TL;DR
Enterprise Wi-Fi networks often struggle with performance issues caused by hidden and exposed terminals, which standard distributed access methods can't effectively handle. This paper introduces CENTAUR, a hybrid approach that centrally schedules traffic for these problematic scenarios while using the standard DCF protocol for everything else, significantly boosting overall network performance without requiring client-side changes.
What Problem Does This Paper Address?
Enterprise Wireless Local Area Networks (WLANs) have increasingly adopted centralized control for management and security. However, the data path, responsible for actual data transmission, largely relies on the distributed Coordination Function (DCF) from the 802.11 standard. While theoretical gains from centralized scheduling exist, practical implementation has been hindered by complexity. The core problem this paper tackles is the performance degradation in enterprise WLANs caused by two specific interference scenarios:
- Hidden Terminals: When two stations (or a station and an Access Point) cannot hear each other but are both transmitting to a common receiver, leading to collisions and lost packets.
- Exposed Terminals: When a station can hear another station transmitting but is not part of that transmission. Standard protocols might cause this station to defer transmission unnecessarily, even though it could transmit without causing interference.
These issues, particularly when they occur on the downlink (from AP to client), significantly reduce throughput and overall network efficiency. Existing solutions often require modifications to client devices or the 802.11 standard itself, making them impractical for widespread adoption.
Methods
CENTAUR proposes a hybrid data path that combines centralized scheduling with the existing DCF mechanism. The key ideas are:
- Selective Centralization: Only traffic identified as problematic (downlink hidden or exposed terminals) is sent to a central controller for scheduling. All other traffic (uplink, non-hidden/exposed downlink) uses the standard DCF.
- Epoch-Based Scheduling: Instead of scheduling each packet individually (which introduces significant overhead and latency), CENTAUR groups packets into "epochs" (time windows, typically 5-10ms). This amortizes the scheduling overhead and makes the system more robust to variations in wired network latency between the controller and APs.
- Packet Staggering: For exposed terminals, CENTAUR uses a technique to slightly delay the start of transmissions from different APs within an epoch. This allows for more concurrent transmissions without disabling carrier sensing, which would be problematic.
- Fixed Backoffs: To ensure fairness with DCF traffic, CENTAUR APs use a fixed backoff value that approximates the average deferral time of DCF.
- Conflict Graph: A central controller maintains a conflict graph to understand which AP-client links interfere with each other, enabling informed scheduling decisions.
This hybrid approach aims to leverage the benefits of centralized scheduling for specific problems while retaining the simplicity and compatibility of DCF.
Datasets
The paper does not explicitly mention using pre-existing public datasets for training. Instead, it relies on:
- Production WLAN Traces: Data from enterprise WLANs (UCSD Jigsaw, Microsoft Research, and a major enterprise vendor) were analyzed to understand traffic patterns and the prevalence of hidden/exposed terminals.
- Real-World Testbeds: The authors deployed and evaluated CENTAUR on two distinct testbeds:
- Testbed 1: A multi-floor setup with 30 Soekris nodes and Atheros-based 802.11 NICs.
- Testbed 2: A single-floor setup with 20 VIA nodes and Intel 2915 802.11 ABG NICs.
- SIGCOMM 2004 Conference Traces: HTTP traffic traces were replayed on the testbeds to simulate realistic enterprise workloads.
Key Insights and Ideas
- Hidden and Exposed Terminals are Real Problems: The paper provides empirical evidence that these classic Wi-Fi issues are not just theoretical but significantly impact performance in real-world enterprise networks, especially for downlink traffic.
- DCF is Surprisingly Robust, but Not Perfect: While DCF handles many scenarios well, it falters significantly in the presence of hidden and exposed terminals, where centralized scheduling offers a clear advantage.
- Selective Centralization is Key: A full, always-on centralized scheduler for all traffic is likely too complex and introduces unnecessary overhead. CENTAUR's hybrid approach, targeting only problematic traffic, strikes a better balance.
- Epoch-Based Scheduling Masks Latency: The variability in wired network delays between the controller and APs is a major hurdle for centralized scheduling. Epoch-based scheduling effectively hides this variability, making the system practical.
- No Client Changes Needed: A critical insight is that CENTAUR can be implemented solely by the WLAN vendor, without requiring any modifications to client devices, which is a major barrier to adoption for many advanced Wi-Fi features.
Key Results
- Significant Throughput Gains: CENTAUR achieved substantial throughput improvements, particularly for hidden and exposed terminal links. For example, in HT-only scenarios, gains of up to 6x were observed.
- Overall Network Improvement: Even in mixed topologies, CENTAUR provided aggregate throughput gains of 44-61.5% over DCF for UDP and TCP traffic, respectively.
- Reduced Latency: UDP delay was reduced by up to 47.4% with CENTAUR, especially for hidden terminal links.
- Improved VoIP Quality: Voice-over-IP (VoIP) traffic saw improved Mean Opinion Scores (MOS), with CENTAUR achieving a MOS of 3.75 compared to DCF's 3.35, and significantly improving MOS for hidden terminal links.
- Effective Co-existence: CENTAUR demonstrated good co-existence with unscheduled and uplink traffic, with improvements in scheduled traffic even benefiting unscheduled traffic due to better medium utilization.
Limitations and Future Work
- Epoch Duration Trade-off: The choice of epoch duration involves a trade-off between scheduling efficiency and packet latency. While 5-10ms was found to be a good balance, variable epoch sizes for different applications could further optimize performance.
- Conflict Graph Overhead: While micro-probing for conflict graph updates has minimal overhead, active probing can still be a concern in highly dynamic environments. Future work could explore passive conflict graph generation.
- Uplink Traffic Impact: While CENTAUR improves overall network performance, the paper notes that the gains are higher when the fraction of unscheduled traffic (including uplink) is lower.
- PHY Rate and ARF: While CENTAUR handles different PHY rates and ARF, gains for higher rates or in scenarios with many exposed terminals might be slightly less pronounced as faster transmissions can sometimes mask exposed links.
The authors suggest exploring variable epoch sizes for different application classes and further investigating passive conflict graph generation techniques.
Summary updated