Draconis: Rethinking the Storage Stack

High performance storage systems like file systems, databases and persistent key value stores have very high storage I/O requirements. To maximize the efficiency of disk I/O, these systems usually make certain assumptions about the performance characteristics of the underlying storage medium, and design their disk read/write patterns accordingly. Some of the common assumptions include: sequential accesses are faster than random accesses, minimum read/write size should be a sector (512 bytes), batching writes into a large chunk is more efficient than writing small pieces, etc. Most of these assumptions are based on characteristics of a spinning disk, and therefore usually give applications efficient storage I/O when actually running on spinning disks. However, newer storage technologies like NAND SSDs, phase change memory and memristors all have very different performance characteristics from spinning disks. For example, SSDs do not have moving mechanical parts and therefore do not incur the extra seek penalty when accessing random addresses; the minimum read/write size is an SSD page size, which is typically 4KB; pages need to be erased before writing new content, causing in place updates to be very costly. Not surprisingly, applications written for spinning disks will not have the optimal performance on these newer storage technologies. To make the issue more problematic, different products of the same technology may have very different performance properties. SSDs for instance have different Flash Translation Layers (FTLs), and depends on the translation scheme, random writes can be much slower or equally fast as sequential writes. As a result, applications optimized for one SSD may have much worse performance on another SSD. The problem involves more than just the application layer. Operating systems’ storage I/O subsystem are also designed around a slow spinning disk: the disk interface only has simple sector read/write, I/O scheduling and buffer cache design optimize for sequential accesses and batching, etc. Even when applications make the correct assumptions, the operating system may totally disturb the disk access pattern, leads to non-optimal performance. With applications, operating systems, file systems, drivers, disk controllers and disk hardware all optimizing with only local information, it becomes extremely difficult for the application to have any end to end storage performance guarantees. This project thus aims to revisit storage system design across the stack. The goal of the project is twofold. Firstly, understand the inefficiencies caused by application’s wrong assumptions about the storage hardware and by the interference from multiple storage layers. This will involve measurements of several storage applications on different storage technologies (spinning disks and SSDs) and performance analysis against raw hardware capabilities. Secondly, base on the observations from the measurement study, propose a new storage stack design that optimizes applications’ I/O performance across different storage hardware. This may involve redesign of the OS I/O subsystem, storage API, disk driver/controller or the application.