Micro-architecture Characterization And Performance Optimization Of Blockchain Systems

Blockchain is a new type of computing model,which exihibits many advantages,such as decentralization,distribution and traceability.Blockchain has played a vital role by being deployed in many key industries,such as banking and logistics,and has already attracted widespread attention worldwide.However,the poor performance of the current blockchain system is still a critical issue that needs to be resolved.Only by understanding the characteristics of the blockchain system at the micro-architecture level can we understand the reasons for the poor performance of the blockchain.However,we do not understand the characteristics of the micro-architecture of the blockchain system.The key to understanding the micro-architecture features of a blockchain system is to analyze these micro-architecture events.What is worse,the fact that there are more than 200 micro-architecture events makes it extremely difficult to understand the characteristics.We even lack a systematic approach to identify the important events from a larger number of ones and in turn focus on them.We propose a novel methodology to characterize and benchmark blockchain systems at micro-architecture level.The key is to leverage fuzzy set theory to select the important micro-architecture events after the importance of them is quantified by a machine learning based approach.The selected important events for single programs are used to characterize the programs while the selected common important events form an importance vector which is used to measure the similarity between benchmarks.We also perform correlation analysis between these importance vector and the configuration parameters of the blockchain system,and finally improve the performance of the blockchain system by adjusting the configuration parameters with higher correlation with the importance vector.We use this method to characterize seven benchmarks from Blockbench and six ones from the Caliper benchmark suite.The results show that our method reveals five interesting findings.Moreover,by using the importance characterization results,I increase the transaction throughput of Hyperledger Fabric by 70% as well as decrease the transaction latency by 55%.In addition,we find that three of the seven benchmarks from Blockbench and two of the six ones from Caliper are redundant,respectively.