SABER: Hybrid Data Processing with Heterogeneous Servers

Modern servers in data centres have become increasingly heterogeneous, e.g. combining multi-core CPUs with many-core GPUs. This has implications on the design of future data-intensive systems for stream processing or machine learning: first, systems must exploit all the available parallelism of the hardware, independently of the processing semantics; and, second, instead of offloading computation entirely to an accelerator, systems must fully utilise all heterogeneous processors in a server, thus making accelerators first-class compute elements.

In this talk, I will describe SABER, a new hybrid stream processing engine for CPUs and GPUs. Under a hybrid execution model, SABER executes streaming SQL queries in a data-parallel fashion on all available CPUs and GPUs simultaneously. Instead of statically assigning query tasks to heterogeneous processors, SABER adaptively schedules computation on the best available processor. It parallelises stream queries in a way that suits the properties of the hardware, independently of the window-based query semantics. Our experiments show how SABER's hybrid execution model can aggregate the performance of multiple heterogeneous processors in a server.

This talk is based on a SIGMOD'16 paper.