Research On The Construction And Performance Optimization Of Software-defined Flash Storage System

With the rapid development of Internet technology,the amount of data that the storage system needs to handle is getting larger,and users demand higher performance of the storage system.Emerging flash memory technology has brought breakthrough to the storage system and greatly improved the performance of the external storage system.However,compared with the user's performance requirements of the storage system and the computing ability of the computer system,the storage performance is still far from being desired.External storage devices are still the bottleneck of the entire computer system.In addition,due to the flaws of the external storage devices,flash-based solid-state storage system faces serious challenges in terms of reliability and performance improvement.On the SSD device side,this thesis proposes a software-defined phase-change memory and flash multi-media fusion SSD construction method to improve the lower-case performance and enhance reliability of SSD.The advent of the emerging non-volatile memories,especially phase-change memory,provides a new research area for improving the performance and reliability of SSDs.Through aggregating the fast speed,flexible access mode and non-volatile properties of phase-change memory and software-defined flash memory,we can flexibly use the PCM as cache or the same level storage devices according to the requirements of different application load,thus efficiently alleviating the poor performance of small writes,reducing the small writes update overhead of the flash memory,improving the performance of the SSD device,and enhancing the reliability of the SSD.On the basis of theoretical research,the software-defined multi-media fusion solid-state disk hardware prototype development platform DSAL V1.0 is realized.Through experiments on the prototype system,it is found that when the small writes are dominate,exploiting the PCM as the same level storage devices is beneficial to improving the performance and reliability of multi-media fusion SSD;When the dominate operations in the application are read access,configuring the PCM as the cache of flash is more beneficial to improving the average response time of the requests of the system.For the complex and varied application environment,this thesis proposes an applicationaware programmable solid-state disk(A-SSD),which enables SSDs to adopt different FTL strategies for different applications and and can be dynamically adjusted according to the changes of applications.Due to the flaws of flash memory storage characteristics,the same SSD configuration may have great performance differences under different workloads.And also because of the architecture of traditional Black-box SSD,response spikes(data access response time exceeds a certain threshold,it is considered to be a response spike)are always exist.The application-aware programmable solid-state disk A-SSD is implemented by moving the flash translation layer FTL to the host side,and reserves an interface for the application layer,receiving application layer semantic information,and providing a programmable interface for the application layer,thereby making SSDs can use different FTL policies for different applications and can be dynamically adjusted based on application changes.Using the relevant semantic information obtained from the upper layer,according to the data access attribute,the aggregation programming and data layout optimization methods are adopted,which lays a foundation for improving garbage collection efficiency and achieving wear leveling.According to the perception of the application load characteristics,choose the appropriate timing to perform a global garbage collection and wear leveling strategy for the corresponding SSD.Taking Tencent as an example,a complex and changeable application environment with a large number of users and mixed multi-applications,it is found that the solid-state storage system constructed by A-SSD can always show better performance than traditional SSD and it effectively controls the average response time of the solid-state disk device within 10 ms and eliminate the response spike.For large-scale solid-state storage systems,a software-defined solid-state disk full flash array(S-RAIS)is proposed to improve the performance of the solid-state disk array.In order to improve the performance and reliability of the storage system in the enterprise applications,SSD RAID are always performed,but at the same time the problem of frequent updates of parities and the sharp decline of the service performance of the SSD array caused by multiple mapping is often faced when the SSD array is constructed.Software-defined solid-state disks(SSDs)are used to construct full flash array S-RAIS,and the methods to improve the service performance of SSDs are studied from the aspects of striping mapping,data layout optimization,smooth device updating and data recovery.By merging the strip mapping table of the solid-state disk array with the flash memory transfer layer FTL of the solid-state disk,the strip is constructed directly with the flash memory physical block,which effectively eliminates the multiple mapping problem of the solid-state disk array and enables the array controller to directly perceive the flash memory physical block in the solid-state disk.When the data needs to be updated,we use the out-of-place update feature of the flash memory and the method of appending write to keep the stripe relationship between the old data and the new data,thereby effectively reducing the update overhead of the frequent parity block caused by the updates of the small writes,improving real-time response performance when the data updates are written.Experimental results show that under different load test conditions,s-rais can effectively reduce the erasure times of flash memory by up to 50%,and improve the average write response time of the system by 5%?20%.