Abstract
In recent years, much attention has been focused on deploying service function chains (SFCs), each of which is composed of a set of virtual network functions (VNFs) in a specified order. This is a promising approach for enabling cloud service providers to deploy user service requests more flexibly while saving costs. However, less effort has been directed toward meeting heterogeneous needs, such as high throughput or low latency of user service requests with heterogeneous bandwidth demands, especially in data center networks (DCNs). In this paper, we propose an efficient orchestration algorithm for online SFC requests. It first splits a large flow into a number of subflows and replicates the same number of sub-SFCs. Each subflow is redirected to one of these “parallelized” sub-SFCs, which is termed a sub-user request. Then, each sub-user request is deployed based on a worst-fit strategy, and VNFs in the same SFC are instantiated on the same server to the greatest possible extent. Our algorithm is expected to enable network load balancing, reducing the delay experienced by small flows while improving the acceptance ratio for user requests. Finally, the simulation results show that the proposed algorithm outperforms other comparable algorithms.
Introduction
Modern Data Center Networks (DCNs) are characterized by a diverse traffic mix. They must simultaneously support "elephant flows"—large, long-lived, and throughput-sensitive traffic like video streaming or VM migrations—and "mouse flows"—small, short-lived, and highly latency-sensitive traffic like web searches.
When these flows are orchestrated through Service Function Chains (SFCs), which are ordered sequences of Virtual Network Functions (VNFs) such as firewalls, IDS, or load balancers, they often compete for the same physical links and server resources. Standard multipath routing techniques frequently suffer from "hash collisions," where multiple elephant flows are mapped to the same link, causing significant queuing delays. These delays are particularly detrimental to mouse flows, ruining the user experience. This paper addresses the challenge of online SFC orchestration by balancing the load to satisfy both high-throughput and low-latency requirements.
Methodology: The ONP_SFO Algorithm
The authors introduce the Online Parallelized SFC Orchestration (ONP_SFO) algorithm. Given that SFC orchestration is an NP-hard problem, ONP_SFO employs a heuristic approach that combines flow splitting with strategic resource mapping.
Flow Splitting and SFC Replication
The defining feature of ONP_SFO is the parallelization of SFCs. When a flow's bandwidth demand exceeds a predefined threshold ($k_{sub}$), it is classified as an elephant flow and split into multiple smaller subflows. For each subflow, the original SFC is replicated, creating a set of "parallelized" sub-user requests.
This approach serves two purposes: it allows elephant flows to utilize multiple paths simultaneously, and it creates "gaps" in the network traffic, allowing latency-sensitive mouse flows to pass through with minimal delay.
Worst-Fit Mapping and VNF Locality
To deploy these sub-requests, ONP_SFO utilizes a Worst-Fit Mapping Strategy. While many cloud providers use "best-fit" algorithms to consolidate traffic and save energy, ONP_SFO prioritizes performance by selecting links and servers with the most available resources. This naturally distributes traffic across the fat-tree topology, preventing hotspots.
Furthermore, the algorithm emphasizes VNF Locality. By attempting to host all VNFs of a single sub-SFC on the same physical server, the algorithm reduces internal network traffic (link mapping cost). This efficiency creates a virtuous cycle, leaving more bandwidth available for other requests and improving the overall acceptance ratio.
Experimental Evaluation
The performance of ONP_SFO was evaluated through extensive simulations modeled after real-world data center characteristics.
Simulation Setup
The simulation environment utilized a 4-pod fat-tree DCN topology. This structure consists of 36 nodes, including 16 servers and 20 switches, organized into Core, Aggregation, and Edge layers.
The traffic profile consisted of 20,000 SFC requests following a Poisson distribution. Bandwidth demands were set between 2 kbps–1 Mbps for mouse flows and 1 Mbps–50 Mbps for elephant flows. ONP_SFO was benchmarked against NF-LGT (Nearest First - Local Global Transformation), a standard greedy orchestration baseline.
Key Results
The simulation results demonstrated the clear advantages of the parallelized approach:
- Improved Acceptance Ratio: ONP_SFO consistently outperformed the baseline, particularly in high-load scenarios where the acceptance ratio for elephant flows improved by at least 72%.
- Enhanced Load Balancing: The algorithm maintained a tight distribution of load, with a 95th-percentile link utilization ratio of approximately 0.1, compared to over 0.25 for the baseline.
- Reduced Latency and Cost: Mouse flows experienced significantly lower queuing delays. Additionally, the average mapping cost was reduced by up to 63% because the algorithm successfully avoided long, multi-hop paths between VNFs.
Discussion and Conclusion
The ONP_SFO algorithm demonstrates that parallelization is a highly effective strategy for managing heterogeneous traffic in DCNs. By splitting elephant flows, the network avoids the congestion bottlenecks that typically penalize small, latency-sensitive flows. While the worst-fit strategy may seem counter-intuitive for energy conservation, it is essential for maintaining the performance levels required by modern cloud services.
Future research will focus on managing the packet reordering overhead introduced by flow splitting and extending the orchestration model to distributed data center environments with more complex heterogeneous constraints.
Summary updated