Fabric design for distributed computation

Relate topology, congestion, collective traffic, and failure domains to the communication demands of modern compute clusters.

10 published lessons1 module256 min of guided study
  1. 01
    Fabric design for distributed computationadvanced · 10 lessons

Follow the path or begin with a question. Every lesson is open.

Your notebook0 of 10 lessons completed

Your place, at your pace. Progress is optional and saved only in this browser when available.

Inside this module

Each lesson includes an interactive lab

Fabric design for distributed computation

Relate topology, congestion, collective traffic, and failure domains to the communication demands of modern compute clusters.

  1. 01

    Clos fabrics and a declared bisection

    Count ports and traffic across a precise cut before claiming full fabric bandwidth.

    25 min
  2. 02

    ECMP, flow hashes, and collisions

    Understand stable per-flow path selection without confusing equal-cost routes with equal utilization.

    24 min
  3. 03

    Incast and the budget before feedback

    Estimate synchronized excess arrivals and explain why average utilization can hide short buffer overflows.

    25 min
  4. 04

    ECN marks and end-to-end response

    Trace a congestion signal from queue marking to transport feedback without confusing signaling with extra capacity.

    25 min
  5. 05

    PFC headroom and cyclic waiting

    Separate per-priority link pausing from end-to-end congestion control and reason about resource-dependency cycles.

    26 min
  6. 06

    RDMA completion, ordering, and buffer ownership

    Follow registered memory through submitted work and distinguish local completion from remote application consumption.

    27 min
  7. 07

    RoCEv2, loss recovery, and congestion response

    Separate transport recovery, ECN-based rate control, and optional link protection in a specified deployment.

    26 min
  8. 08

    Collective algorithms meet physical placement

    Relate rank communication, NIC locality, and fabric cuts instead of interpreting one bandwidth number as a topology map.

    27 min
  9. 09

    Failure domains beyond path count

    Evaluate shared dependencies and degraded capacity instead of multiplying independent failure probabilities by habit.

    25 min
  10. 10

    Telemetry boundaries and realistic capacity tests

    Reconcile counters, time windows, and collective completion before claiming fabric utilization or headroom.

    26 min
All engineering modules