Fast Transactions And Recovery With Commodity Storage in Replicated Multicore Databases
Databases traditionally use slow persistent storage like hard disks but some recent databases primarily use main memory instead. They achieve higher throughput and lower latency but are more expensive and limited in terms of dataset size. One example is Silo: it uses data structures optimized for concurrent in-memory access on modern multicore machines resulting in extremely high throughput. This project will extend Silo to support modern persistent storage such as flash and non-volatile memory. The goal is to preserve Silo’s speed for working sets that fit in memory while still allowing good performance on larger datasets. This project is closely related to another project to improve Silo’s recovery speed through replication and integrating the two will be a major component.

Last year I did a UROP with Professor Liskov on a project intended to make writing performant concurrent programs easier with transactional data structures. Most recently I worked at SpaceX improving the scalability of their in-house distributed GPU combustion simulator for rocket engines. I’m interested in high performance and distributed systems and I hope to deepen my understanding about them through this SuperUROP.