TL;DR
This paper investigates how to optimize TCP throughput in Differentiated Services (DiffServ) networks by strategically marking acknowledgment (ACK) packets. The authors demonstrate that simply prioritizing data packets isn't enough; marking ACKs based on network congestion is crucial for achieving target performance, especially for Premium and Assured traffic.
What Problem Does This Paper Address?
The internet relies heavily on TCP for reliable data transfer. TCP is a bi-directional protocol, meaning it uses acknowledgment packets (ACKs) to confirm receipt of data and to control its sending rate. When a network offers different service levels (like Premium, Assured, and Best-Effort) through DiffServ, the core challenge is ensuring that traffic receives its promised quality of service. Previous approaches often focused on marking and prioritizing data packets. However, this paper highlights a critical oversight: congestion on the reverse path, where ACKs travel, can severely degrade TCP performance, even if the forward path is well-provisioned. This is because TCP interprets a lack of ACKs as packet loss, leading it to reduce its sending rate. The problem, therefore, is to understand and address the impact of ACK packet marking and reverse path congestion on overall TCP throughput in DiffServ environments.
Methods
The researchers employed a rigorous experimental approach using a custom-built testbed. They designed a full factorial statistical experiment to systematically evaluate the impact of four key factors:
- Marking of ACKs: For Premium (P), Assured (A), and Best-Effort (B) flows.
- Congestion: On the forward and/or reverse paths (C).
They tested all 3^4 = 81 combinations of these factors, with each experiment replicated three times, resulting in 243 total experiments. Analysis of Variance (ANOVA) was used to quantify the contribution of each factor and their interactions to TCP throughput. Visual tests (Q-Q plots and residual plots) were used to validate the statistical models.
Datasets
Not discussed in the paper. The experiments were conducted on a custom-built network testbed.
Key Insights and Ideas
- ACKs are not second-class citizens: The paper strongly emphasizes that the performance of TCP flows is highly sensitive to the congestion and marking of ACK packets on the reverse path. Simply prioritizing data packets is insufficient.
- Congestion is a two-way street: The impact of congestion on the reverse path is often as significant, if not more so, than on the forward path for TCP performance.
- Context matters for marking: The optimal strategy for marking ACK packets is not universal. It heavily depends on the level of congestion on both forward and reverse paths.
- "Copy the mark" is a practical solution: While optimal strategies require knowing current network conditions (which is hard to obtain), a practical "sub-optimal" strategy of having ACKs inherit the mark of the data packet they acknowledge performs remarkably well and is independent of network specifics.
Key Results
- Premium Flows: Throughput is significantly affected by congestion (31.23%) and the marking of Premium ACKs (10.43%). The interaction between Premium ACK marking and congestion is also a major factor (32.26%).
- When the reverse path is congested, marking ACKs as Premium is crucial.
- Marking ACKs as Assured generally leads to better Premium throughput, regardless of congestion.
- Marking ACKs as Premium is detrimental when both Premium and Assured flows use Premium marking for their ACKs.
- Assured Flows: Throughput is overwhelmingly dominated by congestion (86.38%). The interaction between Assured ACK marking and congestion is also important (7.86%).
- Assured flows perform poorly when their forward path is congested.
- When the reverse path is congested, protecting ACKs by marking them as Premium or Assured is beneficial.
- When the forward path is congested, it's better for ACKs to not belong to a rate-limited class.
- Best-Effort marking for ACKs is generally suboptimal for Assured flows, especially when congestion is present.
- Practical Strategy: The "copy the mark" strategy (where an ACK carries the same mark as the data packet it acknowledges) achieved performance comparable to optimal strategies and significantly outperformed the default Best-Effort ACK marking, showing up to a 20% improvement for Assured flows in some scenarios.
Limitations and Future Work
- Predicting Congestion: The authors acknowledge that their optimal strategies rely on knowing the level of congestion on the reverse path, which is difficult for network providers to obtain in real-time. This motivates the search for sub-optimal, congestion-independent strategies.
- Outliers: The statistical analysis identified some outlying data points, suspected to be due to simultaneous SYN packet loss at the start of experiments. While these were handled, they highlight potential transient behavior issues.
- Future Work: The paper suggests that further research is needed on how to offer better-than-Best-Effort services, especially concerning the provisioning of Differentiated Services networks and the investigation of performance for flows requesting preferential treatment. They also note that while ACKs are smaller, their provisioning still requires coordination between network providers handling egress and ingress traffic.
Summary updated