Optimization Of Prefetching Data In Solid-state Drive Based On Deep Learning Algorithm

Solid-State Drives(SSDs)have become the storage device for many personal computers and enterprise servers.However,the access speed of SSDs still needs to improve for computer systems.With prefetching,SSDs are able to predict future requests before moving data to a faster storage unit.Nevertheless,the large hard disk address space and the interleaved access flow of applications pose challenges for existing prefetchers.The existing prefetchers cannot accurately predict the subsequent request’s logical block address(LBA)and I/O size.In order to learn the internal relationship of the logical block address sequence,this thesis differs the logical block address sequence to the first order to obtain the differential sequence,also called the Delta sequence.As a solution to the problem of the large input space,this thesis uses the top-1000 frequently occurred Deltas as the input to the model.This thesis proposes a model with two components: a self-attention mechanismbased prefetcher and an LBA-I/O size mapping table.The self-attention mechanism-based prefetcher applies a decoder-only framework with the self-attention mechanism to forecast the next LBA.The LBA-I/O size mapping table maps the logical block address to the most frequently appeared I/O size at that logical block address for prediction.Since none of the previous works considered the dynamic changing I/O size during testing,to evaluate the overall performance of the whole model,this work presents an algorithm to simulate the buffer area as I/O size changes dynamically.Experiments show that the accuracy,coverage,and f1 score of the self-attention mechanism-based prefetcher surpass the baselines in most cases.The prediction accuracy of the LBA-I/O size mapping table exceeds the baseline on all datasets.The prefetching accuracy,coverage,and f1 score of the model also outperform the baseline considering variable I/O size.In addition,since using the top-1000 frequently occurred Deltas as the input to the prefetcher is not adaptable to the requirements of different SSDs,we explore the possibility of employing recurrent Deltas as the input to the self-attention mechanism-based prefetcher.Experiments demonstrate that it helps to increase the prefetching accuracy,coverage,and f1 score of the prefetcher using recurrent Deltas as prefetcher inputs.Finally,as it is not realistic to train an individual model for prediction for each disk,this paper explores the possibility of transferring the same model to different disks.Experiments illustrate whether a model trained on one trace file can be migrated to another trace file,depending on the regularity of the access pattern of the latter.