kernel

PACT kernel setup (Linux 6.3)

This folder builds the kernel PACT runs on and the two out-of-tree modules the tiering subsystem needs to work:

• setup_kernel.sh - builds a vanilla Linux 6.3 kernel.
• tierinit/, kswapdrst/ - required kernel modules (details below).

Important: PACT runs on a vanilla (unpatched) Linux 6.3, but the kernel's NUMA-tiering subsystem will not behave correctly out of the box for our two-node (DRAM + CXL-like) setup. The two modules below configure it. Load both before running PACT - run-pact.sh refuses to start if either is missing.

Required kernel modules

tierinit - register the slow tier and demotion targets

On a stock 6.3 kernel, the far-memory NUMA node (node 1) is treated as just another DRAM node, so the memory-tiering subsystem builds no demotion path between the fast tier (node 0) and the slow tier. tierinit fixes this at load time: it assigns the far node a higher abstract distance (FARMEM_ADISTANCE), registers it as a lower memory tier, and re-establishes the demotion targets - without any kernel patch, using kprobes to resolve the internal memory-tiers symbols (hence the CONFIG_KPROBES / CONFIG_KALLSYMS_ALL requirement below). Without tierinit, PACT has no tier structure to demote/promote pages across.

kswapdrst - keep demotion alive under sustained pressure

The kernel's reclaim/demotion daemon (kswapd) increments a per-node kswapd_failures counter and permanently backs off once it gives up, which stalls demotion during long tiering runs. kswapdrst periodically resets kswapd_failures on the fast-tier node (every 5 s) so demotion keeps making progress. (Derived from the Colloid memory-tiering artifact.) Without kswapdrst, demotion can silently stop partway through a workload.

Requirements

These need a kernel built with the following options (the provided setup_kernel.sh enables them automatically):

• CONFIG_INTEL_UNCORE_FREQ_CONTROL=y - uncore-frequency control (used by ../env/modify-uncore-freq.sh).
• CONFIG_KPROBES=y, CONFIG_KALLSYMS_ALL=y

1. Build and install the 6.3 kernel

setup_kernel.sh does everything - clone vanilla v6.3, configure, build, modules_install, install, and update-grub - and then arms a one-shot boot of the new kernel via grub-reboot, keeping your current kernel as the persistent default. So if 6.3 fails to boot (e.g. a missing built-in driver), the next reboot automatically falls back to the known-good kernel.

./setup_kernel.sh        # build + install + arm one-shot boot of 6.3
sudo reboot              # boot into 6.3 once; confirm with: uname -r

Do not run update-grub / grub-set-default yourself here - that would overwrite the one-shot fallback the script set up. Only make 6.3 the persistent default (grub-set-default ...) once you have confirmed it boots.

2. After rebooting into 6.3, build the modules

Each module has its own Makefile. Build them either way:

./build_modules.sh

or build each separately (from setup/kernel/):

make -C tierinit          # -> tierinit/tierinit.ko
make -C kswapdrst         # -> kswapdrst/kswapdrst.ko

3. Load the modules

run-pact.sh loads them automatically: if a module is already loaded it is kept, if its .ko is built it is insmoded, and if it is missing the run aborts and asks you to build it (step 2). To load them manually (from setup/kernel/):

sudo insmod tierinit/tierinit.ko
sudo insmod kswapdrst/kswapdrst.ko

Verify and inspect logs:

lsmod | grep -E 'tierinit|kswapdrst'
sudo dmesg | grep -E 'tierinit|kswapdrst'

Unload (reverse order):

sudo rmmod kswapdrst tierinit