Skip to content
AI Data Center Systems

Lab: Topology-Aware Bandwidth

Goal

Section titled “Goal”

Show why NCCL path quality depends on placement. A job can use GPUs successfully and still run over a slow PCIe or cross-socket path.

Baseline

Section titled “Baseline”

The baseline assigns ranks in a topology-agnostic order. Several collective edges cross slow links.

Optimized

Section titled “Optimized”

The optimized path groups ranks by fast local links first, then crosses the slower boundary with reduced traffic.

Run

Section titled “Run”
Terminal window
python compare.py

Expected Observation

Section titled “Expected Observation”

Both paths move the same logical gradient bytes. The topology-aware path should expose less communication time because more traffic stays on fast links.