RIPPLE: Lifecycle-aware Embedding of Service Function Chains in Multi-access Edge Computing

In Multi-access Edge Computing networks, services can be deployed on nearby edge clouds (EC) as service function chains (SFCs) to meet strict quality of service (QoS) requirements. As users move, frequent SFC reconfigurations are required, but these are non-trivial: SFCs can serve users only when all required virtual network functions (VNFs) are available, and VNFs undergo time-consuming lifecycle operations before becoming operational. We show that ignoring lifecycle dynamics oversimplifies deployment, jeopardizes QoS, and must be avoided in practical SFC management. To address this, forecasts of user connectivity can be leveraged to proactively deploy VNFs and reconfigure SFCs. But forecasts are inherently imperfect, requiring lifecycle and connectivity uncertainty to be jointly considered. We present RIPPLE, a lifecycle-aware SFC embedding approach to deploy VNFs at the right time and location, reducing service interruptions. We show that RIPPLE closes the gap with solutions that unrealistically assume instantaneous lifecycle, even under realistic lifecycle constraints.

💡 Research Summary

The paper tackles the practical challenge of maintaining uninterrupted service function chains (SFCs) in Multi‑access Edge Computing (MEC) environments where users constantly move across base stations (BSs). Existing works assume that virtual network functions (VNFs) become instantly available once requested, but in reality each VNF must pass through a series of lifecycle stages—download, image creation, deployment, start, pause, etc.—that can take seconds to minutes. Ignoring these delays leads to service interruptions during handovers because a VNF may not be running when a user’s traffic reaches it.

To address this, the authors formulate the SFC embedding and reconfiguration problem with explicit lifecycle dynamics and split it into three sub‑problems: (P1) user connectivity forecasting, (P2) lifecycle‑aware VNF placement, and (P3) virtual link mapping.

P1 – Connectivity Forecasting
User mobility is modeled with a Gauss‑Markov process. A long short‑term memory (LSTM) network predicts future positions over a horizon h, while a random‑forest classifier estimates the probability of connecting to each BS at each predicted location. The horizon length is critical; the best performance is observed when h matches the total VNF lifecycle duration (≈12 seconds).

P2 – Lifecycle‑Aware VNF Placement
VNFs are represented as a finite‑state machine (Descriptor → Source → Image → Stopped → Running → Paused). Transition times are taken from empirical studies (e.g., 12 s download, 100 ms deployment, 530 ms start). Building on the Decreasing First Fit (DFF) heuristic, the authors sort edge clouds (ECs) by distance and available resources, then apply a first‑fit policy that also considers the predicted connection probability for each BS. High‑probability BSs trigger deeper lifecycle progression (up to Running), while low‑probability BSs receive only partial preparation (Image or Stopped). Placement proceeds from the tail of the SFC toward the head, and from the farthest feasible ECs inward, minimizing relocations during handovers and preserving resources.

P3 – Virtual Link Mapping
Given the VNF placement, virtual links are mapped greedily: each VNF connects to the closest instance of its successor, respecting a maximum hop distance and the end‑to‑end (E2E) delay constraint. Because DFF already consolidates VNFs on few ECs, this step remains lightweight.

Evaluation
The authors implement an event‑driven simulator with two topologies: a 16‑BS tree and a realistic city graph. Users follow the Gauss‑Markov mobility model, each requiring an SFC of four VNFs (0.1 ms processing per VNF, 1 ms E2E latency budget). Two baselines are compared: (1) Ideal – solves a relaxed embedding problem with zero lifecycle times (theoretical optimum), and (2) Reactive – solves the same problem with realistic lifecycle times but without any forecasting (purely on‑the‑fly).

Key metrics are (i) unsuccessful packets (either exceeding latency or hitting a non‑running VNF) and (ii) burst length (duration of consecutive service interruptions). Experiments vary the mobility correlation parameter α (controlling randomness) and the forecasting horizon h.

Results show that when h ≈ 11 s (matching the total lifecycle), burst lengths shrink dramatically: only 0.05 % of bursts exceed 2 ms, compared with 3 % exceeding 10 s when h = 0. Longer horizons (>12 s) degrade performance due to prediction errors. With α = 0.9 (highly predictable mobility), RIPPLE’s CDF of unsuccessful‑packet ratios almost overlaps the Ideal curve for more than half of the users (difference <0.01 %). Compared to Reactive, RIPPLE reduces the fraction of unsuccessful packets by roughly 30–40 %. The number of VNFs prepared at each EC scales with the aggregated connection probability, ensuring that limited EC resources are not over‑committed.

Insights and Contributions
The study demonstrates that effective MEC SFC management must jointly consider (a) accurate user connectivity forecasts and (b) the temporal cost of VNF lifecycle transitions. Neither aspect alone suffices; their combination yields near‑optimal QoS while respecting resource constraints. The proposed RIPPLE heuristic is modular: the three sub‑problems can be swapped with alternative machine‑learning models or embedding algorithms, allowing extensibility to other objectives such as energy efficiency.

Limitations and Future Work
The approach relies on sufficiently rich training data for the LSTM and random‑forest models; abrupt network changes or rare mobility patterns could degrade forecast accuracy. Moreover, the current model assumes homogeneous VNF resource demands and static transition times; future work could incorporate variable image sizes, dynamic bandwidth for downloads, and multi‑tenant resource sharing. Reinforcement‑learning based policies and more sophisticated stochastic optimization are promising directions to further close the gap between practical and ideal performance.