PERF-03 · Issue #220 · Diagnostic evidence

T-S01 single-process sampled CPU profile

Source-resolved sample attribution for one repeated Strategy B workload, with a paired unprofiled run.

Timing result: inconclusive. This is one sequential pair on a busy host. The profiled median is lower while the ranges overlap; no profiler overhead, speedup, or latency difference is inferred. The profile identifies a sampled CPU hotspot for this exact diagnostic workload, not exact attribution of the original unprofiled latency.

Paired execution

Unprofiled process wall36.136 s30,000 measured calls

Profiled process wall33.309 s30,000 measured calls

Unprofiled per-call median1.1303 msrange 0.8007–2.3337 ms

Profiled per-call median0.9751 msrange 0.8402–2.1842 ms

Both runs used the same executable; each performed one warmup and 30,000 successful measured calls. Process wall includes process launch, output serialization, and teardown. Each per-call timer surrounds the repeated one_run() request body. These wall and per-call figures are descriptive records, not a controlled performance comparison.

Raw paired per-call records (CSV) · Machine-readable summary (JSON) · Reproduction notes

Sampled CPU attribution

PID 23164 only · 130,813 target-process stack hits · denominator for percentages is target stack hits · 0 reported lost events.

#	Function	Inclusive hits	Share	Source
1	`anonymous namespace`::one_run	126,542	96.74%	`reports/algorithm-results/tooling/syn04/strategy_b_measure.cpp:60`
2	`csl::registry::StrategyPort::run`	118,950	90.93%	`src/registry/registry.cpp:65`
3	`csl::synthesis::strategy_b::DirectJointWorkspace::solve`	116,354	88.95%	`src/synthesis/strategy_b/direct_joint.cpp:218`
4	`csl::adapters::convex::Workspace::solve_native`	104,178	79.64%	`src/adapters/convex/highs.cpp:320`
5	`Highs::optimizeModel`	60,301	46.10%	`HiGHS 1.11.0 highs/lp_data/Highs.cpp:958`
6	`Highs::callSolveQp`	47,804	36.54%	`HiGHS 1.11.0 highs/lp_data/Highs.cpp:3837`

Inclusive percentages overlap along call stacks and must not be summed. Symbols were resolved using the matching executable PDB. This supports the sampled path StrategyPort → DirectJointWorkspace → convex Workspace → HiGHS for this repeated T-S01 workload.

Symbol summary (CSV)

Capture and build

CollectorVisual Studio DiagnosticsHub VSDiagnostics 17.14.36412.14,000 Hz configured sampling

Trace33.79 s468,186 events · 211,895 SampleProf · 137,835 StackWalk

Host12th Gen Intel Core i5-12400F6 physical / 12 logical cores

BuildMSVC 19.44.35215.0Release /O2 + PDB · HiGHS 1.11.0

xperf per-logical-CPU utilization over 34 intervals (33.79 s): mean 12.45%, range 0.03–69.32%; includes the target plus unrelated background work. Host was busy with unrelated build and desktop activity. The ETL and full xperf HTML remain local; their byte sizes and hashes are recorded in the JSON sidecar.

Solver source 2a5ab835295dfd22f02b19d440bb8158267d985a; baseline harness provenance 4fc8e28587a7fe0e4130f1cb435cf6502ba0e4e2. The diagnostic harness changed only the repetition count (7 → 30,000); see patch. The exact no-PDB published binary was not used.

Evidence boundary

This single-process profile provides source-resolved evidence that sampled CPU work during the repeated, individually timed T-S01 one_run() operation is concentrated in the public StrategyPort -> Strategy B DirectJointWorkspace -> convex HiGHS QP path. It supports identifying that path as the sampled CPU hotspot for this exact diagnostic workload. It does not quantify exact unprofiled single-call latency attribution, generalize beyond T-S01, or support speedup, comparison, application-scale, or scalability claims. Independent review is still required for issue closure.

One profiled and one unprofiled execution; no alternation.
Changing host load and overlapping timing ranges make the paired timing comparison inconclusive.
Diagnostic repeated T-S01 only; no generalization to other scenarios, throughput, application scale, or scalability.
Independent review is required; issue #220 remains open.

Bundle manifest: SHA-256 checksums.