GC-RAIS:一种基于垃圾回收感知的固态盘阵列

GC-RAIS: Garbage Collection Aware and Redundant Array of Independent SSDs

垃圾回收操作会显著影响固态盘的性能,进而导致固态盘阵列的性能波动.为此,提出一种基于垃圾回收感知的磁盘阵列(GC-RAIS),充分利用固态盘的高随机读特性和固态盘阵列中的热备份盘,以减轻垃圾回收操作对固态盘阵列性能波动的负面影响.当固态盘阵列中某个固态盘正在处理垃圾回收操作时,对于到达该固态盘的读请求采用重构方式处理,即读取同一条带上其他固态盘上的数据重构得到,而对于到达该固态盘的写请求则将写数据临时存放在热备盘中,并更新相应的校验信息.当垃圾回收过程结束后,将被重定向的写数据写回到正确的固态盘中.仿真实验结果表明相对局部垃圾回收LGC策略和全局垃圾回收GGC策略,GC-RAIS分别减少用户I/O请求的平均响应时间达55%和25%.

SSDs are popular in large-scale storage systems to accelerate the system performance because a single SSD cannot satisfy the performance, capacity and reliability requirements of data- intensive computing applications. Thus applying RAID algorithms to SSDs is necessary and promising to build high performance, high capacity and highly reliable SSD^based storage systems. However, garbage collection operations in SSDs have a significant impact on the SSD performance, thus leading to the performance variability in redundant array of independent SSDs （RAIS）. To address this problem, GC-RAIS exploits the high random-read performance characteristics of SSDs and the hot- spare SSD in RAIS to alleviate the negative impact of GC operations on the RAIS performance. When an SSD is in the GC state, the incoming read requests to this SSD are serviced by reconstructing the read data from the other SSDs in the same stripe （read reconstruction）, while the incoming write data is temporally stored on the hot-spare SSD and the corresponding parity is concurrently updated （write redirection）. After the GC process completes, the redirected write data is reclaimed to its correct SSD. The original DiskSim and the MSR SSD simulator are extended to implement the proposed GC- RAIS and the GC-RAIS performance is evaluated with the HPC-like and enterprise realistic workloads. The simulation results show that GC-RAIS significantly outperforms the local garbage collection （LGC） and the global garbage collection （GGC） by 55~ and 25~ on average, respectively. Moreover, GC-RAIS reduces the performance variability for a wide variety of HPC-like and enterprise realistic workloads.