Modernizing Storage Systems for Big Data Applications

Abstract

Existing high-speed non-volatile storage systems leverage entrenched system stack developed for magnetic hard disk drive, leading to suboptimal performance and under-utilized system resources. As data set sizes of applications keep increasing, using conventional system stack for modern storage devices becomes a new performance bottleneck. For example, a database system can spend 80% of time in just fetching data from the storage system. ????? In this talk, Hung-Wei will demonstrate how to achieve significant performance gain through modernizing the hardware/software interface for storage systems. Hung-Wei will introduce HippogriffDB, a GPU-based database system that balances the huge gap between the throughputs of the GPU and the SSD. With HippogriffDB using a GPU friendly query processing model and streamlining the data exchange mechanisms between an NVMe SSD and the GPU, HippogriffDB achieves 250x speedup for a state-of-the-art GPU-based database system. ????? Hung-Wei will also talk about his recent work, KAML that modernizes the conventional block-based I/O with a keyvalue-like interface. With KAML’s interface, we can eliminate the software indirections in translating block addresses and allow the SSD to provide native support of transactions. As a result, KAML can achieve up to 3x speedup for key-value store applications.