Predicting the performance of the 2D SOME-Bus with the deep learning
4.9 Results and discussion
We compare the results obtained with the MFANN neural network and the SVR-RBF model of [11]. These two models represent the best models they have obtained. Table 4.6, Table 4.7, Table 4.8, Table 4.9, and Table 4.10 summarize the performance of our models and those of [11]
respectively for the prediction of the average channel utilization, average channel waiting time, average network latency, average processor utilization and average input waiting time, with regard toR,RM SE,M AE,RAE(%),RRSE(%).
Table 4.6 – Results for prediction of average channel utilization.
Performance measures R MAE RMSE RAE (%) RRSE (%) Our model 0.98 0.02 0.03 16.00 16.40 SVR-RBF [11] 0.92 0.05 0.07 37.12 37.44 MFANN [11] 0.90 0.06 0.08 40.66 43.87
Table 4.7 – Results for prediction of average channel waiting time.
Performance measures R MAE RMSE RAE (%) RRSE (%) Our model 0.98 30.38 76.64 9.30 20.5 SVR-RBF [11] 0.97 42.73 82.62 13.04 21.63
MFANN [11] 0.96 58.88 96.99 17.97 25.39
Table 4.8 – Results for prediction of average network latency.
Performance measures R MAE RMSE RAE (%) RRSE (%) Our model 0.97 211.764 278.441 23.31 24.72 SVR-RBF [11] 0.97 184.60 277.96 19.76 23.31 MFANN [11] 0.96 245.09 331.83 26.24 27.83
Performance measures R MAE RMSE RAE (%) RRSE (%) Our model 0.99 0.02 0.03 16.00 16.40 SVR-RBF [11] 0.94 0.05 0.06 34.96 35.73 MFANN [11] 0.91 0.06 0.07 35.40 40.01
Table 4.10 – Results for prediction of average input waiting time.
Performance measures R MAE RMSE RAE (%) RRSE (%) Our model 0.99 25.71 34.05 13.20 14.70 SVR-RBF [11] 0.97 39.48 53.66 19.96 22.95 MFANN [11] 0.96 49.39 63.04 24.97 26.96
The results of the different prediction models have given decreasing RMSE as the number of epochs increases and the batch size is changed. These results are referenced on the convergence curves of Figure 4.3a, Figure 4.3b, Figure 4.3c, Figure 4.3d, and Figure 4.3e.
100 200 300 400 500
100 200 300 400 500 600
Epochs
It comes out from the results obtained the following observations:
• overall, these results show that our model outperforms the models MFANN and SVR-RBF ofAkay and al.[11] for predicting the average channel utilization, average channel waiting time, average processor utilization, and average input waiting time;
• in the case of the average network latency our model surpasses the MFANN model of [11]
but remains generally equivalent to the SVR-RBF model of the same author. Indeed, for
• our ANN model has the lowest error on all predictions compared to other models;
• the lowest error is obtained for the prediction of the average channel utilization and aver-age processor utilization (RM SE = 0.03);
• the R for all our models is≥0.97while the best models of [11] starts from 0.92;
• the training phase duration of our models varies between 45 and 70 seconds and their execution time on the test set is negligible;
• the lowestRM SE was obtained for of average processor utilization and average channel utilization thanks to the fact that their values range is between 0 and 1;
• the highest RM SE was obtained for average network latency because the gap between the minimum and maximum values of the average network latency is the highest com-pared to that of the other performance measures.
Conclusion
The 2-dimensional Simultaneous Optical Multiprocessor Exchange Bus (2D SOME-Bus) with its high bandwidth and low latency becomes a realistic choice element to provide high comput-ing power. To ensure its effectiveness, it is necessary to evaluate its performance. Speakcomput-ing of performance it is necessary for such an architecture to reliably evaluate different metrics such as average channel utilization, average channel latency, average network latency, average CPU usage and average latency of inputs . This assessment is usually done analytically or statisti-cally and has shortcomings such as the relatively high cost in time.
In this document, we use artificial intelligence methods, specifically deep learning, based on a set of data made available to us and containing performance measures obtained using predic-tion variables such as the report. between the message transfer time and the processing time, the number of nodes, the number of threads, the spatial and temporal distributions to build models able to evaluate these performance measures. Our models show good results on the data set and our results are better than those obtained in the state of the art while being in-expensive in time. This shows that neural networks in particular and artificial intelligence in general can represent an interesting alternative in the prediction of performance measurements of multiprocessor architectures. Interesting perspectives are being considered, such as applying other artificial intelligence methods to the problem, testing the prediction of performance mea-sures using real data, and applying the techniques to other types of multi-media architectures.
processors.
[1] SCHMIDHUBER J.,Deep learning in neural networks : An overview. The Swiss AI Lab IDSIA.
2014.
[2] GABELI F.Algorithms and Parallel Computing. First Ed., Wiley, New York, 2011.
[3] KATSINIS C., NABET B. A Scalable Interconnection Network Architecture for Petaflops Computing. The Journal of Supercomputing. 2004, 27(2), pp. 103–128, DOI :10.1023/B :SUPE.0000009318.91562.b0.
[4] YORAM E. Methods for Performance Evaluation of Parallel Computer Systems.Tech. rep.
Columbia University Academic Commons, 1986,DOI :10.7916/D8959RKP.
[5] KURIAN J. L., Performance Evaluation : Techniques, Tools and Benchmarks. The University of Texas at Austin. 2004.
[6] LAXMI N., BHUYAN, XIODONG Z.,Tutorial on Multiprocessor Performance Measurement and Evaluation. 1994.
[7] AKAY M. F., C. Katsinis. Performance Evaluation of a Broadcast-based Distributed Shared Me-mory Multiprocessor by OPNET Modeler. 2013.
[8] GENBRUGGE D., EECKHOUT L. Statistical simulation of chip multiprocessors running multi-program workloads. In :Proceedings of 25th International Conference on Computer De-sign, IEEE, 2007, 10 pp. 464– 471.
[9] ZAYID E.I.M., AKAY M.F. Predicting the performance measures of a message-passing mul-tiprocessor architecture using artificial neural networks.Neural Computing and Applications, 2012, 23(7-8), pp. 2481– 2491,DOI :10.1007/s00521-012-1267-9.
[10] AKAY M.F., ABASIKELES¸ I. Predicting the performance measures of an optical distribu-ted shared memory multiprocessor by using support vector regression.Expert Systems with Applications. 2010, 37(9), pp. 6293–6301,DOI :10.1016/j.eswa.2010.02.092.
Network World. 2015, pp 241–265DOI :10.14311/NNW.2015.25.013.
[12] ACI C.I., AKAY M.F. A new congestion control algorithm for improving the performance of a broadcast-based multiprocessor architecture.Journal of Parallel and Distributed Compu-ting. 2010, 70(9) pp. 930–940DOI :10.1016/j.jpdc.2010.06.003.
[13] PATTERSON J., GIBSON A. Deep Learning A Practitioner’s Approach. O’REILLY, First Ed., 2017.
[14] BUDUMA N.Fundamentals of Deep Learning Designing Next-Generation Machine Intelligence Algorithms. O’REILLY, First Ed., 2017.
[15] KIRK M.Thoughtful Machine Learning with Python. O’REILLY, First Ed., 2017.
[16] KONSTANTINOS T., KOTAS S., Dimitrios S., Axel J.Designing 2D and 3D Network-on-Chip Architectures. Springer, First Ed., 2014.
[17] MUHAMMAD Q., STEPHEN J.Multicore Technology : Architecture, Recon guration, and Modeling. CRC Press., 2014.
[18] DALLY W.J., TOWLES B.Principles and Practices of Interconnection Networks. Morgan Kauf-mann Publishers Inc., 2004.
[19] ADIGA N.R., BLUMRICH M.A., CHEN D., COTEUS P., GARA A., GIAMPAPA M.E., HEI-DELBERGER P., SINGH S., STEINMACHER-BUROW B.D., TAKKEN T., TSAO M., VRA-NAS P.Blue Gene/L torus interconnection network.IBM Journal of Research and Development.
2005, 49(2.3), pp. 265–276,DOI: 10.1147/rd.492.0265.
[20] OPNET Modeler. OPNET Technologies. Disponible depuis : www.opnet.com. Consulté le 18 Septembre 2018.
[21] UCI Machine Learning Repository. Disponible depuis : http://archive.ics.uci.
edu. Consulté le 18 Septembre 2018.
[22] Ubuntu. Canonical Group. Disponible depuis :https://www.ubuntu.com/. Consulté le 18 Septembre 2018.
[23] Python. Python Software Fundation. Disponible depuis :https://www.python.org/. Consulté le 19 Septembre 2018.
[24] Anconda. Anaconda Distribution. Disponible depuis :https://www.anaconda.com/. Consulté le 19 Septembre 2018.
[25] Tensorflow. Google. Disponible depuis :https://www.tensorflow.org/. Consulté le 19 Septembre 2018.
[26] Keras. Python package. Disponible depuis : https://keras.io/. Consulté le 19 Sep-tembre 2018.
[27] Scikit-learn. Python package. Disponible depuis : https://scikit-learn.org/. Consulté le 19 Septembre 2018.
[28] KDNuggets. Disponible depuis : https://www.kdnuggets.com/2018/05/poll-tools-analytics-data-science-machine-learning-results.html/. Consulté le 28 Septembre 2018.