Data-Driven Guidelines for Buffer Sizing: A Dynamical System Approach for Throughput Time-Series

Proper router buffer sizing is critical for supporting modern applications that require both high throughput and low latency. An excessively large buffer can induce significant delays due to bufferbloat, whereas undersized buffers often reduce throughput. Deriving theoretical guidelines for buffer sizing is particularly difficult in realistic environments where multiple flows using nonlinear algorithms, such as CUBIC TCP, compete. This paper proposes a novel data-driven method for establishing practical buffer sizing guidelines. Unlike conventional theoretical approaches, this method derives guidelines for proper buffer sizing according to a target bandwidth utilization (normalized throughput) using only throughput time-series data. This method relies on time-series measurements from only three points within a broad parameter space. The proposed method models the macroscopic behavior of bandwidth utilization as a one-dimensional discrete dynamical system (where the next state is iteratively computed via a deterministic rule), using the extreme learning machine with a target of bandwidth utilization (ELM-TBU) model. Here, the target of bandwidth utilization is treated as a fixed point of the dynamical system. The minimum buffer size that satisfies this criterion is numerically determined using concepts from dynamical systems theory. By repeating this process for various numbers of flows and bandwidth-delay products, global design guidelines for buffer sizing can be obtained. Using Network Simulator-2, we demonstrate that the proposed method accurately estimates the proper buffer size (i.e., the minimum buffer size that satisfies the target bandwidth utilization). Performance evaluations indicate that the proposed method is highly effective and applicable across a wide range of environments.