eoc-meter: RAPL reads each package's energy-status MSR where powercap lists none #627

Merged

dcharlot merged 2 commits from eoc-meter/rapl-msr into main 2026-11-05 04:53:00 -05:00

Conversation 0 Commits 2 Files changed 2 +825 -88

2026-11-04

Transaction Science

eoc-meter: RAPL MSR domains are per die on Intel, per package on AMD; powercap die zones count ...

The kernel scopes Intel's RAPL energy counters per die. Its perf PMU treats
the package counter as package-scope only on AMD and Hygon ("AMD/HYGON
platforms have a per-PKG package energy counter", rapl_pkg_pmu_is_pkg_scope()
in arch/x86/events/rapl.c), and on a multi-die Intel package (Cascade Lake-AP)
"its RAPL MSRs are die-scope"; its powercap driver then lists one zone per die,
named package-<p>-die-<d> (intel_rapl_common.c). Both paths here under-counted
such a package, the unsafe direction at a ceiling:

- The MSR fallback grouped CPUs by physical_package_id alone, so it read one
  die of a multi-die Intel package. On Intel a counter domain is now
  (physical_package_id, die_id), die_id 0 where the file is absent; on AMD
  and Hygon it stays the package. Each domain is read on its lowest-numbered
  CPU and every domain is summed, all or nothing as before (zero unit, wrap,
  powercap preferred, status text unchanged). A die of a multi-die package
  has its own rail name, package-<p>-die-<d> (MSR), in the liveness rails and
  in errors.
- The powercap name rule accepted only package-<n>, so such a host's die
  zones were no package at all. It now also accepts package-<p>-die-<d>, and
  nothing looser (not psys, dram, package-, package-0-die-,
  package-0-core-0), so every die zone is summed and watched.

The hardware test's oracle follows both rules: its zone check accepts die
zones, and its MSR reader groups Intel CPUs by package and die.

Tests (fake files): an Intel package of two dies is two counters, each read
on its own first CPU, both summed, with distinct rails, and a die whose device
is missing leaves no figure; an AMD package whose CPUs report different
die_ids is one counter; an Intel host with no die_id files, or all on die 0,
is read per package as before; the name rule over die names and near misses;
a fake two-die powercap tree sums both zones and is frozen when one die holds.
Unit tests pass on Linux x86_64 (default and cuda) and on macOS (the CI step,
its `ioreport liveness:` line printed); clippy is clean on macOS and on Linux
in the default build.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

335f7f1fa7

• View at this point in history

2026-11-04 21:20:05 -05:00

Transaction Science

eoc-meter: RAPL reads each package's energy-status MSR where powercap lists none ...

Some kernels ship without the powercap RAPL driver (Ubuntu without
linux-modules-extra, as on some rented bare-metal servers), so
/sys/class/powercap lists no package while the counters are there, and the
RAPL rung was unavailable on hosts that can measure. Where powercap lists no
package, the meter now reads the same counters directly: each package's
MSR_PKG_ENERGY_STATUS (0x611, Intel family 6; 0xC001029B, AMD Zen and Hygon)
through /dev/cpu/<first cpu of the package>/msr, the 32-bit count converted at
the package's own unit (bits 12:8 of the power-unit MSR), so it wraps at 2^32
counts in microjoules. A power-unit MSR that reads zero is no counter (KVM
guests read zeros). Powercap is still preferred, and the liveness rule is
unchanged: every package readable and live, or no figure. The probe records
where its packages are read (the powercap directory, or the filesystem root
for MSR packages), and the counter reads them there. CounterStatus and
rapl_status say why an MSR host without CAP_SYS_RAWIO is Unavailable. Ported
from jouleclaw-energy's RAPL MSR fallback, adapted to eoc's package and probe
types; no new dependency.

The hardware test's oracle reads the MSRs itself where powercap lists no
package (vendor and family, each package's first CPU, the registers at their
offsets), so its verdict and its package bracket cover such a host.

On two bare-metal servers with no powercap zone, as root: a Xeon E-2278G
live, 8.561 J over a 300 ms one-thread burn (28.5 W), inside the oracle's MSR
bracket; two Xeon Silver 4215R sockets, both live, 23.071 J (76.9 W), inside
it. As an unprivileged user (no CAP_SYS_RAWIO) the status is Unavailable, the
oracle agrees, and detect() falls through.

Tests: fake device files (a register's bytes at its offset) cover a two-socket
host read on each package's first CPU, the 32-bit wrap, AMD's registers, a zero
unit, an unknown CPU, a missing device, powercap preferred, and the production
probe over MSR packages (live, and frozen when one package holds). Unit tests
pass on Linux x86_64 (default and cuda) and on macOS (the CI step, its
`ioreport liveness:` line printed); clippy is clean on macOS and on Linux in
the default build.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

0b08e3d161

• View at this point in history

2026-11-04 21:01:11 -05:00