TL;DR
This paper introduces Oasis, a novel network stack that allows legacy operating systems to seamlessly leverage overlay networks for improved packet delivery. By providing fine-grained control over routing preferences and dynamically selecting the best overlay or underlay path, Oasis aims to unlock the full potential of overlay services for a wider range of applications.
What Problem Does This Paper Address?
The proliferation of overlay networks, powering applications like Akamai, Kazaa, and BitTorrent, has demonstrated their utility. However, the widespread adoption of general-purpose overlay packet delivery services has been hampered by the lack of an "overlay-aware" network stack. Existing operating systems treat all network traffic uniformly, forcing users to either rely on custom overlay protocols or accept the limitations of the underlying Internet. This creates a significant barrier for end-users who want to benefit from the enhanced reliability, performance, or specialized routing offered by overlays without resorting to complex configurations or custom software. The paper argues that this missing piece – an intelligent, adaptable network stack – is the critical bottleneck preventing overlays from reaching their full potential.
Methods
Oasis introduces a system and toolkit that acts as an intermediary between legacy applications and overlay networks. Its core architecture comprises several key components:
- User Control Interface: Allows users to specify routing preferences, such as "route all traffic on port 80 via the Web optimizing overlay" or performance goals like "use an overlay that limits loss rate to at most 1%."
- Overlay Manager: Manages a list of available overlays and dynamically selects the most appropriate one based on user preferences and real-time performance metrics.
- Monitor: A brokering agent that collects end-to-end path metrics for different overlays and helps the Overlay Manager make informed decisions.
- Virtual Network Interface (VNI): A crucial component that intercepts network packets from user-space, allowing Oasis to inspect and redirect them.
- Dynamic Code Execution: Oasis can safely execute custom overlay code modules supplied by Overlay Service Providers (OSPs). This allows for extensibility and the integration of diverse overlay functionalities without modifying the core Oasis system.
- Name-Based and Header-Based Interception: Oasis intercepts traffic either by manipulating DNS resolution (name-faking) or by inspecting packet headers, providing flexibility in how traffic is identified and redirected.
- Fallback to Underlay: A critical design principle is to "do no harm." If no overlay can improve performance beyond the existing Internet, Oasis seamlessly falls back to standard underlay routing.
Datasets
Not discussed in the paper. The evaluation section describes experiments using ping probes to .edu domains and fetching files (index.html, Fedora Core 4 ISO) from .edu mirrors.
Key Insights and Ideas
- The "Missing Link" is the Network Stack: The paper's central thesis is that the lack of an overlay-aware network stack is the primary impediment to overlay adoption, not the overlays themselves.
- User-Centric Control is Key: Empowering users with fine-grained control over their traffic routing, based on specific goals and preferences, is essential for making overlays accessible and useful.
- Dynamic Overlay Selection: The ability to dynamically choose the best overlay (or underlay) for a given flow, based on real-time performance, is a significant advantage over static routing.
- Safe Extensibility via Sandboxed Code: Allowing OSPs to provide custom code modules that run in a sandboxed environment is a clever way to achieve extensibility without compromising security.
- "Do No Harm" Principle: Ensuring that Oasis does not degrade performance or fault tolerance compared to the existing Internet is a crucial design goal for user trust and adoption.
- Minimalist Design for Interoperability: Oasis aims for a "thin waist," similar to IP, to remain agnostic to specific overlay implementations and facilitate independent evolution of overlays and the core system.
Key Results
- Low Latency Overhead: Experiments showed that the median additional latency for individual packets due to Oasis is around 2ms, representing about 4% of the average latency.
- Negligible Impact on Transfer Times: For both short (index.html) and long (20MB ISO files) TCP flows, the median increase in transfer time was 0, indicating that the per-packet overhead does not significantly affect overall transfer performance.
- Reasonable CPU Utilization: Oasis' CPU utilization remained below 20% even at a download rate of 10Mbps. However, upload rates above 2.5Mbps showed a steep increase in CPU usage, making it unusable above 3Mbps, attributed to the overhead of packet interception.
- Scalable DNS Interception: The overhead incurred by DNS lookups with up to 100 name-based rules was comparable to that of normal packets.
Limitations and Future Work
- Upload Performance Bottleneck: The steep increase in CPU utilization at higher upload rates is a significant limitation, suggesting that the current packet interception mechanism on the upload path needs optimization.
- GUI for User Preferences: Currently, user preferences are specified via a Perl script, and the authors plan to replace this with a graphical user interface for better usability.
- PlanetLab Integration: The paper mentions efforts to make Oasis available for PlanetLab, indicating a focus on research and development in distributed environments.
- Developing New Overlay Services: The authors are actively exploring the use of Oasis for building new services like multipath routing, congestion control, and incentive-compatible mechanisms for cooperative routing.
- Public Release: Oasis and its toolkit are being made available for public use, which will likely foster further research and development in this area.
Summary updated