Learnings from an Under the Hood Analysis of an Object Storage Node IO Stack

Conventional object-stores are built on top of traditional OS storage stack, where I/O requests typically transfers through multiple hefty and redundant layers. The complexity of object management has grown dramatically with the ever increasing requirements of performance, consistency and fault-tolerance from storage subsystems. Simply stated, more number of intermediate layers are encountered in the I/O data path, with each passing layer adding its own syntax and semantics. Thereby increasing the overheads of request processing. In this paper, through comprehensive under-the-hood analysis of an object-storage node, we characterize the impact of object-store (and user-application) workloads on the OS I/O stack and its subsequent rippling effect on the underlying object-storage devices (OSD). We observe that the legacy architecture of the OS based I/O storage stack coupled with complex data management policies leads to a performance mismatch between what an end-storage device is capable of delivering and what it actually delivers in a production environment. Therefore, the gains derived from developing faster storage devices is often nullified. These issues get more pronounced in highly concurrent and multiplexed cloud environments. Owing to the associated issues of object-management and the vulnerabilities of the OS I/O software stacks, we discuss the potential of a new class of storage devices, known as Object-Drives. Samsung Key-Value SSD (KV-SSD) [1] and Seagate Kinetic Drive [2] are classic industrial implementations of object-drives, where host data management functionalities can be offloaded to the storage device. This leads towards the simplification of the over-all storage stack. Based on our analysis, we believe object-drives can alleviate object-stores from highly taxing overheads of data management with 20-38% time-savings over traditional Operating Systems (OS) stack.