Narasimhan. An anal- ysis of traces from a production mapreduce cluster. In 2010 10th IEEE/ACM International Conference on Cluster, Cloud and Grid Computing , ...

work page 2001

[4] [4]

doi: 10.1109/ CCGRID.2010.112. S. Kazemi, S. Yatawatta, S. Zaroubi, P. Lampropoulos, A. De Bruyn, L. Koopmans, and J. Noordam. Radio interferometric calibration using the sage algorithm. Monthly Notices of the Royal Astro- nomical Society, 414(2):1656–1666,

work page 2010

[5] [5]

1109/MSP.2009.934719. C. Marco, C. Fabio, D. Alvise, G. Antonia, G. Alessio, G. Francesco, M. Alessandro, M. Elisabetta, M. Salvatore, and P. Luca. The glite workload management system. In Journal of Physics: Conference Series, volume 219, page 062039. IOP Publishing,

work page arXiv 2009

[6] [6]

Mechev, J

A. Mechev, J. B. R. Oonk, A. Danezi, T. W. Shimwell, C. Schrijvers, H. Intema, A. Plaat, and H. J. A. Rottgering. An Automated Scalable Framework for Distributing Radio Astronomy Process- ing Across Clusters and Clouds. In Proceedings of the Interna- tional Symposium on Grids and Clouds (ISGC) 2017, held 5-10 March, 2017 at Academia Sinica, Taipei, Taiwan...

work page 2017

[7] [7]

doi: 10.1093/mnras/stu1368. H. Sanjay and S. Vadhiyar. Performance modeling of parallel appli- cations for grid scheduling. Journal of Parallel and Distributed Computing, 68(8):1135 – 1145,

work page doi:10.1093/mnras/stu1368

[8] [8]

doi: https://doi.org/10.1016/j.jpdc.2008.02.006

ISSN 0743-7315. doi: https://doi.org/10.1016/j.jpdc.2008.02.006. URL http://www. sciencedirect.com/science/article/pii/S0743731508000464. T. Shimwell, H. R¨ ottgering, P. N. Best, W. Williams, T. Dijkema, F. De Gasperin, M. Hardcastle, G. Heald, D. Hoang, A. Horneﬀer, et al. The LOFAR Two-metre Sky Survey-I. Survey description and preliminary data release...

work page doi:10.1016/j.jpdc.2008.02.006 2008

[9] [9]

T. W. Shimwell, C. Tasse, M. J. Hardcastle, A. P. Mechev, W. L. Williams, P. N. Best, H. J. A. R¨ ottgering, J. R. Callingham, T. J. Dijkema, F. de Gasperin, D. N. Hoang, B. Hugo, M. Mirmont, J. B. R. Oonk, I. Prandoni, D. Raﬀerty, J. Sabater, O. Smirnov, R. J. van Weeren, G. J. White, M. Atemkeng, L. Bester, E. Bon- nassieux, M. Br¨ uggen, G. Brunetti, K...

work page internal anchor Pith review Pith/arXiv arXiv

[10] [10]

Templon and J

J. Templon and J. Bot. The dutch national e-infrastructure. To ap- pear in Proceedings of Science edition of the International Sym- posium on Grids and Clouds (ISGC) 2016 13-18 March 2016, Academia Sinica, Taipei, Taiwan, Oct

work page 2016

[11] [11]

org/10.5281/zenodo.163537

URL https://doi. org/10.5281/zenodo.163537. G. van Diepen and T. J. Dijkema. DPPP: Default Pre-Processing Pipeline. Astrophysics Source Code Library, Apr

work page doi:10.5281/zenodo.163537

[12] [12]

URL http://lofar.ie/wp-content/ uploads/2018/03/station_data_cookbook_v1.2.pdf. W. Williams, R. Van Weeren, H. R¨ ottgering, P. Best, T. Dijkema, F. de Gasperin, M. Hardcastle, G. Heald, I. Prandoni, J. Sabater, et al. LOFAR 150-MHz observations of the Bo¨ otes ﬁeld: catalogue and source counts. Monthly Notices of the Royal Astronomical Society, 460(3):2385–2412,

work page 2018

[13] [13]

C. Witt, M. Bux, W. Gusew, and U. Leser. Predictive performance modeling for distributed computing using black-box monitoring and machine learning. CoRR, abs/1805.11877,

work page internal anchor Pith review Pith/arXiv arXiv

[14] [14]

doi: https://doi.org/ 10.1006/jpdc.1996.0151

ISSN 0743-7315. doi: https://doi.org/ 10.1006/jpdc.1996.0151. URL http://www.sciencedirect.com/ science/article/pii/S0743731596901513. L. T. Yang, X. Ma, and F. Mueller. Cross-platform performance prediction of parallel applications using partial execution. In Pro- 16 ceedings of the 2005 ACM/IEEE conference on Supercomputing , page

work page doi:10.1006/jpdc.1996.0151 1996