Zoned Hybrid Storage:Architecture And Optimizations

In the era of big data analytics,the huge amount of constantly generated data raises an ever-increasing demand for high-performance and large-capacity storage solutions.Deploying solid state drives(SSDs)as a cache layer or a fast storage tier together with hard disk drives(HDDs)has been prevalent in cost-effective storage systems.The major objective of this hybrid approach is to simultaneously achieve high performance and reduce the total cost of ownership(TCO)for storage systems.In recent years,shingled magnetic recording(SMR)has been proposed to achieve higher areal density to respond to the demand of exponentially increased storage capacity.However,SMR drives endure severe performance jitter under non-sequential writes(NSWs)due to its zone-based internal structure and the costly media cache cleaning.Meanwhile,SSD has been facing rapid development in the underlying flash memory.This incurs SSD reliability reduction and internal resource utilization stagnation.The new emerging SMR and flash technologies have opportunities to cooperatively satisfy the demand of high-performance and large-capacity storage systems,but also pose challenges for existing hybrid storage designs.In this paper,we present a complete scenario for building cost-effective hybrid storage systems with SSDs and SMR drives.Our main contributions can be summarized as follows.We propose ZoneTier,a zone-based storage tiering and caching co-design,to effectively control all the NSWs by leveraging the host-aware property of HA-SMR drives.ZoneTier exploits real-time data layout of SMR zones to optimize zone placement,reshapes NSWs generated from zone demotions to SMR preferred sequential writes,and transforms the inevitable NSWs to cleaning-friendly write traffics.Our experiments show that ZoneTier can utilize SSD efficiently,minimize zone relocation overhead,and shorten performance recovery time from media cache cleaning.The techniques used in ZoneTier can be easily extended to other deployment scenarios of SMR drives.In ZoneTier,proactively caching NSWs into SSD incurs severe SSD endurance which further violates the system reliability of hybrid storage.We propose SMRC,an endurable SMR-oriented SSD Caching framework.SMRC leverages the intrinsic host-aware property of HA-SMR drives to filter both sequential writes and innocuous NSWs out of SSD during cache admission.Our experiments show that SMRC can effectively reduce SSD write traffics without exacerbating the write amplification in HA-SMR drives,finally provide higher reliability for both SSD and ZoneTier.In ZoneTier,redirecting intensively accessed data to SSD gives rise to the aggregation of flash requests in the SSD device-level queue.We propose PIOS,a Parallelizable I/O Scheduler,to improve internal resource utilization for SSDs from the perspective of I/O scheduling.PIOS pinpoints the access conflicts in the device-level request queue and schedules the I/O requests in a fine-grained way by cooperating with flash translation layer(FTL).Our experiments show that PIOS can significantly improve SSD performance,which is of great benefit to the performance of both SSD and ZoneTier.In ZoneTier,redirecting intensively accessed data to SSD also results in that static wear-leveling cannot get enough idle time to complete at runtime.We propose Roble,a strategy of Regional over-wear for bettering leveling,to minimize the data migration overhead of static wear-leveling.Roble rotationally centralizes hot data placement into a single region to increase cold data aggregation and leverages the internal parallelism to improve data migration efficiency in flash chips.Our experiments indicate that Roble can significantly reduce data migration overhead while achieving SSD wear-leveling,which is of great benefit to the reliability of both SSD and ZoneTier.