Lost in translation : giving SMR drives a second chance

Shingled Magnetic Recording (SMR) drives offer higher data density and lower cost than Conventional Magnetic Recording (CMR) drives, but with a limitation: they only accept sequential writes. To handle random writes, SMR drives rely on a Shingled Translation Layer (STL) that converts random writes into sequential ones. In Device-Managed SMR drives, the STL resides in the drive's firmware, whereas in Host-Managed SMR drives, the STL resides in the host operating system. Anecdotal evidence suggests that replacing CMR drives with Device-Managed SMR drives leads to unacceptable I/O performance. However, there are no detailed root cause analyses of this degradation. This shallow evidence has led to the misconception that SMR drives are inherently inferior to their CMR counterparts and must only be used with applications that write sequentially. This dissertation demonstrates that the real issue is not the shingled medium itself, but rather the design of the STL. This design is influenced by the limited resources available on the device to keep the drive affordable. We show that moving the STL to the host OS can make SMR drives competitive with CMR drives while requiring no application or filesystem changes. Our host-based STL improves 99ᵗʰ⁺ percentile tail latency by ~40× compared to on-device implementations. In RAID reconstruction workloads previously cited as evidence of SMR failure, our solution completes reconstruction ~1.4× faster than CMR drives and ~11× faster than Device-Managed SMR.