[Paper Review] Cost minimization of network services with buffer and end-to-end deadline constraints

Victor Millnert, Enrico Bini+1 more|arXiv (Cornell University)|Sep 15, 2016

Cloud Computing and Resource Management3 citations

TL;DR

This paper proposes a cost-minimization framework for virtual network function (VNF) service chains under end-to-end deadline and buffer constraints in cloud environments. It models dynamic resource scaling as an optimization problem, deriving control strategies that minimize operational costs while guaranteeing latency bounds through analytical queueing and scheduling models, with key results showing stable performance under constant input flows and disturbance handling via adaptive on-time extension.

ABSTRACT

Cloud computing technology provides the means to share physical resources among multiple users and data center tenants by exposing them as virtual resources. There is a strong industrial drive to use similar technology and concepts to provide timing sensitive services. One such is virtual networking services, so called services chains, which consist of several interconnected virtual network functions. This allows for the capacity to be scaled up and down by adding or removing virtual resources. In this work, we develop a model of a service chain and pose the dynamic allocation of resources as an optimization problem. We design and present a set of strategies to allot virtual network nodes in an optimal fashion subject to latency and buffer constraints.

Motivation & Objective

• Address the challenge of minimizing operational costs in cloud-based virtual network function (VNF) service chains while meeting strict end-to-end deadlines.
• Model the dynamic allocation of virtual machines across VNFs as an optimization problem under variable workloads and buffer constraints.
• Develop control strategies that ensure end-to-end delay bounds are met by analyzing queueing behavior and processing time dynamics.
• Account for transient disturbances such as modeling errors or workload spikes by extending on-time durations to maintain service guarantees.
• Establish a foundation for future adaptive schemes under time-varying request patterns and uncertain service rates.

Proposed method

• Formulates a mathematical model of service chains composed of interconnected VNFs, each with configurable numbers of virtual machines.
• Models the input stream as a constant-rate request flow with end-to-end deadline constraints per packet.
• Introduces a service model where each VNF processes requests in FIFO order with non-preemptive execution, emphasizing queueing delay over computation time.
• Derives a cost model that accounts for compute and storage allocation, leading to a minimization objective over the number of active machines.
• Uses deterministic queuing theory and network calculus to analyze delay bounds, particularly focusing on buffer overflow and disturbance handling.
• Proposes an adaptive on-time extension strategy (Equation 39) to handle unexpected workload surges by extending active periods based on disturbance size and idle time availability.

Experimental results

Research questions

• RQ1 How can virtual network functions be dynamically scaled to minimize cost while ensuring end-to-end delay constraints are met?
• RQ2 What is the impact of queueing delay versus computation time on end-to-end latency in service chains with constant input flows?
• RQ3 How should resource allocation be adjusted when unexpected disturbances (e.g., modeling errors or workload spikes) occur?
• RQ4 What control strategy enables stable operation under buffer constraints while minimizing the number of active virtual machines?
• RQ5 How can the system maintain service guarantees when the required processing time exceeds normal on-periods due to disturbances?

Key findings

• The proposed control strategy ensures end-to-end deadline compliance by dynamically extending the on-time of VNFs based on disturbance size and available idle time.
• The model shows that queueing delay dominates over computation delay in determining end-to-end latency, justifying a focus on buffer and scheduling management.
• Disturbances such as modeling errors or workload spikes can be handled by extending the on-period using Equation (39), which accounts for full and partial off-periods needed to process extra work.
• The cost increase due to disturbance handling is constant and does not affect the optimization solution, as it is independent of the decision variables.
• The framework provides a stable baseline for future adaptive schemes under time-varying workloads, with potential extension to feedback-controlled service rate adaptation.
• Simulation results (not quantified in text) confirm that the model maintains delay bounds and avoids over-provisioning through precise on-time extension.