TL;DR
This paper investigates the causes of burstiness in internet traffic at short time scales (milliseconds), a phenomenon that impacts network performance. The authors demonstrate that TCP's self-clocking mechanism, combined with network queueing, can create ON-OFF traffic patterns, leading to this burstiness, especially for large bulk transfers. They also show that ideal TCP pacing can significantly smooth this traffic.
What Problem Does This Paper Address?
Internet traffic is known to be "bursty," meaning it arrives in uneven, clustered patterns rather than a smooth, predictable flow. While this burstiness has been studied at longer time scales (hours) and attributed to factors like heavy-tailed session durations, its causes at short time scales (milliseconds) remained less clear. This short-scale burstiness is crucial because it can significantly impact the performance of network devices like routers and switches, potentially leading to congestion and increased latency. Previous research had offered conflicting explanations, and the role of the dominant transport protocol, TCP, in this short-scale phenomenon was not fully understood. This paper aims to pinpoint the specific TCP mechanisms responsible for short-scale burstiness and explore ways to mitigate it.
Methods
The core of the paper's methodology revolves around MultiResolution Analysis (MRA), specifically using energy plots. MRA decomposes traffic signals into different time scales, and energy plots visualize the variance of these signals at each scale. A key innovation here is using these plots to compare traffic burstiness against a Poisson process as a baseline. A traffic process is deemed "bursty" if its energy plot is higher than that of a Poisson process at a given scale.
The authors also employ synthetic traffic models to isolate and study specific mechanisms. They develop simplified models of TCP behavior, including:
- Periodic interarrivals: Simulating a TCP flow that transmits packets at a constant rate within a round-trip.
- One-level ON-OFF interarrivals: Modeling TCP's behavior without cross-traffic, where packets within a window are sent back-to-back.
- Two-level ON-OFF interarrivals: Incorporating cross-traffic, which breaks down TCP's window into smaller bursts.
Finally, they use a semi-experimental approach by artificially applying pacing to real traffic traces to evaluate its effectiveness.
Datasets
The primary dataset used for evaluation is a unidirectional OC-48 trace from an Internet2 backbone link (Abilene network), collected using a DAG 4.23 monitor. This trace was analyzed in segments, and subsets of flows were created based on characteristics like TCP type, RTT, capacity, and flow size. The authors also reference other NLANR traces for validation.
Key Insights and Ideas
- TCP Self-Clocking is a Primary Culprit: The paper strongly argues that TCP's self-clocking mechanism, where a sender transmits a new packet upon receiving an acknowledgment (ACK), is a key driver of short-scale burstiness. This is particularly true when combined with network queueing.
- ON-OFF Patterns Emerge Naturally: TCP self-clocking, especially with network queueing, can shape packet interarrivals into ON-OFF patterns, creating strong correlations and burstiness within time scales up to the Round-Trip Time (RTT) of the connection.
- Bandwidth-Delay Product Matters: Bulk TCP transfers with a large bandwidth-delay product relative to their congestion window size are more prone to exhibiting this burstiness. They can build up large windows, leading to clustered packet transmissions.
- Aggregation Doesn't Automatically Smooth: Contrary to classical assumptions about point processes, aggregating many TCP flows doesn't necessarily converge to a Poisson stream. The burstiness of individual flows can persist in the aggregate traffic.
- Pacing is a Powerful Solution: Implementing TCP pacing at the sender can significantly reduce short-scale burstiness. The effectiveness of pacing is highly dependent on the minimum pacing timer; a 1ms timer is shown to be much more effective than a 10ms timer.
- RTT as a Burstiness Boundary: The paper observes that the short-scale burstiness in aggregate traffic often extends up to the RTT of the dominant TCP flows, and the energy plot shows a characteristic dip around this scale.
Key Results
- Synthetic Models Show Burstiness: The paper demonstrates through energy plots that their synthetic models of TCP self-clocking (both one-level and two-level ON-OFF) exhibit significant burstiness in sub-RTT scales, outperforming the Poisson baseline.
- Real Trace Analysis Confirms Findings: Analysis of the OC-48 trace reveals that short-scale burstiness extends up to approximately 200ms. This burstiness is primarily driven by large TCP flows (over 15KB) with a high "burstiness ratio" (bandwidth-delay product divided by typical window size, > 2.0).
- Pacing Effectiveness: Ideal TCP pacing completely smooths traffic in sub-RTT scales. Practical pacing with a 1ms minimum timer significantly reduces burstiness, bringing it close to Poisson levels, while a 10ms timer is less effective, leaving residual burstiness.
Limitations and Future Work
- Open-Loop Models: The synthetic TCP models used are "open-loop," meaning they don't account for the impact of network state (like congestion) on TCP behavior. This is acknowledged as a simplification.
- Semi-Experimental Pacing: The pacing experiments are semi-experimental, meaning they artificially modify traffic traces without fully simulating the feedback loops of a real TCP connection (e.g., how pacing might affect throughput or packet loss).
- Trace Representativeness: The authors note that the NLANR traces, primarily from research networks, might not be fully representative of commercial Internet traffic in terms of flow size and capacity distributions.
- Future Work: The paper suggests that further investigation into the impact of pacing on TCP dynamics and the analysis of more diverse traffic traces would be valuable. Understanding how different TCP congestion control algorithms interact with pacing and self-clocking could also be a fruitful area.
Summary updated