We built SplinterDB to address two trends. The first is the modern storage hardware offers significantly higher bandwidth and lower latency. The second is that modern applications at VMware and elsewhere store fine-grain data, such as metadata. Current state-of-the-art key-value stores such as RocksDB fail to fully exploit the capabilities of these devices on these types of workloads.

SplinterDB is designed around a novel data structure, the mapped Bε-tree, which combines ideas from the theory of external memory hash tables with those from storage systems. Mapped Bε-trees expose I/O and CPU concurrency and reduce write amplification without sacrificing query performance.

SplinterDB outperforms RocksDB by up to a factor of 9x on insertion workloads and up to 80% on query workloads.

This talk is part of the ¡Databases! – A Database Seminar Series.

Bio:
Alex is a researcher at VMware. His work bridges storage systems and external memory data structures. He is a co-creator and research lead of SplinterDB.

More Info: https://db.cs.cmu.edu/seminar2022#db11