Impact of background traffic on performance of high-speed TCP variant protocols

Sangtae Ha, Long Le, Injong Rhee, Lisong XuComputer Networks 2007 · DOI

TL;DR

This paper investigates how background network traffic affects the performance of various high-speed TCP (HSTCP) variants. The authors found that background traffic, particularly with variable flow sizes and RTTs, can significantly improve fairness among HSTCP flows but also introduce instability for some protocols. Ultimately, the amount of background traffic, rather than its variability, is the primary factor influencing HSTCP fairness.

What Problem Does This Paper Address?

The internet's backbone is increasingly characterized by high-speed, long-distance links. Traditional TCP struggles to efficiently utilize this capacity, leading to the development of numerous "high-speed TCP variants" (e.g., BIC-TCP, CUBIC, FAST, HSTCP, H-TCP). While these protocols aim to improve performance, their evaluations have often been conducted in simplified network environments, lacking realistic background traffic. This omission is problematic because background traffic can introduce complex dynamics like queue fluctuations and packet loss patterns that significantly impact congestion control algorithms. Without considering these factors, performance evaluations might be misleading, leading to incorrect conclusions about a protocol's real-world behavior. This paper aims to bridge this gap by systematically evaluating HSTCP variants under various realistic background traffic conditions.

Methods

The researchers constructed a laboratory network testbed using a dumbbell topology with a bottleneck link. They employed FreeBSD machines as routers, with one router emulating the bottleneck (400 Mbps bandwidth, 1-2 MB buffer) and the other introducing per-flow delays to simulate wide-area network conditions.

Key methodological aspects include:

  • Background Traffic Generation: Two types of background traffic were used:
  • Long-lived flows: Emulated using iperf (TCP-SACK).
  • Short-lived flows: Emulated using a modified SURGE tool, generating Web sessions with file sizes following log-normal and Pareto distributions.
  • Propagation Delay Emulation: Dummynet was extended to assign random, exponentially distributed round-trip times (RTTs) to background flows, mimicking real-world network variability.
  • High-Speed Flow Implementation: Implementations of BIC-TCP, CUBIC, FAST, HSTCP, H-TCP, and Scalable TCP were used. A fix was developed for selective acknowledgment handling in Linux kernels to improve performance for high-speed protocols.
  • Experimental Setup: Experiments involved running two high-speed flows of the same type, with fixed RTTs for these flows. The buffer size of the bottleneck router was varied between 1 and 2 MB.
  • Metrics: Key performance metrics included link utilization, throughput, packet loss rates, queue fluctuations, fairness (intra-protocol, RTT-fairness, TCP-friendliness), and stability (measured by the coefficient of variance of total forward traffic rates).

Datasets

The paper does not use pre-existing datasets for training. Instead, it generates its own network traffic and experimental conditions. The background traffic generation relies on statistical distributions (log-normal and Pareto) for flow sizes and exponential distributions for inter-arrival times, which are stated to be representative of Internet traffic characteristics based on existing measurement studies (e.g., [1, 4, 13, 24]). The paper also refers to a previous work [14] for some background traffic generation details.

Key Insights and Ideas

  • Background Traffic is Crucial: The study strongly reinforces the idea that evaluating HSTCP protocols without background traffic is "dangerous" and can lead to incorrect conclusions. Realistic network conditions are essential for accurate performance assessment.
  • Fairness Improvement: The presence of background traffic, especially with variable flow sizes and RTTs, generally improves the fairness of most HSTCP protocols. This is attributed to the increased randomness in packet losses, which helps break synchronization among competing flows.
  • Amount vs. Variability: The amount of background traffic has a more significant impact on protocol fairness than the variability (e.g., RTT or flow size distribution) of that traffic.
  • Protocol-Specific Vulnerabilities:
    • H-TCP: Trades stability for fairness. While fair, it becomes unstable and exhibits low link utilization under variable background traffic due to loss synchronization.
  • FAST: Shows instability and unfairness, particularly in small buffer or long-delay networks, and its behavior can be unpredictable with background traffic. Its performance is sensitive to the amount of background traffic.
  • HSTCP: Requires a larger amount and more variability in background traffic to achieve convergence.
  • STCP: Benefits significantly from background traffic, improving its link utilization and reducing loss synchronization.
  • Web Traffic Impact: High-speed flows do not significantly reduce the bandwidth usage of background Web traffic.

Key Results

  • Link Utilization:
    • Without background traffic, most HSTCP protocols achieve around 95% link utilization. STCP and HSTCP show some degradation at higher RTTs. FAST exhibits abnormal behavior at 324 ms RTT, with flows not exceeding 200 Mbps.
  • With Type II background traffic (chosen for its significant impact), link utilization changes dramatically for some protocols. STCP's utilization improves substantially, while H-TCP and FAST drop significantly.
  • Fairness:
    • Intra-protocol fairness: Generally improves with background traffic for most protocols. H-TCP and CUBIC show good performance.
  • RTT-fairness: Most protocols show improvement with background traffic. BIC-TCP and STCP see significant gains. H-TCP and FAST maintain RTT-fairness close to 1, meaning they share bandwidth equally regardless of RTT, which the authors note might not be ideal.
  • TCP-friendliness: Also improves with background traffic for most protocols. H-TCP and CUBIC show the most improvement.
  • Stability (CoV): H-TCP exhibits much higher CoV values than other protocols (except FAST and SACK) across various RTTs, indicating significant fluctuations in throughput. RTT variations in background traffic play a role in H-TCP's instability.
  • FAST Behavior: FAST's performance is sensitive to the amount of background traffic, not its size distribution. Even small amounts can lead to incorrect RTT estimations and unfairness.

Limitations and Future Work

  • Testbed Realism: The authors acknowledge that their laboratory testbed, while more realistic than previous studies, does not perfectly replicate the entire Internet. They make no claim about the absolute realism of their models, only that they are more realistic than those without background traffic.
  • Protocol Coverage: The paper tested a selection of high-speed TCP protocols and did not include all published variants.
  • Background Traffic Types: While they explored several types, the full spectrum of Internet background traffic patterns may not have been covered.
  • Future Work: The authors suggest further investigation into why protocols behave as they do, which depends on their specific performance tradeoffs. They also aim to improve their testing methodology for greater objectivity, simplification, and automation to facilitate reproducible research. They are interested in exploring how different testbed setups might yield different results.

Summary updated

BibTeX
@article{ha2007impact,
  author = {Sangtae Ha and Long Le and Injong Rhee and Lisong Xu},
  title = {{Impact of background traffic on performance of high-speed TCP variant protocols}},
  journal = {Computer Networks},
  year = {2007},
  doi = {10.1016/j.comnet.2006.11.005},
}
Figure