Menu
DFTracer documentation links, artifact log locations, step numbering, and how to read session logs
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Basé sur la classification professionnelle SOC
Communication-component (MPI/collective/network) bottleneck-to-optimization mappings, papers, and L1/L2/L3 strategies for the dftracer optimization pipeline
Compute-component bottleneck-to-optimization mappings, papers, and L1/L2/L3 strategies for the dftracer optimization pipeline
Key literature, bottleneck-to-optimization mappings, and strategies for the dftracer I/O optimization pipeline
Memory-component bottleneck-to-optimization mappings, papers, and L1/L2/L3 strategies for the dftracer optimization pipeline
VPIC-Kokkos-specific knowledge: build/annotation quirks, the two-stage deck-compile pattern, deck sizing for smoke vs validation runs, and the measured MPI-communication-bound diagnosis on Tuolumne. Load this skill whenever working with vpic-kokkos (github.com/lanl/vpic-kokkos).
Install and build dftracer into the session using the MCP install tool and record tool pitfalls early.
| name | dftracer-references |
| description | DFTracer documentation links, artifact log locations, step numbering, and how to read session logs |
Use these two papers when diagnosing bottlenecks and proposing optimizations
(see also .agents/skills/dftracer-io-optimization/SKILL.md for the full mapping table):
Use these links when implementing DFTracer behavior:
HARD RULE — every log goes under <workspace>/artifacts/. That includes logs
you produce yourself: build output, run stdout/stderr, saved Bash output, and
scratch diagnostics. Never leave a log only in the terminal. Never write logs to
<workspace>/tmp/ (reserved for wrapper scripts and scratch inputs) or anywhere
outside the session workspace. Name your own logs <step>_<what>.log (e.g.
opt3_run.log, opt1_build.log) so session_final_report can collect them.
Every session_* tool call automatically writes a stage log to:
<workspace>/artifacts/<NN>_<stage_name>.log
Step numbering:
| NN | Log file | Session tool |
|---|---|---|
| 01 | 01_session_create.log | session_create |
| 02 | 02_session_detect.log | session_detect |
| 03 | 03_session_configure.log | session_configure |
| 04 | 04_session_build_install.log | session_build_install |
| 05 | 05_session_run_smoke_test.log | session_run_smoke_test |
| 06 | 06_session_copy_annotated.log | session_copy_annotated |
| 07 | 07_session_patch_build.log | session_patch_build |
| 08 | 08_session_annotate_source.log | session_annotate_source |
| 09 | 09_session_install_dftracer.log | session_install_dftracer |
| 10 | 10_session_build_annotated.log | session_build_annotated |
| 11 | 11_session_run_with_dftracer.log | session_run_with_dftracer |
| 12 | 12_session_split_traces.log | session_split_traces |
| 13 | 13_session_analyze_traces.log | session_analyze_traces |
Each log contains: timestamp, run_id, step number, stdout, stderr, exit code, and any stage-specific fields (commands, file counts, patch paths, etc.).
After each step, review the corresponding log to verify:
Annotation correctness check: after session_annotate_source, inspect
<workspace>/annotation.patch — it is a unified diff of every file modified.
Use patch --dry-run -p1 < annotation.patch from source/ to confirm it applies cleanly.
On failure: read the log for the failing step, diagnose from stderr, then call the
individual session_* tool directly (not the full pipeline) to retry that step.
