Characterizing and Optimizing the Networking Stack in Databases

Abstract

Databases are latency-critical applications, and client-database communication is a significant contributor to the end-to-end latency. However, the database community has paid little attention to the networking overhead in databases. This thesis focuses on the overhead from the network stack in the server. I characterize the contributions of different components in the database server to the end-to-end latency, focusing on the networking stack. I observe that in transactions involving a single read query, the server network stack accounts for almost 15\% of the total end-to-end latency in VoltDB. Most of this overhead comes from TCP packet processing, interrupt handling, context switches, and I/O multiplexing. Additionally, this work also explores avenues to optimize the networking stack overhead. I find that moving networking to the userspace by bypassing the kernel can significantly reduce the networking stack overhead. This switch in the network stack can help achieve a significant improvement in throughput and lower latency for both the benchmarks used. While the thesis is focused on server networking stack, similar optimization can be applied to client side if necessary hardware (CPU, NIC) is available.