diff --git a/claude-plugin/skills/aae/knowledge-base/040-parhip-bit-identical-speedup.md b/claude-plugin/skills/aae/knowledge-base/040-parhip-bit-identical-speedup.md index 1166590..26fbbc5 100644 --- a/claude-plugin/skills/aae/knowledge-base/040-parhip-bit-identical-speedup.md +++ b/claude-plugin/skills/aae/knowledge-base/040-parhip-bit-identical-speedup.md @@ -380,7 +380,7 @@ time changed the target. on the social graphs (enwiki 1.047 to 1.010, youtube 1.052 to 1.028), and a reserve of 2 or 4 times the needed hash table size also slowed refinement by 1 to 8%. On the mesh (hugetrace) the loss stayed, and a run-time switch that never built the lists, in the same binary, still showed - 7% more refinement work than the base, while two diag binaries that differ by one getenv call + 7% more refinement work than the base, while two diag binaries that differ by one environment-variable lookup differed by 3%. So part of such a loss is a binary effect, not memory: separate the two with a run-time switch in one binary before concluding footprint, and let the timed net on the full set decide. diff --git a/knowledge-base/040-parhip-bit-identical-speedup.md b/knowledge-base/040-parhip-bit-identical-speedup.md index 1166590..26fbbc5 100644 --- a/knowledge-base/040-parhip-bit-identical-speedup.md +++ b/knowledge-base/040-parhip-bit-identical-speedup.md @@ -380,7 +380,7 @@ time changed the target. on the social graphs (enwiki 1.047 to 1.010, youtube 1.052 to 1.028), and a reserve of 2 or 4 times the needed hash table size also slowed refinement by 1 to 8%. On the mesh (hugetrace) the loss stayed, and a run-time switch that never built the lists, in the same binary, still showed - 7% more refinement work than the base, while two diag binaries that differ by one getenv call + 7% more refinement work than the base, while two diag binaries that differ by one environment-variable lookup differed by 3%. So part of such a loss is a binary effect, not memory: separate the two with a run-time switch in one binary before concluding footprint, and let the timed net on the full set decide.