Abstract
In the context of the ever-developing virtual network, OpenStack officially launches its new network component-Quantum, but the comparative performance between singlehost and multi-host deployment is yet unknown. The paper introduces detailed deployment strategies of single-host and multi-host after analyzing Quantum work mechanism and the construction polices of virtual networks. And then, we probe into the influence of various deployments on the reliability of network service, and conduct connectivity and performance tests on different deployment plans. Experiment results show that our virtual network multi-host deployment plan improves virtual network service reliability and efficiency compared with single-node node.
Introduction
As cloud computing infrastructures mature, the efficiency and reliability of virtual networking have become critical performance indicators. In early iterations of OpenStack, the networking component—then known as "Quantum" (now Neutron)—often relied on a centralized architecture. This research addresses the inherent limitations of such designs, specifically the Single Point of Failure (SPoF) and performance bottlenecks associated with a dedicated Network Node. By evaluating a transition to a multi-host deployment strategy, this study aims to quantify improvements in both service availability and network throughput under high-traffic conditions.
Methodology: Single-host vs. Multi-host Deployment
The core of this investigation lies in comparing two distinct architectural paradigms within the OpenStack Grizzly release.
Single-host Architecture
In the traditional single-host deployment, a dedicated physical node is designated as the "Network Node." This node hosts all Layer 3 (L3) routing, DHCP services, and metadata agents. While simple to manage, this creates a significant bottleneck; all traffic between virtual machines (VMs) on different subnets or traffic destined for the external internet must pass through this single point. If this node fails, the entire cloud's network connectivity is severed.
Multi-host Architecture
The proposed multi-host strategy decentralizes these services. By installing L3 and DHCP agents directly on every compute node, the network services are distributed across the cluster. Each compute node manages the traffic for its local VMs, effectively turning every host into a localized network gateway. This architecture significantly reduces the "blast radius" of a failure, as a node crash only affects the VMs residing on that specific hardware.
To further bolster reliability, the authors implemented an agent scheduling mechanism. This model designates "default" and "coordinating" agents. If a local agent fails, the system can dynamically reroute traffic through a neighbor node's agent, ensuring continuous connectivity for the affected VMs.
Experimental Setup
Performance was evaluated using the Distributed Internet Traffic Generator (D-ITG). This tool allowed for the simulation of diverse traffic patterns, including intra-node VM communication, inter-node VM communication, and external host access. The primary metrics measured were latency and packet loss under varying loads, specifically testing the system's response as communication volume scaled up to 1,000 times the baseline to simulate high-volume enterprise environments.
Results and Analysis
The experimental results highlight three key advantages of the multi-host deployment:
- Elimination of SPoF: Connectivity tests confirmed that the failure of a single network agent no longer results in a total cloud outage. The coordination model successfully maintained VM reachability during simulated agent failures.
- Scalability and Latency: While multi-host setups may introduce slight overhead for small packets due to increased agent interaction, they demonstrate superior performance under high load. As traffic volume increased, the latency growth in the multi-host setup was significantly flatter than the exponential growth seen in the single-host bottleneck.
- Resource Efficiency: By distributing the processing load across the entire physical cluster, the multi-host plan avoids the hardware saturation common in centralized network nodes, leading to more efficient utilization of the underlying infrastructure.
Conclusion
The transition from a single-node to a multi-host network architecture in OpenStack represents a significant step toward enterprise-grade reliability. While the decentralized model requires more complex scheduling logic, the gains in service availability and high-load performance are substantial. Future research will focus on optimizing the redundancy ratios to balance safety with resource overhead and developing more sophisticated load-balancing algorithms for agent coordination.
Summary updated