TL;DR
This paper introduces CoCo, a framework for efficiently consolidating modularized network functions (vNFs) in Network Function Virtualization (NFV). CoCo tackles the performance overhead and resource inefficiency introduced by breaking down traditional network functions into smaller, modular components, enabling more compact and performant service function chains.
What Problem Does This Paper Address?
Network Function Virtualization (NFV) allows replacing dedicated hardware network appliances with virtualized Network Functions (vNFs). To accelerate development, vNFs are increasingly being modularized into smaller "elements," forming Modularized Service Function Chains (MSFCs). While this modularity offers benefits like easier development and elasticity, it introduces significant challenges:
- Performance Overhead: Each inter-element communication between vNFs (often in separate Virtual Machines) incurs latency due to packet transfer, significantly degrading overall chain performance. MSFCs require more such transfers than traditional monolithic Service Function Chains (SFCs).
- Resource Inefficiency: Deploying many small vNF elements can consume substantially more hardware resources compared to a single monolithic vNF, compromising efficiency.
Existing solutions have addressed parts of this problem, like merging common elements or consolidating NFs onto CPU cores, but none have holistically tackled the specific challenges of MSFC consolidation, particularly the optimal placement of these modular elements and efficient scaling.
Methods
CoCo introduces a framework with two main components: the CoCo Controller and the CoCo Infrastructure.
- CoCo Controller:
- Optimized Placer: This component addresses the placement problem. It models the performance cost of inter-VM packet transfer and uses a 0-1 Quadratic Programming algorithm to find an optimal placement strategy. The goal is to consolidate adjacent elements on the same VM to minimize inter-VM packet transfers and thus reduce latency.
- Individual Scaler: This component handles the elasticity of MSFCs. It proposes two innovative scaling strategies:
- Push-aside Scaling Up: Instead of creating a new VM for a scaled-out element, this strategy intelligently migrates an upstream or downstream element to a less-loaded VM, freeing up resources on the overloaded element's VM to scale itself up. This avoids introducing new inter-VM hops and associated latency.
- Greedy Scaling Out: If push-aside scaling isn't feasible, this method greedily places new replicas on existing, underutilized VMs to save resources. If no existing VM has enough capacity, new VMs are provisioned.
- CoCo Infrastructure:
- Automatic Scheduler: This component ensures fair resource allocation when multiple elements are consolidated on the same CPU core. It uses an incrementally adaptive algorithm that matches the processing speed of each element with its packet arrival rate, inferred from buffer variations. This avoids the need for manual priority configurations.
Datasets
Not discussed in the paper. The evaluation uses simulations and a testbed with specific hardware configurations.
Key Insights and Ideas
- Modularization's Double-Edged Sword: While modularity in NFV offers flexibility, it inherently introduces performance bottlenecks through increased inter-component communication. CoCo directly addresses this trade-off.
- The Power of Placement: The paper highlights that where modular vNF elements are placed is crucial for performance. CoCo's 0-1 Quadratic Programming approach optimizes this placement to minimize latency.
- Intelligent Scaling: Traditional scaling often means spinning up new resources. CoCo's "push-aside" scaling up is a clever way to rebalance existing resources, achieving scalability without the typical performance penalty of new inter-VM hops.
- Fairness in Consolidation: When multiple vNF elements share a CPU core, ensuring fair resource allocation is vital. CoCo's automatic scheduler achieves this by dynamically matching processing speed to traffic, rather than relying on manual configurations.
Key Results
- Performance Improvement: CoCo significantly reduces inter-VM packet transfer costs. For Topology 1, it reduces the total sum of Delayed Bytes (DB) by 2.46x compared to a random strategy and 1.64x compared to a greedy strategy. For Topology 2, it achieves 1.52x and 1.21x improvements, respectively.
- Efficient Scaling: The push-aside scaling up strategy achieved a converged latency of 6 ms, compared to 11 ms for a traditional method, and improved resource efficiency by 1.5x.
- Reduced Failure Rate: CoCo improves the placement failure rate, reducing it from 7.8% (random) and 6.5% (greedy) to 5.8% for Topology 2.
- Throughput-CPU Mapping Linearity: The paper found a strong linear correlation between processing speed and CPU utilization for different vNF elements, simplifying the modeling process.
Limitations and Future Work
- Multi-core Placement: The current placement analysis assumes one CPU core per VM. The authors suggest extending CoCo to handle VMs with multiple CPU cores.
- Intra-core Analysis: While CoCo focuses on inter-VM costs, future work could explore optimizing intra-core performance, including cache management and element isolation.
- Timeout for Scaling: The paper mentions that administrators can set timeouts for scaling to avoid frequent scaling events, but this is a configuration parameter rather than an inherent limitation of the core algorithm.
- Complex Topologies: While CoCo uses optimization solvers that can handle complex topologies, the paper focuses on evaluating simpler examples. Further evaluation on more intricate network structures would be beneficial.
Summary updated