DOI: 10.17587/prin.17.261-269

A Method for Determining a Rational Computational Load in Heterogeneous Data-Stream Processing Systems

R. Yu. Demkin, Researcher, demkin.rdi-kvant@bk.ru, Research and Development Institute "Kvant", Moscow, 125438, Russian Federation

Corresponding author: Roman Yu. Demkin, Researcher, Research and Development Institute "Kvant", Moscow, 125438, Russian Federation, E-mail: demkin.rdi-kvant@bk.ru

The paper considers the issue of load distribution on each of the used types of computing devices in a heterogeneous computing system. An analytical model of the mutual influence of various types of computing devices as part of a heterogeneous computer is proposed. Based on the considered model, an approach to load distribution is formed, which can provide acceleration of data processing compared to homogeneous solutions. The results of experiments for heterogeneous systems consisting of CPU and GPU are presented, demonstrating the relationship between CPU resource utilization and GPU performance, based on which the parameters of the proposed analytical model were cross-validated. The application of the model for the rational selection of the number of full-weight threads on CPU and GPU for solving real sets of problems is demonstrated. The advantage of the proposed approach in terms of the speed of tasks solving compared to independently executed homogeneous solutions is experimentally shown. Thus, the article proposes a comprehensive method for evaluating and optimizing the load on a heterogeneous computer, describes all the necessary metrics and their analytical dependencies, and provides experimental evidence of the practical significance of the proposed approach.

Keywords: heterogeneous computing system, GPU, computational workload, weight thread (WT), lightweight thread (LWT), operations per second (OPS), peak computational performance, load distribution

pp. 261—269

For citation:
Demkin R. Yu. A Method for Determining a Rational Computational Load in Heterogeneous Data-Stream Processing Systems, Programmnaya Ingeneria, 2026, vol. 17, no. 5, pp. 261—269. DOI: 10.17587/prin.17.261-269.

References:

1. Korneev V. V. Parallel Programming, Programmnaya Ingeneria, 2022, vol. 13, no. 1, pp. 3—16. DOI: 10.17587/prin.13.3-16 (in Russian).
2. Zahran M. Heterogeneous Computing: Here to Stay, Communication of ACM, 2017, vol. 60, no. 3, pp. 42—45. DOI: 10.1145/3024918.
3. Vouillon J. Lwt: a cooperative thread library, Workshop on ML, 2008, рр. 3—12. DOI: 10.1145/1411304.1411307.
4. Draper N., Smith H. Applied Regression Analysis. 3rd edition, Moscow, Dialektica, 2017, 912 p. (in Russian).
5. Chatterjee S., Simonoff J. S. Handbook of regression analysis. NJ., John Wiley & Sons, 2013, 252 p. DOI: 10.1002/9781118532843.
6. Oppenheim A. V., Schafer R. W. Digital Signal Processing, Moscow, Tehnosfera, 2012, 1048 p. (in Russian).
7. Kalman R. E. A New Approach to Linear Filtering and Prediction Problems, ASME Journal of Basic Engineering, 1960, vol. 82, no. 1, pp. 35—45. DOI: 10.1115/1.3662552.
8. Bianchi G., Sorrentino R. Electronic filter simulation & design. NY., McGraw-Hill Professional, 2007, 606 p.
9. Akaike H. A new look at the statistical model identifica­tion, IEEE Transactions on Automatic Control, 1974, vol. 19, no. 6, pp. 716—723. DOI: 10.1109/TAC.1974.1100705.
10. Efron B. Bootstrap Methods: Another Look at the Jack-knife, Annals of Statistics, 1979, vol. 7, no. 1, pp. 1—26. DOI: 10.1214/AOS/1176344552.
11. Hwang K. Advanced Computer Architecture: Parallelism, Scal­ability, Programmability. NY., McGraw-Hill Professional, 1993, 771 p.
12. Wang Z., Yang J., Melhem R. et al. Simultaneous Multikernel GPU: Multi-tasking throughput processors via fine-grained sharing, IEEE International Symposium on High Performance Computer Architecture (HPCA), 2016, vol. 1, pp. 358—369. DOI: 10.1019/HPCA.2016.7446078.