On Augmenting TCP/IP Stack via eBPF

Abstract

As the data center networks' bandwidth-delay product is increasing and the applications are moving to nano services (with many small flows), managing flows in the network is becoming more challenging. Current TCP/IP stack faces fundamental limitations to meet these challenges. First, it lacks the ability to accurately estimate the network state under dynamic network settings. Second, the current stack is not flexible enough to be extended easily. In this work, we propose a framework, Augmenter, that augments (i.e., increases the network visibility of) the TCP/IP stack to address these challenges. Leveraging eBPF, Augmenter gathers the state of ongoing flows and uses this information to manage other flows that are currently active or arriving in the future. We present one specific use case of setting the initial congestion window of flows dynamically based on network conditions. Our initial tests, show that Augmenter can improve the application performance by up to 1.4x compared to the fixed initial window-based solutions. Implementing Augmenter in the TCP/IP stack itself is not trivial. Augmenter employs eBPF to implement its desired functionality as it enables introducing such changes relatively easy. We discuss potential challenges and solutions in designing and implementing Augmenter applications.