Ranking and new database architectures

Database platform support for efficient ranking can have positive performance implications for a number of rank-aware applications, including data exploration, social network analysis, and keyword search. This talk highlights two pieces of work where ranking appears in the research area of database architectures for new hardware. First, we highlight the Bw-tree, a new high-performance B+-tree supporting sorted key-sequential access. The Bw-tree is re-architected to run efficiently on new hardware: its in-memory operations are completely latch-free, removing blocking behavior while also improving multi-core cache behavior, while its storage layer implements a novel log-structured flash storage layer for that exploits fast sequential writes and mitigates adverse performance impact of random writes. Second, we highlight a new classification technique for identifying âĂIJcoldâĂİ (infrequently accessed) data in main-memory database systems. Using a log of sampled record accesses, our technique estimates record access frequencies using exponential smoothing. This classification approach is very efficient: it is able to accurately identify hot and cold records among 1M records in sub-second time from a log of 1B record accesses on a workstation class machine.