This work aims to explore different methods for transaction parallelization on the MultiversX blockchain, combining existing solutions from other blockchains with various machine learning techniques.
The study evaluates multiple clustering algorithms, with the main objective of finding a deterministic method for determining the number of groups of independent transactions.
During the experiment, the algorithms were tested in terms of both stability and execution time, measured across different scenarios: number of transactions, number of shared addresses, software libraries used for matrix operations, and the hardware architectures employed.
The experiment also proposes an architectural solution—multiple components are designed and tested to enable processing transactions in parallel, even under conditions where conflicts may occur.
The work also presents an approach to reduce the number of conflicts that may arise when simulating parallel execution, which is common in the case of smart-contract transactions.
The study tests a neural network that predicts, based on a transaction’s information, the resources accessed by that transaction.
Keywords: blockchain, machine learning, clustering, smart contracts, neural network.