Towards Improving Endurance and Performance in Flash Storage Clusters

Abstract

NAND flash-based Solid State Devices (SSDs) provide high performance and energy efficiency and at the same time their capacity continues to grow at an unprecedented rate. As a result, SSDs are increasingly being used in high end computing systems such as supercomputing clusters. However, one of the biggest impediments to large scale deployments is the limited erase cycles in flash devices. The natural skewness in I/O workloads can results in Wear imbalance which has a significant impact on the reliability, performance as well as lifetime of the cluster. Current load balancers for storage systems are designed with a critical goal to optimize performance. Data migration techniques are used to handle wear balancing but they suffer from a huge metadata overhead and extra erasures. To overcome these problems, we propose an endurance-aware write off-loading technique (EWO) for balancing the wear across different flash-based servers with minimal extra cost. Extant wear leveling algorithms are designed for a single flash device. With the use of flash devices in enterprise server storage, the wear leveling algorithms need to take into account the variance of the wear at the cluster level. EWO exploits the out-of-place update feature of flash memory by off- loading the writes across flash servers instead of moving data across flash servers to mitigate extra-wear cost. To evenly distribute erasures to flash servers, EWO off-loads writes from the flash servers with high erase cycles to the ones with low erase cycles by first quantitatively calculating the amount of writes based on the frequency of garbage collection. To reduce metadata overhead caused by write off-loading, EWO employs a hot-slice off-loading policy to explore the trade-offs between extra-wear cost and metadata overhead. Evaluation on a 50 to 200 node SSD cluster shows that EWO outperforms data migration based wear balancing techniques, reducing up to 70% aggregate extra erase cycles while improving the write performance by up to 20% compared to data migration.