Gentle Slow Start to Alleviate TCP Incast in Data Center Networks

Sheeba Memon, Jiawei Huang, Hussain SaajidSymmetry 2019 · DOI

TL;DR

TCP Incast, a performance bottleneck in data centers, is exacerbated by the aggressive "slow start" phase of TCP. This paper introduces "Gentle Slow Start" (GSS), an algorithm that moderates the congestion window's growth during slow start, significantly reducing Incast probability and improving network goodput by up to 8x.

What Problem Does This Paper Address?

Modern data centers rely heavily on TCP for communication. However, when a large number of TCP flows simultaneously send data to the same receiver (a scenario known as TCP Incast), it can lead to congestion collapse. This happens because switches have limited buffer space, and the rapid increase in data transmission during TCP's "slow start" phase quickly overwhelms these buffers, causing packet loss and timeouts. Existing solutions primarily focus on the "congestion avoidance" phase, neglecting the critical slow start period where many short-lived "mice" flows reside. This oversight means that even with optimized congestion avoidance, TCP Incast remains a significant problem, especially in data centers where short flows are prevalent.

Methods

The core of this paper is the Gentle Slow Start (GSS) algorithm. GSS aims to mitigate TCP Incast by modifying the slow start phase. Instead of the standard exponential increase of the congestion window (cwnd) by one segment per acknowledgment (ACK), GSS dynamically adjusts the cwnd growth rate based on real-time network congestion.

Here's how it works:

  • Congestion State Estimation: GSS uses the round-trip time (RTT) to estimate the congestion state. It calculates a congestion factor 'α' using the current RTT, maximum RTT, and minimum RTT.
  • α = (RTT - RTT_min) / (RTT_max - RTT_min)
  • Gentle Congestion Window Adjustment: The congestion window for the next round trip (cwnd_{i+1}) is updated as:
  • cwnd\_{i+1} = cwnd\_{i} + cwnd\_{i} * (1 - α)
    • When 'α' is close to 0 (low congestion), the growth is faster, approaching exponential.
  • When 'α' is close to 1 (high congestion), the growth becomes more additive, similar to congestion avoidance, thus "gentler."
  • Integration: GSS is designed to be a simple modification, requiring only a few lines of code to be integrated into existing TCP protocols. The authors demonstrate its integration with state-of-the-art data center transport protocols like DCTCP, D2TCP, and L2DCT, as well as the widely used TCP NewReno.

Datasets

Not discussed in the paper. The evaluation relies on simulations using Network Simulator 2 (NS2).

Key Insights and Ideas

  • Slow Start is the Culprit: The paper convincingly argues that the aggressive nature of TCP's slow start phase is a primary driver of TCP Incast, especially in data centers dominated by short flows. This challenges the common focus on congestion avoidance.
  • Real-time Congestion Awareness is Key: By using RTT as a proxy for congestion, GSS can adapt its window growth dynamically, preventing the rapid buffer exhaustion that leads to Incast.
  • Simplicity Enables Wide Adoption: The algorithm's minimal changes to existing TCP stacks make it highly practical for deployment in real-world data center environments.
  • Broad Applicability: GSS isn't tied to a specific protocol; it enhances the performance of various TCP variants, demonstrating its versatility.

Key Results

The paper presents compelling results from NS2 simulations across different scenarios:

  • Reduced Incast Probability: GSS significantly decreases the probability of TCP Incast, especially under high flow concurrency. Figure 3 shows that GSS allows for a much larger number of concurrent flows before Incast becomes probable, even on lower bandwidth links.
  • Improved Network Goodput:
    • In a basic performance test (Figure 4), GSS eliminated Time-Out events and reduced transfer times by approximately 0.01s compared to standard TCP.
  • For multiple rounds of requests (Figure 5), TCP NewReno saw a 7x reduction in transfer time with GSS, and DCTCP also showed significant speedups.
  • Figure 6 demonstrates that with GSS, DCTCP and TCP NewReno can support many more concurrent flows while maintaining high goodput, achieving up to an 8x increase in network goodput in some cases.
  • Enhanced Web Search and MapReduce Performance:
    • In web search applications (Figure 7), GSS eliminated Time-Out ratios for both DCTCP and NewReno, leading to lower search response times.
  • In MapReduce simulations (Figure 8), GSS improved goodput, especially under background workload conditions, outperforming original DCTCP and NewReno.
  • Protocol Agnostic Improvement: Figure 9 shows that GSS enhances the performance of DCTCP, D2TCP, and L2DCT, achieving a 35x incremental increase in network goodput for 100 workers when combined with these protocols.

Limitations and Future Work

  • RTT Measurement Accuracy: The authors acknowledge that accurately measuring RTT in data centers can be challenging due to host delays and virtualization. However, they point to recent advancements in RTT measurement techniques that could address this.
  • Initial Congestion Window: The paper notes that many protocols initialize the congestion window aggressively (10-50 segments). GSS sets the initial cwnd to 2, which is a deliberate choice to be less aggressive.
  • Future Work: The authors plan to:
  • Deploy GSS in a real Linux-based TCP stack testbed for evaluation.
  • Further optimize the congestion window adjustment mechanism within the GSS algorithm.

Summary updated

BibTeX
@article{memon2019gentle,
  author = {Sheeba Memon and Jiawei Huang and Hussain Saajid},
  title = {{Gentle Slow Start to Alleviate TCP Incast in Data Center Networks}},
  journal = {Symmetry},
  year = {2019},
  doi = {10.3390/sym11020138},
}
Figure