Implementing an Append-Only Interface for Semiconductor Storage

Solid-state disks are currently based on NAND flash and expose a standard disk interface. To accommodate limitations of the medium, solid-state disk implementations avoid rewriting data in place, instead exposing a logical remapping of the physical storage. We present an alternative way to use flash storage, where an append interface is exposed directly to software. We motivate how an append interface could be used by a database system. We describe bit rot and space reclamation in direct-attached logstructured storage and give details of how to implement the interface in a custom controller. We then show how to make append operations idempotent, including how to create a fuzzy pointer to a page that has not yet been appended (and therefore whose address is not yet known), how to detect holes in the append sequence, and how to scale out read operations. 1 The Case for a Log-Structured Database Most database systems store data persistently in two representations, a database and a log. These representations are based in large part on the behavior of hard disk drives, namely, that they are much faster at reading and writing pages sequentially than reading and writing them randomly. The database stores related pages contiguously so they can be read quickly to process a query. Writes are written sequentially to the log in the order they execute to enable high transaction rates. Inexpensive, high-performance, non-volatile semiconductor storage, notably NAND flash, has very different behavior than hard disks. This makes it worthwhile to consider other storage interfaces than that of hard disks and other data representations than the combination of a database and log. Unlike hard disks, semiconductor storage is not much faster at sequential operations than at random operations, so there is little performance benefit from laying data out on contiguous pages. However, semiconductor storage offers enormously more random read and write operations per second per gigabyte (GB) than hard drives. For example, a single 4GB flash chip can perform on the order of 10,000 random 4KB reads per second or 5,000 random writes per second. So, in a shared storage server cluster with adequate I/O bandwidth, 1TB of flash can support several million random reads per second and could process those requests in parallel across physical chips. This compares to about 200 random reads for a terabyte of hard drive capacity. However, per gigabyte, the raw flash chips are more than ten times the price of hard drives. Copyright 2010 IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works must be obtained from the IEEE. Bulletin of the IEEE Computer Society Technical Committee on Data Engineering