Transaction Processing at Scale
Online transaction processing (OLTP) is critical for applications including finance, e-commerce, social networks, and healthcare. The increasing performance demands of these applications require OLTP to scale massively. Concurrency control is a major scalability bottleneck in such systems.
This talk presents three projects that identify and help resolve scalability challenges. First, I present a scalability study of concurrency control on a simulated 1000-core processor and show the bottlenecks that constrain the scaling of classic algorithms. Then, I present a new protocol called TicToc that removes the bottleneck of central timestamp allocation on multicore processors. The key technique is data-driven timestamp management that dynamically calculates each transaction's timestamp based on its data access pattern. Finally, I present Sundial, a distributed concurrency control scheme that mitigates the bottleneck of long network latency through a lightweight caching protocol. The talk ends with a vision of transaction processing in the era of cloud computing and internet of things.
Xiangyao Yu is a postdoctoral associate at Computer Science and Artificial Intelligence Lab (CSAIL) at Massachusetts Institute of Technology (MIT) under the supervision of Prof. Michael Stonebraker. He acquired his Ph.D. at MIT in 2017 and B.S. at Tsinghua University in 2012. His research interest centers on databases with additional expertise in computer architecture. He won the best Ph.D. thesis award in EECS at MIT and has three best paper awards or nominations.