| 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.
|
dftracer optimization knowledge base
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: 10 system, 7 software, 13 workload entries.
A second, orthogonal axis: metric_scope
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.
Rules
- Record only measured results —
metric, before, after are required.
Record failures too: knowing a lever did nothing is a result.
- Every entry carries a citation: paper (preferred) > official docs > web.
session:<run_id> marks a result measured in-house, never external evidence.
- Apply optimizations one at a time and measure each, or the attribution
is worthless.
- A
metric_scope="system" entry always carries its paired app-metric proof
— see the non-degradation guard above.
