AI Fabric Load Balancing: ECMP, Letflow, CONGA and ConWeave

AI Fabric Load Balancing: ECMP, Letflow, CONGA and ConWeave

Simulate Load-Balancing Behavior — Not Traditional Router Configuration

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

Instead, this lab uses network simulation to experiment with AI/RDMA fabric load-balancing mechanisms.

You can change traffic load, select different load-balancing schemes, tune parameters, run simulations, and directly observe their effects on path utilization, flow completion time, tail latency, uplink balance, and buffering.

You will compare ECMP, Letflow, CONGA, and ConWeave under the same traffic conditions and see why different approaches to path selection can produce very different results.

Technologies and Focus Areas

In this lab, you will work with the following load-balancing schemes:

• ECMP
• Letflow
• CONGA
• ConWeave

You will analyze how these mechanisms affect:

• Path selection quality
• Path utilization
• Flow completion time
• Tail latency
• ToR uplink balance
• Buffering behavior
• Packet reordering trade-offs

What Will You Learn?

The included workbook guides you through four main experiments that demonstrate how different load-balancing strategies behave under identical network conditions.

1. ECMP vs ConWeave

See how static 5-tuple hashing in ECMP can create path collisions, while congestion-aware path selection can move traffic toward better available paths.

This experiment helps you understand why a simple hash-based mechanism may not always provide efficient balance in AI fabrics.

2. Compare All Four Schemes

Run ECMP, Letflow, CONGA, and ConWeave against identical traffic and compare their behavior.

You will evaluate differences in:

• Large-flow tail latency
• Uplink balance
• Flow distribution across the fabric
• Overall load-balancing effectiveness

3. Change Network Load

Compare behavior at 20% and 80% offered load and observe why better path balancing does not always translate directly into the same improvement in flow completion time.

This experiment highlights how performance gains can depend on traffic intensity, congestion level, and the type of workload.

4. Tune ConWeave

Change ConWeave’s path-switching and timer behavior and observe the trade-offs between:

• Large-flow tail latency
• Small-flow performance
• Path balance
• VOQ buffering

This experiment shows that improving one metric may come at the cost of another, especially when packet steering and queue behavior become more dynamic.

Analyze the Results

The lab produces graphs and measurable metrics so you can see exactly what changes when a different load-balancing decision is made.

You will be able to analyze:

• Average slowdown
• p99 slowdown
• ToR uplink imbalance
• ConWeave VOQ occupancy
• Path utilization efficiency
• Latency trade-offs across different schemes

Learning Outcome

By the end of this lab, you will understand why ECMP hashing alone may not be enough for AI fabrics, how congestion-aware mechanisms can make better path decisions, and why improving path balance may introduce other costs such as packet reordering and buffering.

You will also gain practical insight into how different load-balancing schemes affect not only throughput, but also latency, fairness, and stability across modern AI/RDMA network fabrics.