菜单
Cross-session, citation-backed knowledge base of every MEASURED dftracer optimization, partitioned into system-centric (L3), software-centric (L2), and workload-centric (L1) findings. Load this FIRST in any optimization loop to avoid re-deriving what is already known.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
基于 SOC 职业分类
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/llnl/dftracer-agents --skill dftracer-optimization-kb命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
正在显示 SKILL.md
| name | dftracer-optimization-kb |
| description | Cross-session, citation-backed knowledge base of every MEASURED dftracer optimization, partitioned into system-centric (L3), software-centric (L2), and workload-centric (L1) findings. Load this FIRST in any optimization loop to avoid re-deriving what is already known. |
Step 1 of every optimization loop is recall, not proposal. Call
opt_kb_lookup(system=..., workload=..., software=..., scope=...) before
generating any proposal, and cite prior results in the proposal table.
Scopes and what they transfer to:
| Scope | Level | Transfers to | File |
|---|---|---|---|
| system | L3 | any workload on that system | system.md |
| software | L2 | any workload linking that software, any system | software.md |
| workload | L1 | that application, any system | workload.md |
Recorded: 11 system, 7 software, 26 workload entries.
scope (above) answers who inherits a finding. metric_scope answers a
different question — which metric moved:
| metric_scope | What it measures |
|---|---|
app (default) | The app's own trace: epoch/I-O time, app-observed bandwidth. |
system | A filesystem/system-level outcome: aggregate achieved bandwidth, |
| reduced filesystem load — currently a trace-derived proxy, not real | |
| OST/MDT-side telemetry (this pipeline has no Lustre-admin monitoring access). |
Non-degradation guard (MANDATORY): a metric_scope="system" entry MUST
carry a paired app_metric/app_before/app_after — opt_kb_record rejects
one without it. If the paired app metric regressed more than 2.0%, the verdict is force-set to regression
regardless of how good the system-side number looks. A system optimization
that costs the app is not a win — never apply/keep one where this guard fired.
metric, before, after are required.
Record failures too: knowing a lever did nothing is a result.session:<run_id> marks a result measured in-house, never external evidence.metric_scope="system" entry always carries its paired app-metric proof
— see the non-degradation guard above.Language-agnostic dftracer annotation rules — Rule 0 (what to skip), coverage, error paths, backend priority order, and wrapper handling
Python annotation rules for dftracer — decorator usage, initialize/finalize, comp types, class methods, and quick checklist
Lessons learned from dftracer annotation sessions — real errors, root causes, and exact fixes. Loaded by each per-file annotation subagent at startup. Updated by the pipeline recipe after every session (Step 8).