RIKEN-RCCS · yoshifuminakamura · Sep 18, 2026 · Sep 18, 2026 · Sep 18, 2026
diff --git a/docs/guides/profiler-support.md b/docs/guides/profiler-support.md
@@ -58,6 +58,47 @@ bk_profiler <tool> [options] -- <command ...>
 `bk_resolve_profiler_tool` / `bk_resolve_profiler_level` を使う。既存運用や site-local
 調整のために app 固有変数が必要な場合は、共通変数の上書きとして helper に渡す。
 
+## Workflow Stage Timing
+
+`bk_profiler` は各 `fapp` / `ncu` acquisition と report export の時間を自動記録する。
+アプリの `profile.sh` に開始・終了時計、JSON生成、artifact登録を実装する必要はない。
+SBDとGENESISも同じ共通経路を使い、MPI rank選択やcontainer起動条件はアプリ側で組み立てる。
+
+`run.sh` では、実験条件と出力先を決めた時点で共通実行contextを設定する。
+contextはsubshellやpipelineへ引き継がれ、作業ディレクトリを変えても出力先は変わらない。
+
+```bash
+bk_run_context --results-dir "${PWD}/results" --exp "$experiment"
+bk_run --log benchmark.log -- mpirun -np "$ranks" ./application
+bk_profiler ncu --archive "${BK_RUN_RESULTS_DIR}/padata.tgz" -- ./application
+```
+
+独自のMPI/container起動が必要な場合は、組み立てた完全なコマンドを
+`bk_profile_execute --tool <tool> --phase <collect|export|plan> -- <command ...>`
+へ渡す。`--profile <name>` で対象profileを識別でき、`--log <file>` でコマンドの
+stdout/stderrだけをファイルへ送れる。NSYSのcollect/exportにもこの入口を使う。
+NCU plan生成は `bk_generate_ncu_plan` へ従来のgenerator引数を渡す。
+いずれも時計操作や記録ファイルの管理は共通層が担当する。
+
+共通層は実験条件ごとの `results/workflow_timing_*.json` にstage、tool、profile、
+UTC開始・終了時刻、monotonic時計による経過秒、終了コードを保存する。
+更新はlockとatomic replacementを使い、並列コマンドの追記を保持する。
+`scripts/result.sh` が現在の実行sessionの記録を `timing_observations` へ自動追加し、
+既存のapp独自timing observationも保持する。Resultの `Exp` によりscopeを絞り、
+既存のMeasurement Artifacts送信・console表示経路を使う。公開面には追加しない。
+context未設定の `bk_profiler` はarchiveの出力ディレクトリを使い、job単位のscopeとなる。
+一つのresultsディレクトリは一つの実行sessionで使用する。
+
+記録はコマンドのwall timeであり、MPI起動やprofilerのreport finalizationを含む。
+queue待ちやarchive圧縮は含まず、アプリのFOMを置き換えない。profile用の再実行時間を
+そのままアプリのslowdown比とは解釈せず、同条件の通常実行と比較する。
+計測用helper自身の起動・記録には小さな追加コストがある。
+
+失敗・timeoutはコマンドの終了コードを保持する。finishを書けずに中断したstageは
+`running` と開始時刻だけが残るため、未完了として扱い、成功やゼロ秒とみなさない。
+任意の記録処理が失敗しても実行コマンドの終了コードを変えない。schedulerによる強制終了後の
+転送可否はjobのartifact upload方針に依存する。
+
 ## 3. 共通語彙としての level
 
 `single/simple/standard/detailed` は Benchkit の共通語彙として扱う。

diff --git a/programs/genesis/profile.sh b/programs/genesis/profile.sh
@@ -428,7 +428,8 @@ genesis_run_container_ncu_acquisition_profile() {
     echo "bk_run_ncu_acquisition_profile: profile='${profile_name}' kernel='${kernel_regex}' skip=${launch_skip} count=${launch_count}" >&2
     echo "bk_profiler[ncu]: starting ${rep_name} level=${profiler_level} inside container rank 0" >&2
     set +e
-    "${profile_cmd[@]}" </dev/null 2>&1 | tee "$profile_log"
+    bk_profile_execute --tool ncu --phase collect --profile "$profile_slug" -- \
+        "${profile_cmd[@]}" </dev/null 2>&1 | tee "$profile_log"
     profiler_status=${PIPESTATUS[0]}
     set -e
 
@@ -446,12 +447,14 @@ genesis_run_container_ncu_acquisition_profile() {
             --csv \
             --print-units base \
             --print-fp || return 1
-        "${import_cmd[@]}" > "${rep_dir}/profile_raw.csv" 2> "${rep_dir}/profile_raw.csv.log" || true
+        bk_profile_execute --tool ncu --phase export --profile "${profile_slug}/raw" -- \
+            "${import_cmd[@]}" > "${rep_dir}/profile_raw.csv" 2> "${rep_dir}/profile_raw.csv.log" || true
 
         import_cmd=()
         genesis_build_container_once_command app_cmd import_cmd \
             ncu --import "$report_file" --page details || return 1
-        "${import_cmd[@]}" > "$stage_dir/reports/ncu_import_${rep_name}.txt" 2>&1 || true
+        bk_profile_execute --tool ncu --phase export --profile "${profile_slug}/details" -- \
+            "${import_cmd[@]}" > "$stage_dir/reports/ncu_import_${rep_name}.txt" 2>&1 || true
     fi
 
     cp -R "$rep_dir" "$stage_dir/raw/${rep_name}"
@@ -610,9 +613,9 @@ genesis_generate_ncu_plan() {
                 "${discovery_cmd[@]:$GENESIS_APPTAINER_PAYLOAD_INDEX}"
             )
             genesis_build_container_rank0_profile_command nsys_payload discovery_cmd nsys_profile_cmd
-            "${nsys_profile_cmd[@]}" 2>&1 | tee "$nsys_log" >&2
+            bk_profile_execute --tool nsys --phase collect -- "${nsys_profile_cmd[@]}" 2>&1 | tee "$nsys_log" >&2
         else
-            nsys profile \
+            bk_profile_execute --tool nsys --phase collect -- nsys profile \
                 --force-overwrite=true \
                 --trace=cuda \
                 --sample=none \
@@ -633,10 +636,11 @@ genesis_generate_ncu_plan() {
         if genesis_find_apptainer_payload_index discovery_cmd; then
             genesis_build_container_once_command discovery_cmd nsys_stats_cmd \
                 nsys stats --force-export=true --report cuda_gpu_kern_sum --format csv --output "$nsys_csv" "$nsys_report" || return 1
-            "${nsys_stats_cmd[@]}" >/dev/null
+            bk_profile_execute --tool nsys --phase export -- "${nsys_stats_cmd[@]}" >/dev/null
             nsys_stats_status=$?
         else
-            nsys stats --force-export=true --report cuda_gpu_kern_sum --format csv --output "$nsys_csv" "$nsys_report" >/dev/null
+            bk_profile_execute --tool nsys --phase export -- \
+                nsys stats --force-export=true --report cuda_gpu_kern_sum --format csv --output "$nsys_csv" "$nsys_report" >/dev/null
             nsys_stats_status=$?
         fi
         if [ "$nsys_stats_status" -ne 0 ]; then
@@ -670,7 +674,7 @@ genesis_generate_ncu_plan() {
         esac
     fi
 
-    "$python_bin" "${SCRIPT_DIR}/scripts/profiling/generate_ncu_plan.py" \
+    bk_generate_ncu_plan \
         --nsys-csv "$discovery_csv" \
         --out-discovery "$discovery_json" \
         --out-plan "$plan_json" \

diff --git a/programs/genesis/run.sh b/programs/genesis/run.sh
@@ -28,6 +28,7 @@ input=${header}.inp
 resultsdir=${SCRIPT_DIR}/results
 artifactsdir=${SCRIPT_DIR}/artifacts
 mkdir -p ${resultsdir}
+bk_run_context --results-dir "${resultsdir}" --exp "$exp"
 export BK_INPUT_INFO_FILE="${resultsdir}/input_info.json"
 export BK_INPUT_INFO_ITEMS_FILE="${resultsdir}/.input_info_items.jsonl"
 bk_reset_input_info
@@ -162,7 +163,7 @@ run_genesis_nvidia_gpu() {
     genesis_configure_ncu_profile "$system_name" "$profiler_tool_var" "$profiler_level_var" "$module_var" || return 1
 
     echo "Running ${system_name} as NVIDIA GPU benchmark run without profiler"
-    "${nvidia_mpi_cmd[@]}" ./${binary} ${input}.sub 2>&1 | tee ${output}
+    bk_run -- "${nvidia_mpi_cmd[@]}" ./${binary} ${input}.sub 2>&1 | tee ${output}
     genesis_run_configured_ncu_profiles "$system_name" "${nvidia_mpi_cmd[@]}" ./${binary} ${input}.sub || return 1
 }
 
@@ -203,7 +204,9 @@ case "$system" in
     export PARALLEL=${nthreads}
     export OMP_NUM_THREADS=${nthreads}
 	echo "${mpi_cmd} ./${binary} ${input}.sub"
-	${mpi_cmd} ./${binary} ${input}.sub
+	# Intentional word splitting: mpi_cmd contains the launcher and its arguments.
+	# shellcheck disable=SC2086
+	bk_run -- ${mpi_cmd} ./${binary} ${input}.sub
 	[[ -f ./stdout.1.0 ]] && cp ./stdout.1.0 ${output}
 	[[ -f ./stderr.1.0 ]] && cp ./stderr.1.0 ${stderr}
     ;;

diff --git a/programs/qws/run.sh b/programs/qws/run.sh
@@ -109,6 +109,7 @@ case "$system" in
                 mpiexec -n 2 ./main "${qws_case1_args[@]}" > CASE1
                 print_results output.${PJM_JOBID}/0/2/stdout.2.0 CASE1 2 >> ../results/result
                 if bk_profiler_enabled "$qws_profiler_tool"; then
+                    bk_run_context --results-dir ../results --exp CASE0
                     bk_profiler "$qws_profiler_tool" --level "$qws_profiler_level" --archive ../results/padata0.tgz --raw-dir pa -- mpiexec -n 1 ./main "${qws_case0_args[@]}" > CASE0.profile
                 # else
                 #     emit_qws_dummy_padata ../results/padata0.tgz
@@ -118,6 +119,7 @@ case "$system" in
                 mpiexec -n 8 ./main "${qws_case7_args[@]}" > CASE7
                 print_results output.${PJM_JOBID}/0/1/stdout.1.0 CASE7 4 >> ../results/result
                 if bk_profiler_enabled "$qws_profiler_tool"; then
+                    bk_run_context --results-dir ../results --exp CASE7
                     bk_profiler "$qws_profiler_tool" --level "$qws_profiler_level" --archive ../results/padata0.tgz --raw-dir pa -- mpiexec -n 8 ./main "${qws_case7_args[@]}" > CASE7.profile
                 # else
                 #     emit_qws_dummy_padata ../results/padata0.tgz

diff --git a/programs/sbd/README.md b/programs/sbd/README.md
@@ -17,9 +17,22 @@ and then uses that plan for rank-0 Nsight Compute acquisition when
 Use `BK_SBD_NCU_PROFILE_MODE=discovery-only` to stop after the NSYS discovery
 and plan generation step. `BK_SBD_NCU_PLAN_TOP_K` controls how many kernels are
 selected; discovery-only defaults to all kernels, while discovery defaults to
-the top three. Generated NCU archives are attached to the `mult` section as
+the top three. The RIKYU system default overrides this selection to the top one.
+Generated NCU archives are attached to the `mult` section as
 measurement artifacts.
 
+## Workflow stage timing
+
+SBD uses the common execution helpers for its ordinary benchmark, NSYS
+collection/export, NCU plan generation and NCU collection/export. Benchkit
+automatically logs and stores their stage timings as Measurement Artifacts;
+the application does not manage timing records. See the
+[shared profiler guide](../../docs/guides/profiler-support.md#workflow-stage-timing)
+for scope, failure handling and interpretation. The internal Davidson FOM
+remains independent of workflow wall time.
+
+## Benchmark recipes
+
 The RIKYU recipe uses the H2O cc-pVDZ FCIDUMP with the `1em7` selected alpha
 determinant file (about 628 million product determinants), one MPI rank per
 B200 GPU, 32 OpenMP threads per rank, and the rank-distributed/index-reordered/NCCL

diff --git a/programs/sbd/profile.sh b/programs/sbd/profile.sh
@@ -146,8 +146,8 @@ sbd_run_rank0_nsys_discovery() {
   local profile_status
 
   echo "Running SBD NSYS kernel discovery for automatic NCU plan generation" >&2
-  set +e
-  mpirun -np "$n_ranks" bash -lc '
+  profile_status=0
+  bk_profile_execute --tool nsys --phase collect --log "$log_file" -- mpirun -np "$n_ranks" bash -lc '
     rank=${OMPI_COMM_WORLD_RANK:-${PMIX_RANK:-${SLURM_PROCID:-0}}}
     local_rank=${OMPI_COMM_WORLD_LOCAL_RANK:-${SLURM_LOCALID:-0}}
     export CUDA_VISIBLE_DEVICES="${local_rank}"
@@ -157,9 +157,7 @@ sbd_run_rank0_nsys_discovery() {
       exec nsys profile --force-overwrite=true --trace=cuda --sample=none -o "$report_base" ./diag "$@"
     fi
     exec ./diag "$@"
-  ' bash "$report_base" "$@" > "$log_file" 2>&1
-  profile_status=$?
-  set -e
+  ' bash "$report_base" "$@" || profile_status=$?
 
   if [ "$profile_status" -ne 0 ]; then
     echo "SBD NSYS kernel discovery failed with status ${profile_status}" >&2
@@ -225,14 +223,12 @@ sbd_generate_ncu_plan() {
       return 1
     fi
 
-    set +e
-    nsys stats --force-export=true \
+    nsys_stats_status=0
+    bk_profile_execute --tool nsys --phase export -- nsys stats --force-export=true \
       --report cuda_gpu_kern_sum,cuda_api_sum \
       --format csv \
       --output "$nsys_stats_base" \
-      "$nsys_report" >/dev/null
-    nsys_stats_status=$?
-    set -e
+      "$nsys_report" >/dev/null || nsys_stats_status=$?
     if [ "$nsys_stats_status" -ne 0 ]; then
       echo "SBD NSYS CUDA summary export failed with status ${nsys_stats_status}" >&2
       return "$nsys_stats_status"
@@ -263,7 +259,7 @@ sbd_generate_ncu_plan() {
     esac
   fi
 
-  if ! "$python_bin" "${SBD_BENCHKIT_ROOT}/scripts/profiling/generate_ncu_plan.py" \
+  if ! bk_generate_ncu_plan \
     --nsys-csv "$discovery_csv" \
     --out-discovery "$discovery_json" \
     --out-plan "$plan_json" \
@@ -394,14 +390,14 @@ sbd_run_rank0_ncu_profile() {
     ' bash "$profile_base" "${#ncu_level_args[@]}" "${ncu_level_args[@]}" \
       "$kernel_regex" "$launch_skip" "$launch_count" "$@"
   )
-  set +e
+  profiler_status=0
   if [ "$profile_timeout_seconds" -gt 0 ]; then
-    timeout --kill-after=60s "$profile_timeout_seconds" "${profile_cmd[@]}" > "$profile_log" 2>&1
+    bk_profile_execute --tool ncu --phase collect --profile "$profile_slug" --log "$profile_log" -- \
+      timeout --kill-after=60s "$profile_timeout_seconds" "${profile_cmd[@]}" || profiler_status=$?
   else
-    "${profile_cmd[@]}" > "$profile_log" 2>&1
+    bk_profile_execute --tool ncu --phase collect --profile "$profile_slug" --log "$profile_log" -- \
+      "${profile_cmd[@]}" || profiler_status=$?
   fi
-  profiler_status=$?
-  set -e
 
   if [ "$profiler_status" -eq 0 ]; then
     echo "bk_profiler[ncu]: completed ${rep_name} level=${profiler_level}" >&2
@@ -411,12 +407,13 @@ sbd_run_rank0_ncu_profile() {
 
   report_file=$(bk_profiler_find_ncu_report "$rep_dir" || true)
   if [ -n "$report_file" ]; then
-    ncu --import "$report_file" \
+    bk_profile_execute --tool ncu --phase export --profile "${profile_slug}/raw" -- ncu --import "$report_file" \
       --page raw \
       --csv \
       --print-units base \
       --print-fp > "${rep_dir}/profile_raw.csv" 2> "${rep_dir}/profile_raw.csv.log" || true
-    ncu --import "$report_file" --page details > "$stage_dir/reports/ncu_import_${rep_name}.txt" 2>&1 || true
+    bk_profile_execute --tool ncu --phase export --profile "${profile_slug}/details" -- \
+      ncu --import "$report_file" --page details > "$stage_dir/reports/ncu_import_${rep_name}.txt" 2>&1 || true
   fi
 
   cp -R "$rep_dir" "$stage_dir/raw/${rep_name}"

diff --git a/programs/sbd/run.sh b/programs/sbd/run.sh
@@ -1,3 +1,3 @@
 #!/bin/bash
 set -euo pipefail

@@ -70,6 +70,8 @@
     ;;
 esac
 
+bk_run_context --results-dir "${RESULTS_DIR}" --exp "${experiment}"
+
 for input_file in fcidump.txt "${determinant_file}"; do
   if [[ ! -f "${INPUT_DIR}/${input_file}" ]]; then
     echo "SBD input not found: ${INPUT_DIR}/${input_file}" >&2
@@ -105,12 +107,12 @@
 )
 
 if [[ "${system}" == "RC_FX700" ]]; then
-  mpirun -np "${n_ranks}" -bind-to numa ./diag "${diag_args[@]}" \
-    > diag.log 2>&1
+  bk_run --log diag.log -- \
+    mpirun -np "${n_ranks}" -bind-to numa ./diag "${diag_args[@]}"
 else
-  mpirun -np "${n_ranks}" bash -lc \
+  bk_run --log diag.log -- mpirun -np "${n_ranks}" bash -lc \
     'export CUDA_VISIBLE_DEVICES=$OMPI_COMM_WORLD_LOCAL_RANK; exec "$@"' \
-    bash ./diag "${diag_args[@]}" > diag.log 2>&1
+    bash ./diag "${diag_args[@]}"
 fi
 
 davidson_time=$(grep -E 'Elapsed time for davidson ' diag.log | tail -n 1 | awk '{print $(NF-1)}')

diff --git a/result_server/tests/test_result_detail_template.py b/result_server/tests/test_result_detail_template.py
@@ -397,6 +397,74 @@ def test_public_surface_keeps_padata_archive_links(self, app):
         assert "Measurement Artifacts" in html
         assert f'href="/results/{filename}"' in html
 
+    @pytest.mark.parametrize("collection", ["sections", "overlaps"])
+    @pytest.mark.parametrize("uploaded", [True, False])
+    @pytest.mark.parametrize("public_surface", [True, False])
+    def test_profile_metadata_links_respect_surface_and_upload_state(
+        self, app, collection, uploaded, public_surface,
+    ):
+        basenames = ["kernel_discovery.json", "ncu_plan.json", "padata_kernel.metadata.json"]
+        filenames = [
+            "measurement_artifact_20250101_120000_"
+            f"12345678-1234-1234-1234-123456789abc_{name}"
+            for name in basenames
+        ]
+        result = {
+            "code": "demoapp",
+            "_server_uuid": "12345678-1234-1234-1234-123456789abc",
+            "_server_timestamp": "20250101_120000",
+            "fom_breakdown": {
+                collection: [{
+                    "name": "kernel",
+                    "artifacts": [
+                        {"type": "file_reference", "path": f"results/{name}"}
+                        for name in basenames
+                    ],
+                }],
+            },
+        }
+
+        with app.test_request_context():
+            html = _render_result_detail(
+                result, {}, filenames if uploaded else [], public_surface=public_surface,
+            )
+
+        for basename, filename in zip(basenames, filenames):
+            if public_surface:
+                assert basename not in html
+            else:
+                assert "Profile metadata" in html
+                assert f"results/{basename}" in html
+                if uploaded:
+                    assert f'href="/results/{filename}"' in html
+                else:
+                    assert f"{filename} not uploaded" in html
+                    assert f'href="/results/{filename}"' not in html
+
+    @pytest.mark.parametrize("artifact_path", [
+        "../metadata.json",
+        "results/../metadata.json",
+        "/results/metadata.json",
+        "artifacts/metadata.json",
+        "results/bad name.json",
+        "results/metadata.txt",
+    ])
+    def test_invalid_profile_metadata_paths_are_omitted(self, app, artifact_path):
+        result = {
+            "_server_uuid": "12345678-1234-1234-1234-123456789abc",
+            "_server_timestamp": "20250101_120000",
+            "fom_breakdown": {
+                "sections": [{
+                    "name": "kernel",
+                    "artifacts": [{"type": "file_reference", "path": artifact_path}],
+                }],
+            },
+        }
+        with app.test_request_context():
+            context = build_result_detail_context(result, {})
+
+        assert context["measurement_artifact_rows"] == []
+
     def test_timing_observation_artifact_is_linked_on_console_surface(self, app):
         result = {
             **FULL_RESULT,