Load Balancing Playground
Page 4.1 explained the four LB tiers in prose. This page is the hands-on version — pick a method, toggle between web and AI traffic, fail a spine link, and watch what each method actually does on a 4-spine fabric.
How to use it:
- Start with ECMP + Web traffic — flows balance across all four spines. The default everyone assumes. ✔
- Switch to AI training — same ECMP, same fabric. Watch the polarization. Six of eight elephants land on Spine 0. Throughput collapses.
- Try Flowlet DLB + AI traffic — the ASIC notices the load, reassigns flows at flowlet boundaries, all four spines pull load again.
- Try Packet Spraying (TELB) — every packet picks its own path. Maximum spread; needs hardware reorder buffer on the receiver side. This is what AWS SRD / Google Falcon / MRC / UEC do.
- Fail a spine — watch how each method reacts. ECMP needs BGP to rebuild the group (seconds). DLB sees the queue go infinite and reassigns instantly.
Load Balancing Simulator
Pick a method. Pick a traffic pattern. Watch what happens.
RoCE v2 collapses 4 of the 5 hash fields (same src/dst IPs, same UDP dest port 4791). Six of eight elephants land on Spine 0. The fabric runs at ~30% effective throughput while three spines sit idle.
What to take away
| Observation | What it means |
|---|---|
| ECMP works for web traffic and fails for AI traffic | The 5-tuple hash needs entropy. RoCE v2 has almost none. |
| DLB rebalances within ~1 ms | Switch silicon sees its own queue depth. Adaptive routing pays for itself. |
| GLB sees one hop further | Catches downstream congestion DLB can't see locally. |
| TELB / packet spraying is the long-term answer | Per-packet spread + HW reorder. UEC / MRC / SRD / Falcon. |
| Failure recovery is a different problem | BFD / BGP / IP FRR react to dead paths; they don't balance load. |
Once you've played with the simulator, the prose on page 4.1 should read as a deeper explanation of what you just saw.
Next: Cluster Sizing & Cabling → — back to chapter 3. Port counts, optics, cable lengths, and the BOM math that turns the design you've learned into a build plan.