Program Execution Time In Multi-core System Modeling Based On Generalized Additive Models

In present work,we tackle the problem of evaluating the timing impact from shared resource in multi-core embedded system.Due to the growing demand of complex functionality and performance,system-level design becomes a tendency for modern system design.On the other hand,for building faithful system-level model,the necessary back-annotating step poses a great demand on performance analysis,especially bounding the system timing behavior.Consequently,capturing the timing behavior in embedded system,especially in multi-core system,become necessary but challenging.The purpose of this paper is to build a mathematical model that can be used to predict the program contention delay in multi-core systems,given the basic hardware and software information in the early stages of system design.The input data of this model can be obtained through simulating software in hardware emulation platform,and finally returns the contention delay corresponding to the input data.The main works of this paper includes:modeling theory research,verification platform building.The modeling theory research includes:?1?the modeling processes of generalized additive models specific to predicting contention delay.?2?research on the influence of sampling granularity on the adjusted of obtained generalized additive models.The experimental results show that,as the sampling period increases,the adjusted R² of the generalized additive model is inversely proportionally reduced.?3?research on the influence of sampled data volume on obtained generalized additive models.The experimental results show that,as the data volum increases,the GCV value of the generalized additive model is inversely proportionally reduced.?4?Research on the influence on the generalized additive model when changing the concurrent executing software.The experimental results show that when the amount of data and the frequency of read/write operation of the parallel program are close to the measured program,the model has an adjusted R² above 0.9,but the model adjusted R² decreases linearly with the increase of the write data amount.The simulation platform built in this paper is based on a dual-core MicroBlaze system accessing shared memory through the AXI bus.The arbitration method is Round Robbin.The AES encryption software was used as the test program to explore the application scope of the model,and the program selected from the Testbench^[34]of the M?lardalen WCET research group was used as a parallel program.In summary,the contribution of this paper is to explore the application of the generalized additive model in the program execution time prediction of multi-core embedded systems.The generalized additive model can adapt to the fact that the conflict delay does not necessarily obey the normal distribution.Under the fine-grained sampling method,the adjustment of the prediction function fitted by the generalized additive model can reach more than 0.9.