4. PerfSpect baseline report

Capture the platform once, before any measurement. Set the power profile first (Set the power profile) so the baseline records the configured state, not the default one.

If this environment uses an outbound proxy, set LAB_HTTP_PROXY/LAB_HTTPS_PROXY in config/lab.env first. Scripts build NO_PROXY for cluster/node addresses, but manual kubectl commands may still need shell no_proxy/NO_PROXY.

Run it

scripts/run-perfspect.sh              # LAB_DEFAULT_NODE
scripts/run-perfspect.sh <node>       # a specific worker

Three stages: install Intel PerfSpect on the worker if it is not there, run the full configuration report as root, copy it back to the controller.

Output lands in:

results/perfspect/<node>/baseline-<UTC timestamp>/     the report (html, json, xlsx)
results/perfspect/<node>/latest                        symlink to the newest

PerfSpect reads MSRs and DMI, so it needs root on the worker — you will be prompted for the worker's sudo password on the worker's own prompt.

What to check in the report

Report section	What it must say	Why
BIOS	version and date recorded	The comparison baseline.
CPU	model, sockets, cores, `Hyper-Threading` (reference: `Disabled`)	Confirms step 1; must match `LAB_THREADS_PER_CORE` in `config/lab.env`.
NUMA	one node per socket	Confirms SNC off.
Frequency	base, all-core turbo, `Turbo: Enabled`	Encoder throughput scales with all-core turbo, so a different turbo bin changes the density result legitimately.
Power	`System Profile: Performance`, governor `performance`, EPB 0	A `powersave` governor typically costs 1–2 streams.
C-states	only C0/C1 exposed	C6 exit latency causes dropped frames on pinned cores.
Prefetchers	all enabled (default)	Disabled prefetchers change L3 hit ratio and DRAM traffic.
Uncore	max/min frequency	Affects the DRAM and UPI numbers directly.
Memory	DIMM count, size, type and configured speed, channels populated	The theoretical DRAM peak in the summary is `channels × MT/s × 8`. A half-populated machine halves it.
Kernel / OS	version, `isolcpus`/`nohz_full` if any	Reference: Ubuntu 24.04.4, kernel 6.17.0-19-generic.
PMU	not in use by another agent	PCM cannot read uncore counters if something else holds the PMU.
Accelerators	—	Unused here; H.264 encoding is pure software (libx264).

Set the power profile

The report above only reads. Five power/frequency controls decide how much frequency the cores a run is given actually reach, and none of them is visible in the result as anything but a lower stream count:

Setting	Reference value	Where it lives	Why it changes the number
P-state driver	`intel_pstate=active`	kernel argument + `/sys/devices/system/cpu/intel_pstate/status`	`active` is HWP: the hardware picks the P-state from EPP. `passive` hands control to the governor instead. Different control loop, different frequency under a bursty encoder.
Scaling governor	`performance`	`cpu*/cpufreq/scaling_governor`	`powersave` ramps down between frames. Typically costs 1–2 streams.
EPB (PerfSpect-selected source)	`0`	PerfSpect `config` report (`Energy Performance Bias`)	The coarse BIOS/OS hint, 0 = performance, 15 = maximum saving. PerfSpect selects the active source and reports the effective value.
EPP (MSR 0x774)	`0`	`cpu*/cpufreq/energy_performance_preference`	The HWP hint, 0 = performance. Only exists while the driver is in `active` mode.
ELC	`latency`	TPMI, PerfSpect only	Efficiency Latency Control, Sierra Forest and newer. Decides how eagerly the part trades frequency for efficiency at low utilisation.

scripts/configure-power.sh applies all five, from the worker power and frequency block of config/lab.env, using PerfSpect for the MSRs and sysfs for the driver:

scripts/configure-power.sh                    # LAB_DEFAULT_NODE
scripts/configure-power.sh <node>             # a specific worker
scripts/configure-power.sh --verify <node>    # read back only, change nothing

Five stages: install PerfSpect if absent, record the current configuration, switch the P-state driver, apply governor/EPB/EPP/ELC, and install the unit that re-applies them at boot. It ends by reading back P-state from sysfs and governor/EPB/EPP/ELC from PerfSpect config, and exits non-zero if required settings did not take.

Not every setting exists on every part. --elc needs Sierra Forest or newer and --epp needs the driver in active mode; each is applied on its own so an unsupported one cannot take the others down with it. Both report what happened and leave the rest in place. Set LAB_POWER_ELC=skip or LAB_POWER_PSTATE_DRIVER=skip to stop trying.

Using it as evidence

Two rules that make the reports comparable:

Capture before the workload, not during. This is a configuration report, not a profile. Live behaviour comes from Prometheus (06-observability.md) and the MCP profiler (13-mcp-profiling.md).
Recapture after any BIOS or kernel change and keep both. The timestamped directory names make the history readable.

Next: 05-ffmpeg-mxl-container.md.