Load-aware Optimization On Cache Management For SSD

In recent years,with the rapid development of semiconductor technology,flash-based solid-state drives have received extensive attention and research in industry and academia.Compared with traditional mechanical hard disks,SSDs have advantages of fast access speed,low energy consumption,low noise,and strong shock resistance,and they have been widely used in military,medical,and electronic products.Due to the physical characteristics of flash memory,SSDs have many constraints such as write-before-erase,asymmetric read and write latency,and limited erase count.In order to reduce the impact of these constraints on the performance of SSDs,caching technol-ogy is introduced in SSDs,which enabled some user I/O requests to be completed in the cache,reduced read and write operations on SSDs and improved I/O performance of SSDs.At present,the cache management algorithm has become a hot issue in the research of solid state drives.The traditional cache management algorithms of SSDs are designed to improve the cache hit rate or reduce the cache replacement cost,and often ignore the impact of the state of the underlying flash chips on the performance of the cache algorithm.The optimization of replacement strategy in the cache management algorithm is the main topic of this paper.Considering the load characteristics of the underling flash chips,we proposed a load-aware cache replacement algorithm.The main research contents are as follows:(1)Load-aware Optimization on Cache Strategy for SSDIn order to reduce the impact of I/O access latency on the unbalanced workloads of flash chips,we proposed a load-aware Cache Replacement(LCR)algorithm based on cache replacement cost,the main constraint is the load condition of fl ash chips.The basic idea is to give a higher priority to cache the blocks on overloaded flash chips and replace the data blocks on the chip with light load.The LCR algorithm divides the cache into a Working Region and a Destaging Region,the Working Region caches the data blocks with high frequency,and the Destaging Region keeps the data blocks which are candidates for eviction.Finally,we evaluate the performance of LCR scheme by using a trace-driven simulator with multiple real-world workloads.And results show that compared with the classic cache replacement algorithms LRU and GCaR-CFLRU,the LCR algorithm can reduce the average system response time by as much as 39.2%and 12.3%,respectively.(2)Load-aware Optimization on Cache Strategy for SSD RAIDRAID5 array system based on SSDs face the problem that small write requests are frequently updated,which cause parity updating frequently,inducing additional read and write I/O operations,which not only increase system access latency,but also shorten the lifetime of RAID system.To solve this problem,we proposed a Stripe-level Cache Management Algorithm based on Load Awareness(SLCache),the cache granularity is in stripes,the data belonging to the same stripe is stored in the same stripe node,and user I/O read and write requests are also cached.For the design of the cache replace-ment strategy,two main factors are concerned.First,the number of dirty data pages included in the stripe;Second,the load of the SSDs for dirty data blocks and parity blocks.Replacement objects is selected based on the principle that the replacement cost is minimal.At the same time,in the data write-back operation,different parity up-dating strategies are selected according to the number of dirty data pages.Finally,we implemented the SLCache algorithm on the trace-driven SSD simulator Disksim+SSD Extension,and compared with LRU-RAID5,our approach can reduce the number of erasures by 15.9%,while reduce the average system response time by 20.3%and re-duce the read overhead of the parity update 16.1%.