RED PERFORMANCE EVALUATION USING STOCHASTIC MODELLING AND FLUIDBASED

  • Slides: 14
Download presentation
RED PERFORMANCE EVALUATION USING STOCHASTIC MODELLING AND FLUID-BASED ANALYSIS APPROACHES Hussein Al-Zubaidy Tariq Omari

RED PERFORMANCE EVALUATION USING STOCHASTIC MODELLING AND FLUID-BASED ANALYSIS APPROACHES Hussein Al-Zubaidy Tariq Omari System and Computer Engineering Carleton University {hussein, tomari} @sce. carleton. ca May 10, 2006 CCECE 06

Random Early Detection (RED) Queue Management Scheme RED is used in modern computer networks

Random Early Detection (RED) Queue Management Scheme RED is used in modern computer networks to alleviate some of the problems that the old tail drop suffers from. 2

Objective l l Outline a comprehensive RED performance analysis to provide a better understanding

Objective l l Outline a comprehensive RED performance analysis to provide a better understanding of the RED algorithm. Using stochastic modeling and Fluid based analysis to answer the following questions: 1. What performance enhancements does RED have over Tail Drop. 2. How well RED performs with bursty arrival traffic (e. g. , TCP) 3. What is the effect of adding UDP traffic on RED performance. l 3 Verify the accuracy of these models using simulation.

Contributions l Modeling RED router with a mixed TCP/UDP traffic using two analysis techniques:

Contributions l Modeling RED router with a mixed TCP/UDP traffic using two analysis techniques: 1 - Fluid-based analysis and, 2 - Stochastic modeling. l l 4 Study RED performance in a TCP based network with added UDP traffic. Study the effect of the added UDP on system stability.

RED Analysis: Stochastic-Based Approach l l l 5 The router is modeled by a

RED Analysis: Stochastic-Based Approach l l l 5 The router is modeled by a simple queue with a single input stream of bursty traffic. The packets arrival process is modeled as a batch Poisson process with random burst size (B). For smooth traffic, the same model used with burst size equal 1.

Drop Probability for RED and TD The queue occupancy defines a Markov chain that

Drop Probability for RED and TD The queue occupancy defines a Markov chain that has a stationary distribution (π). The drop probabilities for both Tail Drop and RED routers are: 6

RED Analysis: Fluid-Based Analysis A router with M TCP input flows and added UDP

RED Analysis: Fluid-Based Analysis A router with M TCP input flows and added UDP flows with aggregate rate λUDP is modeled by a system of M+2 differential equations that can be solved numerically: … (1) … (2) … (3) 7

Simulation setup s 0 s 1 10 Mbps 20 msec Bottleneck RED Router d

Simulation setup s 0 s 1 10 Mbps 20 msec Bottleneck RED Router d 0 d 1 4 Mbps 60 msec sn 8 s 0 ↔ d 0 : UDP si ↔ di : TCP; for i =1, … n) and n = 40 dn

Results and Discussion Normalized TCP throughput vs λUDP /C. RED doesn’t starve TCP even

Results and Discussion Normalized TCP throughput vs λUDP /C. RED doesn’t starve TCP even with very high UDP rates. 9

Drop probability for different UDP/TCP rates. l l l 10 RED queue can handle

Drop probability for different UDP/TCP rates. l l l 10 RED queue can handle small UDP rates compared to C. Higher UDP rates will induce instability and cause the drop probability to vary wildly and increase the drops. This results in an unfairness to TCP flows and performance deterioration.

Queue size for RED and TD routers 11

Queue size for RED and TD routers 11

Effect of Adding UDP Traffic 12

Effect of Adding UDP Traffic 12

Conclusions l l l 13 l Two analytic approaches were used to quantify the

Conclusions l l l 13 l Two analytic approaches were used to quantify the benefits claimed that RED has on Tail Drop (TD). The analysis shows that a well tuned RED outperforms TD and fixes some of its deficiencies. RED reduces the bias against bursty TCP traffic by increasing the drop of UDP traffic. RED decreases queuing delay but increases jitter. RED queue can handle small UDP rates compared to the link capacity in addition to the TCP flows. For higher UDP rates, RED became unstable and the drop probability starts oscillating wildly. However, this will not starve the TCP flows. Simulation confirms the above conclusions and matches well with the findings obtained from analysis.

Thank you Q&A 14

Thank you Q&A 14