Research On Real-time Performance Optimization Technology Of Multi-core System Considering Application Scenarios

Multi-core chip system is the mainstream development direction in the industry.However,the current multi-core chip system will encounter three major problems in the actual operation.Firstly,due to the failure of Dennard's scaling law,the chip will encounter serious heating problems,which will limit its performance and reduce its reliability.Secondly,because advanced chips have a large number of cores,if only the chip temperature is considered for the decentralized assignment of tasks,the communication distance and delay between the system cores may be too large,and ultimately lead to the degradation of system performance.Finally,there are many types of application scenarios encountered when the computer system is running,and due to the different requirements of different application scenarios for system performance,the system will encounter the problem that the performance optimization strategy does not match the requirements.In order to solve these problems and optimize the overall performance of the system,a real-time performance optimization technique for multicore systems considering application scenarios is developed in this thesis.Firstly,in order to determine the current application scenarios of the system in realtime,the technique constructs and trains the Long-Short-Term-Memory network for application scenario identification based on the statistics data of the system performance counting monitor as input.Then,according to the characteristics of the identified application scenarios,this technology establishes a multi-core chip power model,a thermal model,and a performance model that considers the communication delay.By establishing an optimization model that comprehensively considers communication latency and thermal effects,and using greedy algorithms to approximate the optimization problem in real-time,this technology can assign tasks to the core of achieving optimal performance for multi-core systems.By establishing an optimization model that comprehensively considers communication latency and thermal effects,and using greedy algorithms to approximate the optimization problem in real-time,this technology can assign tasks to the core of achieving optimal performance for multi-core systems.For high-performance scenarios,this technology takes the absolute performance of the multicore system as the optimization goal and maximizes the overall operating frequency of the system under the condition that the chip temperature does not exceed the threshold.For the continuous scenario,this technology takes the operating energy efficiency ratio of the system as the optimization goal and adjusts the frequency of each core under the condition of considering the static power consumption to achieve the highest performance per watt of the system.For periodic scenarios,this technology aims at the shortest task execution time and improves the task execution speed under the condition that the average power of the core does not exceed the given power limit.Experimental results show that this technology can accurately identify application scenario types: with 12 data from the performance count monitor as input,the recognition accuracy of scenario categories for 6 application scenarios is up to 94.3%.By considering the chip thermal effect and communication delay effect,this technology can reasonably assign tasks to the core.Compared with the two methods without considering the communication delay,the performance of the proposed method is up to 39% and 28%,respectively.Finally,our experiments show that the performance of multi-core chip system is greatly improved in all aspects after using this technology: in the highperformance scenario,the system frequency is increased by more than 50%;In the continuous scenario,the system energy efficiency ratio increased by 18.2%;In the periodic scenario,the time to complete the same workload was reduced by 25.5%.