Skip to content

PoC: evaluate PiPNN bulk indexing and graph search on ARM - #1122

Draft
dor-forer wants to merge 1 commit into
mainfrom
codex/pipnn-arm-poc
Draft

dor-forer wants to merge 1 commit into
mainfrom
codex/pipnn-arm-poc

Conversation

@dor-forer

Copy link
Copy Markdown
Collaborator

Describe the changes in the pull request

Add an isolated ARM proof of concept evaluating PiPNN for bulk index construction and graph search. This draft contains benchmark drivers, pinned-source build patches/instructions, data preparation, raw measurements and reports under poc/pipnn-arm/. It does not add a production backend or change VectorSimilarity APIs.

On 100k normalized 1536D DBpedia vectors, serial queries and one build thread, medians of two runs are:

Configuration Construction seconds* p95 at >=95% recall p95 at >=99% recall
HNSW M16, efConstruction=100 59.4 0.72 ms 4.57 ms
HNSW M16, efConstruction=200 119.2 0.63 ms 2.23 ms
PiPNN alpha=1.1, one tree 62.6 0.77 ms 1.85 ms

*Construction boundaries differ: PiPNN includes serialization but excludes query-process graph loading; HNSW is queryable after construction. Compiler arithmetic flags also differ. These are implementation measurements, not equivalent ready-to-serve or algorithm-only speedups. Search efforts are a finite sweep; the cheaper HNSW configuration's 99% point could improve with a denser sweep.

The result supports further high-recall validation, not replacing HNSW. Pruning tuning did not eliminate the 95%-recall latency gap against the stronger HNSW build. Two trees increased build cost without a compelling advantage over the selected single-tree candidate.

Validation

  • 19 complete ARM runs across both phases; phase 1 covers DEEP 96D and DBpedia 1536D, each with 100k vectors.
  • Distance checks and small brute-force query smoke tests passed on ARM.
  • Selected settings repeated; recall checked against recomputed subset ground truth with exact-overlap and near-tie exclusions.
  • Both summaries regenerated from packaged logs and compared with saved results; final artifact review performed independently.
  • Existing working-checkout changes excluded. No new native build was needed to package this draft.

Start with poc/pipnn-arm/PHASE2_RESULTS.md and RESULTS.md. Host-specific runner/data paths are documented in the README; datasets, binaries and graph files are not committed. Upstream license attribution is included.

Before production consideration: untouched-query validation, denser effort sweeps, aligned compiler settings, common readiness timing, larger/multithreaded runs, and update/delete/serving integration. Identical PiPNN RSS readings across configurations remain qualified in the report.

Which issues this PR fixes

None; exploratory draft PoC.

Main objects this PR modified

  • poc/pipnn-arm/: experimental harnesses, patches, documentation and measurement evidence only.

Mark if applicable

  • This PR introduces API changes
  • This PR introduces serialization changes

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant