Comparing AI Fabric Congestion Control: DCQCN, TIMELY and HPCC

Comparing AI Fabric Congestion Control: DCQCN, TIMELY and HPCC

Simulate Protocol Behavior — Not Traditional Router Configuration

This is not a traditional Cisco or Juniper lab where you connect to routers and switches and configure protocols through the CLI.

Instead, this lab uses a network simulator to compare the three congestion control algorithms most often discussed for RDMA and RoCEv2 AI training fabrics: DCQCN, TIMELY and HPCC. You change parameters, generate different traffic patterns, run the simulation and compare how each algorithm reacts to the same congestion event.

Lab Objective

The goal is to understand why AI fabrics need end-to-end congestion control on top of PFC, and how the signal each algorithm uses changes its behavior. DCQCN reacts to ECN marks from the switches, TIMELY reacts to round-trip time measured at the hosts, and HPCC uses in-band network telemetry (INT) to read the exact queue length and link utilization on every hop. By the end of the lab you will be able to explain which algorithm converges faster, which keeps queues shorter and which is easier to deploy on an existing fabric.

Topology

The simulated topology is a two-tier leaf-spine fabric with GPU servers attached to the leaf switches over RoCEv2 links. Several senders transmit to a single receiver to create an incast hotspot, and a separate victim flow shares part of the path so you can see how congestion spreads beyond the hotspot.

• Leaf-spine fabric with configurable link speed
• Many-to-one incast traffic toward one receiver
• A victim flow that crosses the congested leaf
• PFC available as the lossless safety net

What Can You Configure

• Select DCQCN, TIMELY or HPCC as the congestion control algorithm
• Tune ECN Kmin and Kmax thresholds for DCQCN
• Set the low and high RTT thresholds for TIMELY
• Change the target link utilization (eta) for HPCC
• Modify the incast ratio, link speed and traffic duration

Lab Tasks

1. Baseline With PFC Only

Run the incast without end-to-end congestion control and observe PFC PAUSE frames, head-of-line blocking and the throughput drop of the victim flow.

2. DCQCN

Enable DCQCN and measure how quickly ECN feedback reduces sender rates, how deep the queue grows before marking starts and how often PFC is still triggered.

3. TIMELY

Switch to TIMELY and compare its RTT-gradient based reaction. Note how sensitive it is to RTT thresholds and measurement noise.

4. HPCC

Switch to HPCC and observe how precise INT feedback lets senders converge to the available bandwidth in very few RTTs while keeping queues near zero.

5. Side-by-Side Comparison

Repeat the same scenario with different incast ratios and link speeds and record queue depth, flow completion time, fairness and PFC pause counts for each algorithm.

Analyze the Results

Each run produces graphs and metrics so you can compare the algorithms on the same traffic. You will see that DCQCN is the most widely deployed because it only needs ECN on the switches, TIMELY avoids switch support but depends on accurate timestamps, and HPCC gives the tightest control at the cost of INT-capable hardware.

Outcome

After completing this lab you will be able to choose and tune a congestion control algorithm for an AI back-end network, explain the trade-offs between ECN, delay and telemetry based feedback, and predict how each one interacts with PFC during incast. This lab builds on the PFC, ECN and DCQCN lab and prepares you for the load balancing and multipathing labs in this series.