Learning-Augmented Online Packet Scheduling with Deadlines

TL;DR

This work addresses online packet scheduling with deadlines under buffer management by introducing a learning-augmented framework that leverages predictions to improve performance. The LAP algorithm computes an offline predicted schedule via a maximum-weight matching, then uses a local test with threshold $\rho$ to switch between following predictions and a baseline online algorithm, achieving $1$-consistency and a bounded robustness of $\gamma_{\texttt{on}} + c$; it also provides a tight trade-off between consistency and robustness and shows the competitive ratio can be a smooth function of the prediction error $\eta$. Theoretical results quantify robustness and consistency guarantees, while experiments on synthetic and real data demonstrate strong performance when predictions are accurate and bounded degradation under worse predictions. The findings have practical impact for QoS-enabled network switches by enabling adaptive buffer management that exploits predictions while maintaining guarantees in adversarial settings, and they open directions for alternative error measures and extensions such as FIFO models.

Abstract

The modern network aims to prioritize critical traffic over non-critical traffic and effectively manage traffic flow. This necessitates proper buffer management to prevent the loss of crucial traffic while minimizing the impact on non-critical traffic. Therefore, the algorithm's objective is to control which packets to transmit and which to discard at each step. In this study, we initiate the learning-augmented online packet scheduling with deadlines and provide a novel algorithmic framework to cope with the prediction. We show that when the prediction error is small, our algorithm improves the competitive ratio while still maintaining a bounded competitive ratio, regardless of the prediction error.

Figures (3)

• Figure 1: Illustration of one iteration of the local test (Lines 4-7) of the LAP algorithm. The first two rows with a gray background represent the job following the predicted schedule $\hat{\mathcal{S}}(\hat{\mathcal{J}})^{(t)}$ and the corresponding job in realization $\hat{\mathcal{S}}(\mathcal{J})^{(t)}$ at time $t$, while the last two rows indicate the processed jobs $\mathcal{P}$ (yellow) as well as two potential operations for LAP and the jobs in $\textsc{Opt}(\mathcal{J}_{\le t})^{(\le t)}$. For each local test at time $t$, LAP compares $W(\textsc{Opt}(\mathcal{J}_{\le t})^{(\le t)})$ with $W(\mathcal{P} \cup \hat{\mathcal{S}}(\mathcal{J})^{(t)})$ to decide whether LAP processes either $\hat{\mathcal{S}}(\mathcal{J})^{(t)}$ (green) or $\textsc{OnlineAlg}(\mathcal{J} \setminus \mathcal{P} )^{(t)}$ (orange). In this example, we assume $\hat{\mathcal{S}}(\mathcal{J})^{(t)} \notin \mathcal{P}$ and $\mathcal{D} (t,\mathcal{J}) = \emptyset$.
• Figure 2: The competitive ratio achieved by our algorithm, LAP, and the benchmark algorithms, as a function of the error parameter $\sigma$.
• Figure 3: The competitive ratio achieved by our algorithm, LAP, and the benchmark algorithms, as a function of deadline shift $[-k, k]$.

Theorems & Definitions (22)

• Definition 2.1: Dominance Relation jez2012online
• Lemma 3.1
• proof
• Lemma 3.2
• proof
• Lemma 3.3
• proof
• Lemma 3.4
• proof
• Lemma 3.5