After conducting the first tests and proving our hypothesis on POSDAO at the beginning of 2020, GBC.AI team followed up working on implementing machine learning / deep learning algorithms to optimize the fast-growing Solana blockchain.
Experiments
The previously used pipeline was deliberately not changed to test it in the conditions of different blockchain architecture. We have chosen Solana blockchain for our second blockchain guardian. Two main conceptions are important for proper network functioning — performance and resilience. Performance can be measured by three straightforward metrics — throughput, which is measured by transactions per second, transaction confirmation time, and drop rate — a fraction of non-confirmed transactions.
Similar to our approach with POSDAO, a fault-tolerant simulation model was developed and built. Additionally, we formed a list of different attack scenarios like Clock Frequency, Missing Node, Disconnected Node, Administrative Partitioning and others. The scenarios were sequentially simulated within the local network to generate the most representative sample for training the AI-guardian recommendation algorithm of the GBC v0.2 framework.
In total, while working with Solana blockchain, we have conducted more than 1500 experiments, generated 1.500.000 blocks, training data of more than 12Tb were generated. Of the 89 identified significant dynamic parameters X in the course of the frequency analysis, 6 most significant (for example page_size, account_storage_overhead, gossip_ping_cache_capacity, etc) were detected. Subsequently, the trained recommendation model made a forecast for target parameters based on its analysis of the selected parameters.
Achievements
We deployed the simulation infrastructure of the autogenerating cluster with several nodes and transaction sender worker. After that, we launched the same simulations in a loop on different configs. After that, we made a comparison table and calculated the mean and deviation in each group and saw confirmation of the model’s work by increasing the throughput (~28%) on average and decreasing the drop rate (~11 pp) on the independent hardware.
Results
Our approach showed its usefulness again. GBC.AI team decided to get ready for launching an independent AI module for the test-net of the public blockchain. We went further and started developing GBC v0.3 framework.
Conclusion
Experiments with POSDAO and Solana blockchain allowed us to prove that:
— constant blockchain parameters are not optimal
— build in adaptive parameters selection is a promising solution
In process of building models we formulated machine learning subtasks for parameters selection and showed approach feasibility within two blockchains.
From the very beginning, we aimed to create a blockchain agnostic solution. Such an approach allows us to work with any PoS blockchain. We need to conduct deep research and deeply understand blockchain specifics, but in general process with different blockchains is relatively the same.
Solana is an open-source project implementing a new, high-performance, permissionless blockchain. Created by former chip engineers from Qualcomm it’s designed to process 50,000 transactions per second and create blocks in 400ms. The overarching goal of the Solana software is to demonstrate that there is a possible set of software algorithms using the combination to create a blockchain. The system can support an upper bound of 710,000 TPS on a standard gigabit network and 28.4 million tps on a 40-gigabit network.
