Exploiting intrinsic flash properties to enhance modern storage systems

Abstract

The longstanding goals of storage system design have been to provide simple abstractions for applications to efficiently access data while ensuring the data durability and security on a hardware device. The traditional storage system, which was designed for slow hard disk drive with little parallelism, does not fit for the new storage technologies such as the faster flash memory with high internal parallelism. The gap between the storage system software and flash device causes both resource inefficiency and sub-optimal performance. This dissertation focuses on the rethinking of the storage system design for flash memory with a holistic approach from the system level to the device level and revisits several critical aspects of the storage system design including the storage performance, performance isolation, energy-efficiency, and data security. The traditional storage system lacks full performance isolation between applications sharing the device because it does not make the software aware of the underlying flash properties and constraints. This dissertation proposes FlashBlox, a storage virtualization system that utilizes flash parallelism to provide hardware isolation between applications by assigning them on dedicated chips. FlashBlox reduces the tail latency of storage operations dramatically compared with the existing software-based isolation techniques while achieving uniform lifetime for the flash device. As the underlying flash device latency is reduced significantly compared to the conventional hard disk drive, the storage software overhead has become the major bottleneck. This dissertation presents FlashMap, a holistic flash-based storage stack that combines memory, storage and device-level indirections into a unified layer. By combining these layers, FlashMap reduces critical-path latency for accessing data in the flash device and improves DRAM caching efficiency significantly for flash management. The traditional storage software incurs energy-intensive storage operations due to the need for maintaining data durability and security for personal data, which has become a significant challenge for resource-constrained devices such as mobiles and wearables. This dissertation proposes WearDrive, a fast and energy-efficient storage system for wearables. WearDrive treats the battery-backed DRAM as non-volatile memory to store personal data and trades the connected phone’s battery for the wearable’s by performing large and energy-intensive tasks on the phone while performing small and energy-efficient tasks locally using battery-backed DRAM. WearDrive improves wearable’s battery life significantly with negligible impact to the phone’s battery life. The storage software which has been developed for decades is still vulnerable to malware attacks. For example, the encryption ransomware which is a malicious software that stealthily encrypts user files and demands a ransom to provide access to these files. Prior solutions such as ransomware detection and data backups have been proposed to defend against encryption ransomware. Unfortunately, by the time the ransomware is detected, some files already undergo encryption and the user is still required to pay a ransom to access those files. Furthermore, ransomware variants can obtain kernel privilege to terminate or destroy these software-based defense systems. This dissertation presents FlashGuard, a ransomware-tolerant SSD which has a firmware-level recovery system that allows effective data recovery from encryption ransomware. FlashGuard leverages the intrinsic flash properties to defend against the encryption ransomware and adds minimal overhead to regular storage operations.