PocketData: Benchmarking and Redesigning Mobile Data Storage Systems
Smartphone apps and platforms rely on access to structured and unstructured data. Today, many apps build persistent storage layers on top of embedded SQL databases such as SQLite. However, SQL is not necessarily an ideal choice for persisting many types of data structures. Embedded database engines lack benchmarks that reflect the differences between interactive workloads and the traditional throughput-driven workloads that are used to evaluate database servers supporting big data and web applications. The PocketData project is developing new benchmarks based on smartphone embedded database workloads that can drive innovative next-generation approaches to mobile structured data storage and access.
The usage of embedded databases by smartphone apps differs from traditional database access patterns in two important ways. First, mobile app embedded databases support interactive workloads where queries arrive in bursts separately by long periods of idleness. Thus, raw throughput is a poor way to evaluate their performance, and opportunities exist to reorganize data after interactive bursts to prepare for future requests. We plan to leverage the ongoing work by the ODIn Lab on just-in-time data structures as a way to use idleness and previous queries to iterative improve performance.
In all of these cases, the freedom of the embedded database to couple itself tightly to a specific app may open up new optimization opportunities.Second, mobile app embedded databases support single apps that may issue queries with very specific properties. For example, our initial analysis has demonstrated that many Android apps use SQLite purely as a key-value store. In other cases, query patterns indicate that apps are using existing object-relational mapping libraries ( ORMs) to persist objects, which produce their own distinctive access patterns. We have also seen cases where SQLite database were never updated by apps, indicating that the entire database may serve as a cache of structured app data that is only updated during app upgrades. In all of these cases, the freedom of the embedded database to couple itself tightly to a specific app may open up new optimization opportunities.
To explore these ideas we are beginning by building a smartphone embedded app database benchmark—actually a benchmark generator, which will be able to use traces of database activity to synthesize a benchmark for any app. This feature is critical to ensure that we can support the variety of app database usage patterns and update the benchmark suite easily as they change. We are also evaluating the performance and energy consumption of SQLite, the embedded database provided by default to Android apps, against alternatives such as BerkeleyDB and H-Store.