Deep Buffers Experiment

Buffer-filling congestion control algorithms (CCAs) are designed to continuously probe for free bandwidth. By doing that, they may keep filling up a standing queue at the bottleneck, which leads to self-inflicted queueing delay. When the bottleneck queue has a large queue size, i.e., when the buffer is deep, large queueing delays may be a consequence. The presence of deep buffers in many networks today and the consequences thereof is known as the problem of bufferbloat. A CCA that is resilient against bufferbloat should refrain from inflicting queueing delays that are proportional to the queue size.

Scenario

In the deep buffers experiment, a single flow operates in a static dumbbell network with a queue size that is larger than the bandwidth-delay product (BDP). The flow generates greedy source traffic and uses a CCA. The experiment has one parameter named qlen that sets the size of the bottleneck queue. It can be repeated for different values of qlen to evaluate the influence of the queue size on the operating point.

To summarize the experiment setup:

  • Topology: Dumbbell topology (\(K=1\)) with static network parameters

  • Flows: A single flow (\(K=1\)) that uses a CCA

  • Traffic Generation Model: Greedy source traffic

Experiment Results

Experiment #1

Parameters

Command: ns3-dev-ccperf-static-dumbbell-default --experiment-name=deep_buffers --db-path=benchmark_TcpNewReno.db '--parameters={aut:TcpNewReno,k:1,qlen:80p}' --aut=TcpNewReno --stop-time=15s --seed=42 --qlen=80p --bw=16Mbps --loss=0.0 --qdisc=FifoQueueDisc --rtts=15ms --sources=src_0 --destinations=dst_0 --protocols=TCP --algs=TcpNewReno --recoveries=TcpPrrRecovery --start-times=0s --stop-times=15s '--traffic-models=Greedy(bytes=0)'

Flows

src dst transport_protocol cca cc_recovery_alg traffic_model start_time stop_time
src_0 dst_0 TCP TcpNewReno TcpPrrRecovery Greedy(bytes=0) 0.00 15.00

Metrics

The following tables list the flow, link, and network metrics of experiment #1. Refer to the the metrics page for definitions of the listed metrics.

Flow Metrics

Flow metrics capture the performance of an individual flow. They are measured at the endpoints of a network path at either the source, the receiver, or both. Bold values indicate which flow achieved the best performance.

Metric flow_1
cov_in_flight_l4 0.21
cov_throughput_l4 0.03
flow_completion_time_l4 15.00
mean_cwnd_l4 79.70
mean_delivery_rate_l4 15.38
mean_est_qdelay_l4 43.84
mean_idt_ewma_l4 0.76
mean_in_flight_l4 79.18
mean_network_power_l4 273.48
mean_one_way_delay_l7 1925.14
mean_recovery_time_l4 101.88
mean_sending_rate_l4 15.53
mean_sending_rate_l7 17.53
mean_srtt_l4 58.84
mean_throughput_l4 15.39
mean_throughput_l7 15.39
mean_utility_mpdf_l4 -0.07
mean_utility_pf_l4 2.73
mean_utilization_bdp_l4 4.12
mean_utilization_bw_l4 0.96
total_retransmissions_l4 108.00
total_rtos_l4 0.00

Figures

The following figures show the results of the experiment #1.

Time Series Plot of the Operating Point

Time series plot of the number of segments in flight, the smoothed round-trip time (sRTT), and the throughput at the transport layer.

In Flight vs Mean Operating Point

The mean throughput and mean smoothed round-trip time (sRTT) at the transport layer of each flow. The optimal operating point is highlighted with a star (magenta). The joint operating point is given by the aggregated throughput and the mean sRTT over all flows

Distribution of the Operating Point

The empirical cumulative distribution function (eCDF) of the throughput and smoothed round-trip time (sRTT) at the transport layer of each flow.

Comparison of Congestion Control Algorithms (CCAs)

Figures