CXL-ANNS-KX

Research workspace for graph ANNS on unified CXL–SSD memory (ASPLOS-oriented).

One-line thesis: Full-precision vectors often must live on flash; CXL–SSD gives a unified VA + hot window, but the hard problem is residency / stall control (not “mmap is nice”). Naive page caches, blind graph prefetch, and blindly increasing efSearch amplify SSD misses; semantic pin / bounded promote / iso-recall search control are the intended remedies.

docs/
  plan/          ASPLOS research plan + measured motivation tables
  literature/    CXL–ANNS related-work survey
  notes/         CXL memory + CXL–SSD structure notes
  results/       Canonical motivation findings summary
motivation_exps/ Userspace motivation harness on /dev/vmem0 (CXL–SSD software path)
patches/         DiskANN --no_pq_nav experimental patch

Targets gpu01-class setup: Montage CXL Type-3 + /dev/vmem0 (vmem_sw → NVMe).

cd motivation_exps
make
./motivation vmem-identity
./motivation vmem-verify --dir <populated_laion_dir> --nq 8   # must ALL PASS
./motivation vmem-mot-ef  --dir <dir> --nq 16                 # efSearch 100..1600
./motivation vmem-mot-pin --dir <dir> --nq 16 --cache-mb 64   # pin vs nopin
./motivation vmem-mot-me  --dir <dir> --nq 16                 # M-E1/3/4/5
./motivation vmem-mot12   --dir <dir> --nq 16                 # earlier MOT12 suite

Key logs under motivation_exps/results/. Narrative summary: docs/results/MOTIVATION_FINDINGS.md.

Not in this repo: multi-hundred-GB LAION/DiskANN index blobs (keep on local /mnt/disk0).

Official SSD DiskANN cannot turn off in-memory PQ. Experimental flag --no_pq_nav ranks unseen neighbors with disk full-precision reads (same sector layout / aio beam / node cache).

cd DiskANN_cpp
git apply /path/to/patches/diskann-no-pq-nav/diskann-no-pq-nav.patch
# rebuild search_disk_index, then:
./build/apps/search_disk_index ... --no_pq_nav

See motivation_exps/include/vmem_cxl.hpp (require_cxl_vmem_identity): /dev/vmem0 software backend, expected NVMe BDF, Montage mem0/region0. Adjust constants for your machine.

Motivation / problem framing and measurements are in progress. System design (semantic residency + stall-aware search) is specified in docs/plan/ and not fully implemented as a production runtime yet.