Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LiSFC-Search: Lifelong Search for Network SFC Optimization under Non-stationary Drifts

Zuyuan Zhang, Vaneet Aggarwal, Tian Lan

TL;DR

Preliminary results on synthetic CPN topologies and SFC workloads show that LiSFC consistently reduces SFC blocking probability and improves tail delay compared to non-transfer MCTS and purely learning-based baselines, highlighting its potential as an AI/ML building block for cloud-network convergence.

Abstract

Edge-cloud convergence is reshaping service provisioning across 5G/6G and computing power networks (CPNs). Service function chaining (SFC) requires continuously placing and scheduling virtual network functions (VNFs) chains under compute/bandwidth and end-to-end QoS constraints. Most SFC optimizers assume static or stationary networks, and degrade under long-term topology/resource changes (failures, upgrades, expansions) that induce non-stationary graph drifts. We propose LiSFC, a Lipschitz lifelong planner that transfers MCTS statistics across drifting network configurations using an MDP-distance bound. More precisely, we formulate the problem as a sequence of MDPs indexed by the underlying network graph and constraints, and we define a \emph{graph drift} metric that upper-bounds the LiZero MDP distance. This allows LiSFC to import theoretical guarantees on bias and sample efficiency from the LiZero framework while being tailored to cloud-network convergence. We then design \emph{LiSFC-Search}, an SFC-aware unified MCTS (UMCTS) procedure that uses transferable adaptive UCT (aUCT) bonuses to reuse search statistics from prior CPN configurations. Preliminary results on synthetic CPN topologies and SFC workloads show that LiSFC consistently reduces SFC blocking probability and improves tail delay compared to non-transfer MCTS and purely learning-based baselines, highlighting its potential as an AI/ML building block for cloud-network convergence.

LiSFC-Search: Lifelong Search for Network SFC Optimization under Non-stationary Drifts

TL;DR

Preliminary results on synthetic CPN topologies and SFC workloads show that LiSFC consistently reduces SFC blocking probability and improves tail delay compared to non-transfer MCTS and purely learning-based baselines, highlighting its potential as an AI/ML building block for cloud-network convergence.

Abstract

Edge-cloud convergence is reshaping service provisioning across 5G/6G and computing power networks (CPNs). Service function chaining (SFC) requires continuously placing and scheduling virtual network functions (VNFs) chains under compute/bandwidth and end-to-end QoS constraints. Most SFC optimizers assume static or stationary networks, and degrade under long-term topology/resource changes (failures, upgrades, expansions) that induce non-stationary graph drifts. We propose LiSFC, a Lipschitz lifelong planner that transfers MCTS statistics across drifting network configurations using an MDP-distance bound. More precisely, we formulate the problem as a sequence of MDPs indexed by the underlying network graph and constraints, and we define a \emph{graph drift} metric that upper-bounds the LiZero MDP distance. This allows LiSFC to import theoretical guarantees on bias and sample efficiency from the LiZero framework while being tailored to cloud-network convergence. We then design \emph{LiSFC-Search}, an SFC-aware unified MCTS (UMCTS) procedure that uses transferable adaptive UCT (aUCT) bonuses to reuse search statistics from prior CPN configurations. Preliminary results on synthetic CPN topologies and SFC workloads show that LiSFC consistently reduces SFC blocking probability and improves tail delay compared to non-transfer MCTS and purely learning-based baselines, highlighting its potential as an AI/ML building block for cloud-network convergence.
Paper Structure (16 sections, 5 theorems, 7 equations, 3 figures, 1 algorithm)

This paper contains 16 sections, 5 theorems, 7 equations, 3 figures, 1 algorithm.

Table of Contents

  1. Introduction
  2. Preliminaries and Problem Setup
  3. Network and SFC Model
  4. SFC Planning MDP
  5. Lifelong Task Family
  6. Graph Drift and MDP Distance
  7. Transferable aUCT for Lifelong Search
  8. Algorithm: LiSFC-Search
  9. Estimating Distances
  10. Complexity and Sample Efficiency
  11. Experiments
  12. Experimental Setup
  13. Scenario 1: Load Variation on a Fixed CPN
  14. Scenario 2: Graph Drift and Transfer Across CPN Configurations
  15. Scenario 3: Cross-task convergence with and without lifelong transfer
  16. ...and 1 more sections

Key Result

Lemma 3.3

Assume that rewards and transitions in the SFC MDPs are Lipschitz with respect to resource margins and connectivity: small changes in capacities/bandwidths and topology induce changes in $R$ and $P$ bounded linearly by $\Delta_{\text{cap}},\Delta_{\text{bw}},\Delta_{\text{spec}},\Delta_{\text{edit}}

Figures (3)

  • Figure 1: SFC blocking probability versus normalized offered load on the base CPN topology $G_0$.
  • Figure 2: Average MCTS simulations per decision as a function of graph drift $\Delta G(G_0,G_k)$ between the base topology $G_0$ and perturbed topologies $G_1,G_2, G_3$. LiSFC adapts its reuse of past statistics based on estimated MDP distance, reducing search effort for small drift while gracefully reverting to standard UMCTS for large drift.
  • Figure 3: Running blocking probability of LiSFC and vanilla UMCTS as a function of the decision index on three perturbed CPN configurations $G_1,G_2,G_3$. Each configuration has a different graph drift $\Delta G(G_0,G_k)$ from the base topology $G_0$. LiSFC converges faster and to lower blocking on tasks with small drift, while gracefully degrading towards UMCTS behavior as the drift increases.

Theorems & Definitions (11)

  • Definition 2.1: Network and SFC
  • Definition 2.2: SFC MDP
  • Definition 2.3: Lifelong sequence
  • Definition 3.1: Graph drift $\Delta G$
  • Definition 3.2: MDP distance $d(M,M')$
  • Lemma 3.3: Graph drift as a proxy for MDP distance
  • Theorem 4.1: Lipschitz aUCT zhang2025lipschitz
  • Corollary 4.2: Graph-drift aUCT bound
  • Corollary 4.3: Graph-drift Transfer UCB
  • Definition 6.1: IS estimator of $d$
  • ...and 1 more