TCP and UDP performance for Internet over optical packet-switched networks

Jingyi He, S.-h. Gary Chan

TL;DR

This paper investigates how two key optical packet-switched (OPS) network mechanisms—packet aggregation and deflection routing—impact the performance of TCP and UDP, the workhorses of internet traffic. They find that mixed-flow aggregation boosts throughput for both protocols, while deflection routing significantly improves TCP performance despite introducing out-of-order packets.

What Problem Does This Paper Address?

The internet is transitioning to optical networks to handle ever-increasing bandwidth demands. Optical Packet Switching (OPS) is a promising architecture for this "Optical Internet" because it offers fine-grained bandwidth allocation at the packet level, unlike older wavelength-routed networks. However, the unique characteristics of OPS, such as the lack of optical RAM and the need for efficient packet handling, can significantly affect the performance of familiar internet transport protocols like TCP and UDP. This paper aims to understand these impacts by studying two core OPS mechanisms: packet aggregation (combining multiple IP packets into larger optical packets) and deflection routing (a method to resolve output port contention by sending packets down alternative paths).

Methods

The researchers used the ns-2 network simulator to model an irregular mesh OPS network. They simulated two primary packet aggregation schemes:

  • Mixed-flow aggregation: IP packets from different flows destined for the same egress OPS are combined.
  • Per-flow aggregation: Only IP packets from the same flow are combined. They also simulated a scenario with no aggregation for baseline comparison.

For deflection routing, they introduced a "congestion factor" (p) to represent the probability of a deflected packet being dropped on its alternate path, simulating congestion on deflection routes. They also varied the "deflection cost" (tc), which is the additional delay incurred by a deflected packet. The study focused on TCP (using TCP-Reno with FTP traffic) and UDP (using Poisson packet arrivals) and evaluated performance based on throughput, fairness (using Jain's fairness index), and UDP delay jitter.

Datasets

Not discussed in the paper. The study relies on a simulated network environment rather than real-world datasets.

Key Insights and Ideas

  • Mixed-flow aggregation is a throughput booster: Combining packets from different flows into a single optical packet significantly increases throughput for both TCP and UDP. This is because it smooths out traffic bursts, reducing contention and packet loss at the ingress OPS.
  • Deflection routing is a boon for TCP: Despite causing out-of-order packet delivery, deflection routing substantially improves TCP throughput. This is because packet loss is more detrimental to TCP than out-of-order packets, and deflection helps avoid losses.
  • UDP handles out-of-order packets gracefully: Unlike TCP, UDP doesn't have complex congestion control mechanisms that are sensitive to out-of-order packets. This means deflection routing can lead to even more prominent throughput improvements for UDP compared to TCP.
  • Per-flow aggregation can hurt fairness: While mixed-flow aggregation improves fairness, per-flow aggregation can significantly decrease it, especially at larger aggregation intervals. This is because if a large aggregate packet from a single flow is lost, that flow can be severely starved.
  • Deflection path congestion is a critical factor: The improvement gained from deflection routing is significantly hampered if the deflection paths are congested (high p). This leads to more packet drops, negating the benefits of avoiding the primary path.
  • Deflection cost has a nuanced impact: The extra delay incurred by deflection (tc) doesn't significantly affect TCP or UDP throughput and fairness. However, it directly increases UDP delay jitter, as expected.

Key Results

  • Throughput: Mixed-flow aggregation consistently yielded the highest throughput for both TCP and UDP. Deflection routing improved TCP throughput by approximately 20%.
  • Fairness: Mixed-flow aggregation generally improved fairness, while per-flow aggregation degraded it, especially at larger aggregation intervals. UDP maintained good fairness across most scenarios.
  • Delay Jitter (UDP): Packet aggregation increased UDP delay jitter, with larger aggregation intervals leading to higher jitter. Deflection cost also linearly increased UDP delay jitter.
  • Congestion Factor (p): Higher congestion factors (p) on deflection paths led to lower throughput for both TCP and UDP. TCP throughput dropped sharply as p increased from 0 to 0.1, indicating its sensitivity to packet loss.

Limitations and Future Work

  • Simulation-based: The study relies on simulations, and real-world network conditions might introduce additional complexities.
  • Specific OPS Model: The findings are based on a specific irregular mesh OPS network model with certain parameters. Performance might vary with different network topologies or switch architectures.
  • Deflection Path Congestion: While the study models deflection path congestion using a probability p, a more detailed simulation of background traffic on these paths could provide deeper insights.
  • Future Work: The authors suggest further investigation into the impact of these mechanisms on other internet protocols and more complex OPS network scenarios. Exploring optimal aggregation interval settings for different traffic mixes and network loads would also be valuable.

Summary updated

Abstract

A strong candidate for the future Internet core is optical packet-switched (OPS) network. In this paper, we study the impact of mechanisms as employed in OPS networks on the performance of upper layer Internet protocols represented by TCP and UDP. The mechanisms we investigate are packet aggregation, deflection routing, and ingress buffering. We show that packet aggregation in general improves TCP throughput, and the improvement increases with the aggregation interval (or optical packet size). With the packets destined to the same egress optical switch, aggregation may be done at different granularities: aggregating all the packets (full aggregation), aggregating packets from the same traffic class (per-class aggregation), and aggregating packets from the same flow (per-flow aggregation). We show that with per-class aggregation and per-flow aggregation some flows may be severely penalized in throughput at large aggregation intervals, resulting in significant degradation in TCP fairness, because of the synchronization problem with shared queueing. By using weighted fair queueing (WFQ) at the ingress buffer, in contrast, we show that differentiated QoS (in terms of throughput) can be provisioned for both TCP and UDP traffic even with deflection routing. Deflection routing avoids packet losses, but results in out-of-order packet delivery and increased packet delay jitter. We show that TCP throughput can be significantly improved by deflection routing in spite of the packet reordering, and the UDP packet delay jitter introduced by deflection routing can be alleviated by packet aggregation and ingress buffering. We also show that ingress buffering significantly improves TCP throughput and the ingress buffer only needs a small size (in terms of the number of optical packets).
BibTeX
@article{he2004tcp,
  author = {Jingyi He and S.-H. Gary Chan},
  title = {{TCP and UDP performance for Internet over optical packet-switched networks}},
  journal = {Computer Networks},
  year = {2004},
  doi = {10.1016/j.comnet.2004.02.011},
}
Figure