Performance and Robustness Testing of Explicit-Rate ABR Flow Control Schemes Milan Zoranovic Carey Williamson October 26, 1999

MASCOTS Agenda u Introduction and Motivation u Background Information u Explicit-Rate ABR Traffic Control Schemes (ERICA, ERICA+, DEBRA) u Experimental Methodology u Simulation Results: Performance Testing u Simulation Results: Robustness Testing u Summary and Conclusions

MASCOTS Introduction u Problem Definition and Motivation: u Explicit-Rate (ER) ABR flow control schemes u Many (ER) ABR flow control schemes have been proposed u Performance evaluations are author and scheme dependent u Difficult to do direct comparison u Study Objectives: u Propose set of benchmark network configurations u Evaluate and compare ERICA, ERICA+, and DEBRA strategies on this set of benchmark configurations u Use Asynchronous Transfer Mode -Traffic and Network (ATM-TN) simulator for this purpose

MASCOTS Background u ABR Flow Control Mechanism u There are five classes of service (CBR, VBR (2), UBR, and ABR) u ABR and UBR use the remaining bandwidth u ABR bandwidth varies between minimum bandwidth and the extra bandwidth freed by the VBR traffic sources u ABR flow control schemes are in charge of managing this bandwidth effectively u Resource Management (RM) Cells u Used as mechanism for ABR flow control u RM-cell contains information about the state of the network (CI, ER, CCR, MCR. DIR,…) u The mechanism is called closed-loop u Behavior of ABR flow control:

MASCOTS Background Continued... Data FRM BRM Source Destin. Switch

MASCOTS Explicit-Rate ABR Traffic Control Schemes u The ERICA Algorithm u ERICA (Explicit Rate Indication for Congestion Avoidance) is proposed by Ray Jain et al. u ERICA tries to achieve a fair and efficient allocation of the available bandwidth to competing sources u Each switch monitors the incoming cell rates of each ABR traffic source, the available capacity, and the number of active sources u Aggregate ABR demand vs target load u The ERICA+ algorithm u It uses a target queuing delay rather than a target utilization, and refined parameters for source rate adjustment for faster convergence u The target queuing delay (D), determines the steady state buffer occupancy at the bottleneck link u ERICA+ achieves higher network utilization then ERICA, while only slightly increasing the end-to-end delay

MASCOTS Explicit-Rate ABR Traffic Control Schemes Continued... u The Dynamic Explicit Bid Rate Algorithm (DEBRA) u Based on a rate-based flow control strategy called loss-load curves u Switches compute and provide to traffic sources concise aggregate load information u Sources compute precise transmission rates that provide the best trade off between offered load and the level of packet loss in the network u  = r * (1-p) u  - allocated bandwidth to a current VC u r - requested bandwidth by a current VC u p - loss probability assigned to a current VC u f - a fraction of total capacity requested by current VC u K- controls aggressiveness, responsiveness and convergence

MASCOTS Experimental Methodology u ATM-TN Simulator u Provides cell-level simulation of the ATM-TN traffic flows from traffic sources to traffic sinks u ABR persistent sources u Per-port output-buffered switch model u ERICA, ERICA+ and DEBRA are implemented in the simulator u A set of nine network configurations for performance evaluation u A set of four network configuration for robustness tests

MASCOTS Experimental Methodology Continued... u Performance Metrics u Allowed Cell Rate (ACR): Mbps u Link Utilisation: Percentage u Queue Length: Number of Cells u Throughput: Number of Cells u Cell Loss Ratio (CLR): Percentage u Experimental Design u Performance Testing: each of the algorithms is evaluated on set of nine benchmark scenarios u Robustness Testing: each of the algorithms is evaluated on a set of four benchmark scenarios for testing the robustness

MASCOTS Performance Testing Set of Benchmark Scenarios

MASCOTS Performance Testing Continued... u Simulation results for all the three schemes are shown on One-at-a-Time and Generic Fairness Configuration 1 network scenarios (ACR and Link Utilisation) u One-at-a -Time Network Configuration u LAN network configuration with 30 sources u Start up one at a time, every 10 ms u Test responsiveness, fairness, efficiency, and scalability

MASCOTS Performance Testing Continued … One-at-a-Time: ACR and Link Utilisation ERICAERICA+DEBRA

MASCOTS Performance Testing Continued… Generic Fairness Configuration 1 (GFC1) u Five Switch “Parking-Lot” WAN Network Topology u Used by ATM Forum u There are 23 traffic sources u Purpose: testing for max-min fairness among the sources with different bottlenecks, rates and RTT

MASCOTS Performance Testing Continued… GFC1: ACR and Link Utilisation ERICAERICA+DEBRA

MASCOTS Performance Testing Continued… u Summary of Performance Testing Results u All three algorithms performed well on One-at-a-Time scenario u DEBRA needs more time to converge to a steady-state than ERICA+ on GFC1, but less than ERICA (link utilization) u ERICA+ performs better than its predecessor ERICA u ERICA and ERICA+ did not perform as well as DEBRA during the steady-state on GFC1 (more oscillations for higher rate sources in both ACR and Link Utilization) u ERICA and ERICA+ showed to be very sensitive to parameters configuration (  and D)

MASCOTS Robustness Testing Set of Benchmark Scenarios u Network scenarios with non-cooperative traffic sources u Intentional overuse of underuse of their fair-share u Dishonest and honest traffic sources u Based on Two Sources network scenario

MASCOTS Robustness Testing Continued… Dishonest Sources Scenario: ACR and Throughput ERICA ERICA+DEBRA

MASCOTS Robustness Testing Continued… Honest Sources-One High Scenario: ACR/Throughput ERICAERICA+DEBRA

MASCOTS Robustness Testing Continued… u Summary of Robustness Testing Results u None of the schemes performs properly when sources are greedy and dishonest u ERICA+ is able to avoid congestion on all the scenarios, but do not achieve fairness u ERICA is not very robust - experience both, unfairness and congestion (CLR) when sources are greedy u DEBRA the only one to perform properly on the scenarios with honest and greedy ABR traffic sources

MASCOTS Conclusions and Future Work... u Conclusions u Set of benchmark network configuration is needed for good comparison u Simulation results show: none of the schemes is perfect u ERICA+ performed better than its predecessor ERICA u DEBRA, a new ER ABR flow control scheme is very competitive u Performed as well as ERICA+ on basic set of network configuration u Performed better than ERICA+ on the robustness tests u Future Work u Study ABR performance with more realistic traffic (bursty traffic sources, self-similar traffic, finite traffic sources) u Interaction between TCP and ATM ABR u Improving the DEBRA algorithm (avoiding the buffer overflow problem at the source start-up time) by adding gradual ramp-up feature (THIS ONE WILL BE REMOVED)