Research On Stream Management Of NVMe Multi-Stream SSDs

The non-volatile memory host controller interface specification NVMe(Non-Volatile Memory express)improves the performance of solid state drives(SSDs).The host passes the semantic information such as the expected lifetime of the data which is generated by stream management strategies to a SSD through the newly introduced multi-stream interface of NVMe,which can guide the SSD to organize its internal data layout,so that the number of migrated valid pages during garbage collection(GC)can be reduced,therefore,the write amplification of a SSD can be reduced.The existing manual stream management strategies require experienced developers who are familiar with the application to customize according to its internal working mechanism,and they depend on a specific application.The automatic stream management strategies predict the lifetimes of data according to the characteristics of the requests(e.g.,the sequentiality,frequency,and recency of the requests' logical address characteristics),they don't depend on a specific application and have good scalability.However,the existing automatic stream management strategies don't consider the difference in lifetimes between metadata and data,among different types of metadata,and among different data,and they don't consider the difference in write amplification between sequential and random applications,and they don't consider the problem that the number of streams supported by a SSD is limited.In this context,we propose a mixed automatic stream management strategy MixedStream based on the kernel space I/O stacks.MixedStream combines the types of metadata with the logical address characteristics of the requests and can distinguish the difference in lifetimes between metadata and data,among different types of metadata,and among different data.We propose an automatic stream management strategy UStream based on the user space I/O stacks.UStream can dynamically adapt to different applications and distinguish the difference in write amplification between sequential and random applications.The concept of virtual streams is introduced and the method of splitting into two steps is used to solve the problem that the number of streams supported by a SSD is limited.During the first split,MixedStream completes the mapping of metadata and data to virtual streams,and UStream completes the mapping of sequential and random applications' data to virtual streams.During the second split,both stream management strategies dynamically adjust the mapping relationship between the virtual streams and the streams supported by a SSD according to the average lifetime of data/metadata in each virtual stream.Experiment verification was completed.Test results on a self-developed multi-stream SSD show that compared with two stream management strategies: FStream which is developed only according to the metadata types and AutoStream which is developed only according to the logical address characteristics of the requests,MixedStream reduces the write amplification factor(WAF)by 2.11%-6.64% and 1.96%-7.56%,respectively.MixedStream improves the performance by 1.49%-4.61% and 0.79%-5.11%,respectively.Compared with AutoStream,UStream reduces the WAF by 8.03%-21.57% and improves the performance by 5.88%-16.43%.