Research On I/O Performance Jitter Optimization Algorithm For Solid State Disk

With the increasing application scale of big data,the processing requirements of massive data put forward higher requirements for the storage performance of computer systems.Traditional HDDs based storage systems have low performance and can not meet the demand of large data applications with low response time because of their mechanical properties.At the same time,NAND flash-based SSDs have been widely used for their excellent read-write performance.Compared with HDDs,NAND flash-based SSDs have the advantages of high read-write performance,low power consumption,and low noise,so they are gradually being widely used in large-scale storage systems.However,due to the physical characteristics of the NAND flash media,such as the inconsistent granularity of read,write and erase operations,high latency of erase operations,and limited number of P/E operations,there are many new problems in the practical application of SSDs.One of them is that the internal garbage collection operation will cause performance jitter in the SSDs.Usually this problem will not be noticed by researchers,but it will increase the user I/O request response delay,and affecting the user experience.Large performance jitter will affect the user request completion time,which can cause transaction data errors and even loss of critical data,especially in real-time transactional applications.In order to solve the performance jitter problem caused by garbage collection,we propose a buffer management algorithm CalmWPC for write performance jitter optimization in this paper.The algorithm mainly includes a cluster-based buffer data organization structure,a cluster activity prediction module based on historical data,and a fingerprint library module based on updating the data page between the DRAM space and the flash storage unit in the SSD.In the CalmWPC algorithm,the data organization of the cluster is the basis of the prediction and fingerprint library module.By clustering,the update operation can be concentrated into a certain address range as much as possible,and finally more invalid data pages are generated in the same flash data block.The prediction module calculates the activity of the cluster in the future based on the history of the cluster operation,and uses the specific prediction algorithm to calculate the activity of the cluster in the future.The fingerprint library module acts as a sensing module for updating the data page between the DRAM space and the flash storage unit in the SSD,so that when the request data is written into the DRAM space,the data can be perceived in real time as whether the data is updated data relative to the flash memory.The main idea of the CalmWPC algorithm is to replace the cluster with less activity and more updated data pages in the DRAM space.On the one hand,it can reduce the active data being frequently written back to the flash unit,improve the space recovery efficiency of garbage collection after each cached data write back,and avoid the impact of frequent garbage collection on external request response as much as possible.On the other hand,the garbage collection operation can be performed in a decentralized manner,eliminating many disadvantages of the traditional hard-threshold-based centralized garbage collection operation,and further avoiding the occurrence of write performance jitter caused by the garbage collection operation.The experimental result shows that the algorithm of CalmWPC which proposed in this paper can effectively reduce the write performance jitter,improve the average write access latency,optimize the write amplification,and thus improve the performance of the solid state disk and extend the life of the solid state disk.Take the trace of Financiall as an example,CalmWPC can reduces the write performance cliff by averages of 60.9%and 60.0%compared with LRU and CFLRU.CalmWPC also shortens the response time of LRU and CFLRU by averages of 69.4%and 70.1%,respectively.In addition,the write amplification of the CalmWPC algorithm is on average 16.3%and 15.8%lower than LRU and CFLRU,respectively.