Demos

The programs demonstrated during the lecture, one directory per part of it.

Each demo is small on purpose: it makes one point, and it is meant to be run rather than read. The write-up in each directory says what to expect, what actually happens, and why.

Demo	Part of the lecture	What it shows
00-pitch/	00. Pitch	Ten billion page faults since boot, and a sum of RSS larger than the memory in use.
01-two-views/	01. Two views of one address space	The list of zones and the page table of one process, and which of the two each operation edits.
02-page-fault/	02. The page fault	The four verdicts: minor, major, SEGV_MAPERR, SEGV_ACCERR.
03-frames/	03. Where frames come from	The page cache, and a program squeezed into a 64 MB cgroup until it swaps.
04-memory-sharing/	04. Memory sharing	RSS, PSS and USS for ten processes, and fork() sharing pages until they are written.
05-page-table/	05. Storing the page table	How much page table a process really has, and what a TLB miss costs.

Running them

Every demo that needs building has a Makefile, so make inside its directory is enough. The tools the demos need, beyond a C compiler and make, are:

• ps, pmap and vmstat (procps), to read a process and the system from outside
• smem, for USS and PSS per process
• fincore (util-linux), to see which pages of a file are in the page cache
• systemd-run, to run one program inside a small memory cgroup
• strace, to see the signal a crashing program received
• perf (linux-tools) and sudo, once, for the TLB miss counters

Several demos wait at prompts so that ps can be run on them from a second terminal; they are marked as such in their READMEs. utils/utils.h holds the DIE() macro the sources share.

The numbers quoted in the write-ups were measured on one machine (Ubuntu 25.04, Linux 6.14, glibc 2.41, x86-64, 4 KB pages, 60 GB of RAM, 80 GB of swap). Addresses and exact counts differ from run to run and from machine to machine; the shapes do not.