Performance optimization of hybrid GNN–LLM pipelines for code comparison
DOI:
https://doi.org/10.55630/mem.2026.55.098-106Keywords:
performance optimization, GNN, LLM, code comparison, parallel processing, caching, semantic normalizationAbstract
This study proposes a set of techniques for optimizing the performance of hybrid pipelines combining large language models (LLMs) and graph neural networks (GNN) for comparison of program code fragments. On the basis on previous work [1] on semantic normalization and graph embeddings, the main computational bottlenecks are identified and solutions are proposed including parallel processing, dynamic skipping of redundant operations, caching, and optimization of graph representation. The article contributes to the practical implementation of semantic-based systems for comparing code in real conditions.
References
Damyanov D., Semantic normalization on graph models searching structure matchings in code, Mathematics and Education in Mathematics (to appear).
Donchev, I., An alternative to virtual functions in modern C++. Application in training, Mathematics, Computer Science and Education, Vol. 8, Issue 2, (2025), pp. 144-159
Dong X., Gu Y., Sun Y., Wang L., PASGAL: Parallel And Scalable Graph Algorithm Library, (2024), https://doi.org/10.48550/arXiv.2404.17101
Gale T., Elsen E., Hooker S., The State of Sparsity in Deep Neural Networks, (2019), DOI: 10.48550/arXiv.1902.09574
Hoefler T., Alistarh D., Dryden T., Peste A., Sparsity in Deep Learning: Pruning and growth for efficient, Journal of Machine Learning Research 23, (2021), pp. 1-124, DOI: 10.48550/arXiv.2102.00554
Holm M., Scientific Computing on Hybrid Architectures, Department of Information Technology, Uppsala University, ISSN 1404-5117, (2013), Sweden
Karbowski A., Amdahl’s and Gustafson-Barsis laws revisited, Pomiary Automatyka Robotyka, vol. 29(4), (2025), pp. 35-38, https://doi.org/10.48550/arXiv.0809.1177
Pugacheva D., Ermakov A., Lyskov I., Igor M., Zotov Y., Enhancing GNNs Performance on Combinatorial Optimization by Recurrent Feature Update, Cornell University, (2024), DOI: 10.48550/arXiv.2407.16468.
Petrov M., Dochkova J., Comparison of the PCA and FLD Approaches in Glial Tumors Classification Systems, Baltic J. Modern Computing, Vol. 13, No. 1, (2025), pp. 221-232, https://doi.org/10.22364/bjmc.2025.13.1.12
Ramisetty S., Chandrasekaran T., Eruvaram V. K., Pulicharla M. R., Optimizing Real-Time Data Pipelines for Machine Learning: A Comparative Study of Stream Processing Architectures, World Journal of Advanced Research and Reviews, (2024), pp.1653–1660, DOI: 10.30574/wjarr.2024.23.3.2818
Yang C., Buluc A., Owens J., GraphBLAST: A High-Performance Linear Algebra-based Graph Framework on the GPU, ACM Transactions on Mathematical Software (TOMS), Volume 48, Issue 1, (2022), pp.1-51, https://doi.org/10.1145/3466795
Yin L., Li G., Fang M., Shen L., Huang T., Wang Z., Menkovski V., Ma X., Pechenizkiy M., Liu S., Dynamic Sparsity Is Channel-Level Sparsity Learner, 37th Conference on Neural Information Processing Systems NeurIPS, (2023), pp. 67993 – 68012, https://doi.org/10.48550/arXiv.2305.19454
