Design and implementation of a feedback controller for slowdown differentiation on internet servers

Jianbin Wei, Chengzhong Xu2005 · DOI

TL;DR

This paper introduces a feedback controller for Internet servers that ensures different client classes experience proportional slowdown ratios, meaning their waiting time is proportional to their service time. By dynamically adjusting processing rates based on achieved slowdowns, the controller effectively manages Quality of Service (QoS) and improves system stability, as demonstrated in an Apache web server implementation.

What Problem Does This Paper Address?

In today's internet, users expect different levels of service. Imagine a high-priority customer versus a casual browser – they shouldn't necessarily experience the same waiting times. The paper tackles the challenge of providing differentiated Quality of Service (QoS) on internet servers. Specifically, it focuses on a metric called "slowdown," which is the ratio of a request's waiting time to its service time. The goal is to maintain these slowdown ratios proportionally across different client classes, as defined by pre-set differentiation parameters. This means if class A is supposed to have twice the slowdown of class B, the system should strive to enforce that ratio. Existing systems often focus on minimizing average slowdown, but this paper aims for precise, proportional differentiation.

Methods

The core of the paper is a feedback controller designed to dynamically allocate processing rates to different client classes. Here's how it works:

  • Proportional Slowdown Differentiation (PSD): The system aims to satisfy the condition $S_i(k) / \delta_i = S_j(k) / \delta_j$ for any two classes $i$ and $j$ at sampling period $k$, where $S(k)$ is the average slowdown and $\delta$ is the differentiation parameter.
  • Integral Feedback Control: The controller calculates an "error" based on the difference between the target slowdown ratio and the achieved one. This error is then used by an integral controller to adjust the processing rate for each class. Integral control is chosen to eliminate steady-state errors and make the system robust to noise.
  • Rate Prediction: Before the feedback loop kicks in, an initial processing rate for each class is estimated using queueing theory. This prediction considers the arrival rate and mean service time of requests, ensuring a baseline allocation that prevents overload.
  • Implementation on Apache: The feedback controller was implemented within an Apache web server. Instead of directly controlling CPU allocation, the system manipulates the number of child processes assigned to handle requests from different classes.

Datasets

Not discussed in the paper. The experiments were conducted using a custom setup with an Apache web server and the SURGE traffic generator.

Key Insights and Ideas

  • Dynamic Rate Allocation is Crucial for Proportional Differentiation: Simply setting static rates isn't enough. The system needs to constantly monitor and adjust processing power to maintain the desired slowdown ratios, especially under varying workloads.
  • Integral Control for Robustness: Using integral control for the feedback mechanism is key to achieving stable performance and eliminating persistent errors in slowdown ratios.
  • Queueing Theory for Initial Allocation: Leveraging queueing theory to predict initial processing rates provides a solid starting point, preventing immediate performance issues and making the feedback loop more effective.
  • Process Count as a Proxy for Rate: The clever implementation of controlling the number of Apache child processes effectively translates the abstract concept of "processing rate" into a concrete system mechanism.

Key Results

The experimental results demonstrate the effectiveness of the feedback controller:

  • Target Ratios Achieved: The controller successfully achieved the target slowdown ratios (e.g., 2, 4, and 8) across different workload conditions.
  • Reduced Variance: The feedback controller significantly reduced the variance in achieved slowdown ratios compared to a system without feedback. For instance, with a system load of 50% and a target ratio of 4, the difference between the 95th and 5th percentiles of slowdown ratios dropped from 3.3 to 0.43.
  • Robustness: The controller showed effectiveness and robustness under various environments and workloads.

Limitations and Future Work

  • Control Factor Tuning: The paper mentions that the control factor 'g' (which influences how quickly the controller responds to errors) was tuned using a simulator and set to 1. The process of tuning this factor and its impact on oscillations are discussed in a separate reference [4].
  • Future Work: While not explicitly detailed in this excerpt, the success of this implementation suggests future work could involve scaling this approach to more complex server architectures, exploring different QoS metrics, or integrating it with other network management techniques.

Summary updated

Abstract

Proportional slowdown differentiation (PSD) aims to maintain slowdown ratios between different classes of clients according to their pre-specified differentiation parameters. In this paper, we design a feedback controller to allocate processing rate on Internet servers for PSD. In this approach, the processing rate of a class is adjusted by an integral feedback controller according to the difference between the target slowdown ratio and the achieved one. The initial rate class is estimated based on predicted workload using queueing theory. We implement the feedback controller in an Apache Web server. The experimental results under various environments demonstrate the controller's effectiveness and robustness.
BibTeX
@misc{wei2005design,
  author = {Jianbin Wei and Chengzhong Xu},
  title = {{Design and implementation of a feedback controller for slowdown differentiation on internet servers}},
  year = {2005},
  doi = {10.1145/1062745.1062841},
}
Figure