SEP NOTES
Reference source:https://github.com/nietras/Sep
Purpose:
Reference system for benchmark organization, scope separation, machine specification discipline, and future SIMD-path benchmarking methodology.
Observed useful patterns:
- Benchmark scope separation
Benchmark layers should remain isolated.
Example principle:
core algorithm benchmark
!=
system benchmark
!=
startup benchmark
!=
IPC benchmark
!=
HTTP benchmark
Interpretation for GLYPH:
Occ/rank benchmarks should be isolated from:
- mmap startup
- backend startup
- HTTP layer
- IPC overhead
- shard orchestration
Benchmark signal becomes cleaner.
- Machine specification discipline
Benchmarks without machine context lose value.
Machine profile should always include:
- CPU model
- core/thread count
- RAM
- compiler
- SIMD capability
- kernel
- corpus size
- shard count
Interpretation:
benchmark = contract
- SIMD direction
Future GLYPH acceleration direction:
Occ()
↓
scalar baseline
↓
AVX2
↓
AVX512
Goal:
reduce rank latency
reduce shard fanout overhead
increase deterministic retrieval throughput
Future GLYPH work:
- SIMD Occ layer
- isolated Occ benchmark
- benchmark reproducibility discipline
Strategic interpretation:
benchmark != marketing
benchmark = engineering contract