Host-guided data placement: whose job is it anyway?

Abstract: The increasing demand for SSDs coupled with scaling difficulties have left manufacturers scrambling for newer SSD interfaces which promise better performance and durability. While these interfaces reduce the rigidity of traditional abstractions, they require application or system-level changes that can impact the stability, security, and portability of systems. To make matters worse, such changes are rendered futile with introduction of next-generation interfaces. Further, there is little guidance on data placement and hardware specifics are often abstracted from the application layer. It is no surprise therefore that such interfaces have seen limited adoption, leaving behind a graveyard of experimental interfaces ranging from open-channel SSDs to zoned namespaces. In this paper, we show how shim layers can to shield systems from changing hardware interfaces while benefiting from them. We present Reshim, an all-userspace shim layer that performs affinity and lifetime based data placement with no change to the operating system or the application. We demonstrate Reshim's ease of adoption with host-device coordination for three widely-used data-intensive systems: RocksDB, MongoDB, and CacheLib. With Reshim, these systems see 2-6 times highe write throughput, up to 6 times lower latency, and reduced write amplification compared to filesystems like F2FS. Reshim performs on par with application-specific backends like ZenFS while offering more generality, lower latency, and richer data placement. With Reshim we demonstrate the value of isolating the complexity of the placement logic, allowing easy deployment of dynamic placement rules across several applications and storage interfaces.