Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Support NM Index and NM BruteForce search #909

Open
15 tasks
cqy123456 opened this issue Oct 23, 2024 · 0 comments
Open
15 tasks

Support NM Index and NM BruteForce search #909

cqy123456 opened this issue Oct 23, 2024 · 0 comments
Assignees

Comments

@cqy123456
Copy link
Collaborator

cqy123456 commented Oct 23, 2024

In order to control memory usage of querynode's raw data search service, we no longer copy the raw data to the interim index and bruteforce methods.
1. a offset base interim index

  • ivf_flat_nm(no memory): with fp16/bf16/fp32 in IP/L2 metric, bin1 in hamming/jaccard;
  • More Simd support for fp16/bf16;
  • More Simd support for fp16/bf16/bin1->fp32;
  • async build index;
  • iterator support for all data type;
  • ivf flat cc range search opt:
    use a small max_empty_result_buckets value(default = 2) may lead a empty result(knn search(topk1000)->mask topk 1000 deleted-> use topk=1000 as radius ->empty range search result)
    if metric = l2: pass a bucket if abs(q->centroids - bucket radius) > radius;
    if metric = ip: pass a bucekt if abs(q->centroids + bucket radius) < radius;
  • more ut/e2e tests;

2. bruteforce search with zero copy:

  • More Simd support for fp16/bf16;
  • remove all (fp16/bf16->fp32) conversion, and use fp16/bf16 directly to do brute force;
  • iterator support for binary type;
  • more ut/e2e tests;

3. milvus part

  • Milvus support ivf flat nm as interimin index;
  • update mmap config;
  • data relayout;
  • more ut/e2e tests;
@cqy123456 cqy123456 self-assigned this Oct 23, 2024
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

No branches or pull requests

1 participant