eGossip: Optimizing Resource Utilization in Gossip-Based Clusters through eBPF
Kuo-Guang Tsai, Jerry Chou · 2024
The Extended Berkeley Packet Filter (eBPF) presents a transformative approach to dynamic micro-program loading and execution directly within the Linux kernel, circumventing the need for kernel recompilation. This paper explores the innovative application of eBPF in optimizing resource usage within gossip-based cluster environments. Our primary focus is the deployment of TC eBPF hooks for in-kernel packet cloning, aiming to enhance the efficiency of user-space broadcasts. This approach is particularly pivotal in consensus algorithms like Gossip, which heavily rely on user-space broadcast mechanisms. We explore the implementation of this method, highlighting its ability to boost broadcast speeds significantly. Our findings reveal a marked improvement in CPU utilization, which decreased by 29%, and in Transmit Bandwidth, which increased by a factor of 2.67, in Gossip-based systems. This research demonstrates the potential of eBPF in reducing the overhead of network protocol stacks and system calls, which significantly optimizes the typical user-space broadcast behavior found in distributed consensus algorithms.