| With the rapid development of edge and cloud computing technologies,the types of data generated by applications increase.A related report pointed out that public cloud services will grow by 22% in 2022 and bring more heterogeneous types of data.Heterogeneous types of data require higher performance of storage systems,and require storage systems to provide differentiated services.Researches proposes a multi-level storage system that comprehensively utilizes the characteristics of different storage media to deal with the differences in access mode,size,life cycle and other characteristics of heterogeneous types of data.Recently,because the differences between application loads increase,even a single level in a multi-level storage needs to handle different types of data.Meanwhile,the complexity of the layered storage system itself exacerbates the difficulty of storage system optimization.This thesis focuses on the optimization of multi-level storage systems for heterogeneous types of data.For the optimization problem of multi-level storage systems under heterogeneous types of data,the key is to clarify the characteristics of different types of data that each individual storage level needs to process,and to optimize the storage system according to the characteristics of hardware and software at each level.This work divided the whole optimization problem into three problems of different storage layers: application layer,external storage layer,and on-chip storage layer.For each layer,we make the following contributions:(1)At the application layer,recent studies classify application data hotness based on the historical access information of application data.However,such methods are not suitable for applications with rapidly changing access patterns.To overcome this problem,this thesis proposes a prediction method of storage access behavior of applications based on a machine learning method.This thesis analyzes the characteristics of typical machine learning methods in detail,compares the accuracy and overhead of each machine learning method in access pattern prediction,and selects a suitable learning model.Finally,by further adjusting the parameters of the selected model,a high prediction accuracy of storage access patterns of applications is achieved.(2)On the external storage layer,recent studies mainly focus on the address management,data consistency,software stack optimization,etc.Those studies do not distinguish different types of data,which will result in a decrease in system access performance.To overcome this problem,we propose an adaptive data placement mechanism for hybrid storage systems based on the predicted results of patterns of application data accesses.The adaptive data placement method utilizes the difference between different storage mediums of hybrid storage systems to suit the changeable data access patterns.The proposed method can improve the data access performance of the hybrid storage system in terms of access latency and energy consumption.(3)On the on-chip storage level,sensitivity analysis methods can locate the vulnerable part of the program data to reduce the overhead of the reliability protection.However,traditional methods cannot suit the cases of multi-bit flip errors,which will decrease the analysis efficiency.To overcome this problem,this thesis proposes a sensitivity analysis framework for multi-bit flip errors.After exploring the sensitivity of program data on the on-chip storage layer under multi-bit flip errors,we use a basic block level sensitivity analysis instead of traditional instruction-level analysis methods.Meanwhile,we also use a machine learning method to classify the sensitivity of some basic blocks directly instead if fault simulation experiments.Using this method,we can achieve significantly higher sensitivity analysis efficiency with a little decrease of analysis accuracy.Finally,based on real application data access traces and execution traces,we carried out experimental verification on the server platform.Experimental results show that the proposed prediction method of storage access patterns can achieve high accuracy.The proposed data placement mechanism for hybrid storage systems can improve data access performance of hybrid storage systems using the performance difference of hybrid storage mediums.The proposed sensitivity analysis framework for on-chip storage can achieve a significant improvement in sensitivity analysis efficiency with accurate sensitivity classification. |
