TL;DR
This paper evaluates OpenVZ and User Mode Linux (UML) for virtualizing the ORBIT wireless testbed. They found OpenVZ significantly outperforms UML in terms of system overhead and performance isolation, making it a more suitable choice for deploying simultaneous wireless experiments on shared infrastructure.
What Problem Does This Paper Address?
As the demand for wireless experimentation testbeds like ORBIT grows, efficiently sharing limited resources becomes crucial. While virtualization offers a solution by allowing multiple experiments to run concurrently, achieving true isolation and flexibility in a wireless environment is challenging. This paper addresses the need to select an appropriate virtualization technology that can multiplex system resources effectively without compromising experiment performance or repeatability, especially for integration with larger frameworks like GENI.
Methods
The authors compare two virtualization approaches: OpenVZ (OS-level virtualization) and User Mode Linux (UML) (para-virtualization). They focus on qualitative criteria important for both testbed administrators (ease of administration, interface mapping, network control) and experimenters (support for Linux distributions, root access within containers). Their experimental evaluation involves measuring throughput, delay, and slice isolation under various network conditions. They use tools like iperf for UDP traffic generation and FTP for file transfers, and ICMP (ping) for delay measurements.
Datasets
Not discussed in the paper. The experiments were conducted on the ORBIT wireless testbed using its available hardware and network configurations.
Key Insights and Ideas
- OS-level virtualization shines for wireless: OpenVZ, as an OS-level virtualization mechanism, offers a more efficient and less intrusive way to virtualize wireless testbeds compared to para-virtualization like UML.
- Performance isolation is key: For wireless experiments, the ability to isolate experiments from each other is more critical than achieving the absolute highest throughput. OpenVZ excels in this area, ensuring experiment repeatability.
- Overhead matters for small packets: UML exhibits significant packet processing overhead with small packet sizes, leading to degraded performance, which is not observed with OpenVZ.
- OpenVZ is "close to the metal": The performance of OpenVZ is very close to that of a non-virtualized system, indicating minimal overhead. This is attributed to its tight virtualization mechanism and efficient packet handling.
- Qualitative factors are crucial: Beyond raw performance, ease of administration and flexibility for experimenters are vital for testbed deployment, and OpenVZ scores well on these fronts.
Key Results
- Throughput: OpenVZ and non-virtualized systems showed no distinct difference in throughput, even under saturation. UML's throughput variance increased significantly with offered load, especially near saturation.
- Packet Size Impact: For packet sizes up to 1024 bytes, UML showed significantly higher packet processing overhead, leading to degraded performance. OpenVZ's performance was largely unaffected by packet size.
- FTP Performance: For larger IP frames (FTP transfers), UML performed on par with OpenVZ and non-virtualized systems, as the overhead of larger frames masked the virtualization differences.
- Transmission Delay: OpenVZ added a very small average overhead (around 0.05ms) in absolute delay, with comparable performance across slices. UML was noted to have performance degradation in similar settings in prior studies.
- Slice Isolation: OpenVZ demonstrated reasonable isolation, with little to no correlation in throughput of one slice when the offered load on another slice changed. Cross-coupling was negligible. Process isolation tests also showed OpenVZ successfully containing processes within containers.
Limitations and Future Work
- Hardware Specificity: The authors note that while the results are hardware-specific, the performance trends are expected to hold and scale with hardware capacity and load.
- Future Work: The paper suggests that integration of OpenVZ with the ORBIT framework and measurement library are important next steps. Future research could also involve comparing OpenVZ with Xen and Vservers on newer Intel chipset-based machines. The study also implicitly suggests that further investigation into optimizing UML or other para-virtualization techniques for wireless environments might be warranted, though OpenVZ is presented as the superior choice for their current needs.
Summary updated