ARC-201 · Processor Architecture · Practitioner
NUMA & Multi-Socket Systems — full syllabus
Non-uniform memory access, node topology discovery, and the placement decisions that quietly cost you throughput.
Who this course is for
Platform, database and HPC engineers running multi-socket servers whose placement decisions quietly cost them throughput.
Prerequisites
- Linux administration basics
- ARC-102 or equivalent cache and memory knowledge
- C helpful for the libnuma exercises
Course outline
Day 1 — Topology and measurement
- NUMA node topology, distance matrices and discovery with numactl -H, lscpu and sysfs
- Local versus remote memory latency and bandwidth in practice, measured not assumed
- First-touch allocation policy and its long-lived consequences
- Reading /proc/<pid>/numa_maps to see where pages actually are
Day 2 — Placement policy and diagnosis
- numactl and libnuma in practice: binding, interleaving and preferred nodes
- The effect of placement on real workloads: databases, JVMs, HPC jobs
- Diagnosing cross-node traffic with hardware counters: offcore and uncore events, perf c2c
- When to stop fighting and interleave
- A repeatable topology-first routine for unfamiliar servers
Hands-on labs
- Lab: map a machine's NUMA topology with numactl -H, lstopo and /sys/devices/system/node and draw the distance matrix
- Lab: measure local versus remote latency and bandwidth with a first-touch-controlled benchmark under numactl --membind/--cpunodebind
- Lab: trace a process's page placement through /proc/<pid>/numa_maps and migrate pages with migratepages to prove the effect
- Lab: detect cross-node sharing in a two-thread workload with perf c2c and quantify the remote traffic
Capstone project
Given a throughput problem on a multi-socket machine, find the placement fault: characterise the workload's memory and CPU binding, produce counter evidence for the cross-node traffic, choose between binding, interleaving and first-touch fixes, and deliver before/after distributions showing the change worked — or an honest report of why it did not.
What you leave with
- A topology-discovery routine applicable to any unfamiliar server
- Working command of numactl, libnuma and numa_maps
- Measured local-versus-remote numbers for a real platform
- A diagnostic path for cross-node traffic using perf c2c and uncore counters
Upcoming dates
| Dates | Where | Seats | Early bird | Regular | |
|---|---|---|---|---|---|
| 8 Nov – 9 Nov 20262 full days | RiyadhIn person · KAFD Conference Centre | 11 of 14 | SAR 4,720until 9 Oct | ||
| 15 Nov – 16 Nov 20262 full days | Kuwait CityIn person · Al Hamra Tower | 6 of 14 | KWD 390until 16 Oct | ||
| 22 Nov – 23 Nov 20262 full days | MuscatIn person · Knowledge Oasis Muscat | 11 of 14 | OMR 490until 23 Oct | ||
| 22 Nov – 25 Nov 20264 half-days | Gulf bandLive online · 09:00–13:00 GMT+3 | 11 of 20 | US$900until 23 Oct | ||
| 23 Nov – 24 Nov 20262 full days | OttawaIn person · Kanata North Tech Park | 6 of 14 | CAD 1,710until 24 Oct | ||
| 30 Nov – 1 Dec 20262 full days | TorontoIn person · MaRS Discovery District | 11 of 14 | CAD 1,710until 31 Oct | ||
| 30 Nov – 3 Dec 20264 half-days | Europe bandLive online · 09:00–13:00 CET | 16 of 20 | US$900until 31 Oct | ||
| 30 Nov – 3 Dec 20264 half-days | Americas bandLive online · 13:00–17:00 ET | 5 of 20 | US$900until 31 Oct | ||
| 7 Dec – 8 Dec 20262 full days | LondonIn person · Shoreditch Works | 6 of 14 | GBP 980until 7 Nov | ||
| 7 Dec – 8 Dec 20262 full days | BerlinIn person · Factory Görlitzer Park | 11 of 14 | EUR 1,160until 7 Nov |
Book a seat, or bring this course to your team
Seats can be reserved online; private delivery runs on-site or live online, adapted to your stack.
Questions about fit or prerequisites? Email hello@kernelsystems.academy. To save this syllabus, print this page to PDF from your browser.