Sample Generation Method Of I/O Trace Based On Generative Adversarial Network

I/O performance has become the bottleneck of performance improvement of the HPC systems.As the I/O request record in the running process of the system,I/O trace has been widely used in the field of I/O performance analysis.However,the traditional trace collection tools have shortcomings such as high resource consumption and limited scope of application,which restricts the improvement of the accuracy of I/O performance evaluation.Therefore,the trace generation tool came into being.On the other hand,as one of the most rapidly developing generation models,the generative adversarial network has been successfully applied in the fields of computer vision,encryption and security,and has great potential in the generation of real-valued sequence data.Aiming at the shortcomings of current trace generation tools such as limited application range,we combine the generation of confrontation networks with the field of trace generation,and propose a precise generation method for traces based on the generative adversarial network.Considering the diversity of I/O trace formats,we select core feature columns,and design corresponding preprocessing algorithms to convert trace data into trace vector data that can be processed by neural networks.On the basis of the existing generative adversarial network,we considered the timing characteristics of the trace vector and the characteristics of the generator,discriminator and loss function in the generative adversarial network,and designed a total of eight candidate architectures.Finally,the trace generation accuracy and generation performance of each candidate architecture are evaluated in terms of maximum mean discrepancy(MMD),trace accelerated replay time,and overhead and the architecture with the best generation effect is selected.The experimental results show that the generation architecture with the LSTM as the generator and discriminator and the cross-entropy function as the loss function has the best generation effect.The architecture has excellent generation effects on the six trace data sets,the MMD is stable between 0.012 and 0.058,and the error of the accelerated replay time between the generated trace and the real trace is between 0.08% and 3.23%.It proves that the trace generated by this architecture has good accuracy and can accurately reflect the workload characteristics of real applications so that the generated trace can be directly used for I/O performance analysis.