8000 Improve Vamana index build performance and recall by bkarsin · Pull Request #1032 · rapidsai/cuvs · GitHub
[go: up one dir, main page]
More Web Proxy on the site http://driver.im/
Skip to content

Improve Vamana index build performance and recall #1032

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.

Sign up for GitHub

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

Jump to bottom
Open
wants to merge 14 commits into
base: branch-25.08
Choose a base branch
from
Open

Improve Vamana index build performance and recall #1032

wants to merge 14 commits into from

Conversation

bkarsin
Copy link
Contributor
@bkarsin bkarsin commented Jun 19, 2025

Series of optimizations for the Vamana index build implementation including:

  • Reduce shared memory usage by the main kernels: GreedySearch and RobustPrune
  • Replace cub::blockSort with manual merge that reduces shared memory and improves perf
  • Remove other cub Sort calls to separate k 8000 ernel to improve resource utilization
  • Other changes to resource allocation and management to improve performance

Other changes that significantly improve recall (closing most of the gap with CPU build and giving the option to further improve recall with parameters):

  • Re-worked the RobustPrune kernel to use a multi-pass "occlusion" approach. This closes 60-80% of the recall gap seen between previous GPU Vamana and CPU DiskANN.
  • Changed the "vamana_iters" parameter to a float to allow fractional iterations. This lets us precisely control recall vs. performance to close the gap (or surpass CPU recall).

Further optimizations are underway, but this first round address many needs of users to both improve performance and recall.

bkarsin added 14 commits February 25, 2025 01:31
@bkarsin
Shared-memory reducing optimizations for GreedySearch and RobustPrune
df07c2b
@bkarsin
Draft of work in progress for occlusion pruning
b32ee4e
@bkarsin
Remove debug code
9fa27f9
@bkarsin
New occlusion robustprune code to close recall gap
6684ee4
@bkarsin
Shared-memory reducing optimizations for GreedySearch and RobustPrune
529d11e
@bkarsin
Draft of work in progress for occlusion pruning
815b49f
@bkarsin
Remove debug code
7bdac15
@bkarsin
change iters to float and fix memory allocation for edge_dist_pair
04f90ab
@bkarsin
Merge branch 'rapidsai:branch-25.08' into recall_fix
ced26dc
@bkarsin
clean up code and run clang-format
6bc19ce
@bkarsin
clean up code and run clang-format
e80ef96
@bkarsin
Merge remote-tracking branch 'origin/recall_fix' into vamana_opt
ced2ad4
@bkarsin
Merge remote-tracking branch 'origin/branch-25.08' into recall_fix
2ff9c61
@bkarsin bkarsin requested review from a team as code owners June 19, 2025 01:04
@copy-pr-bot copy-pr-bot
Copy link
copy-pr-bot bot commented Jun 19, 2025

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

@bkarsin bkarsin added improvement Improves an existing functionality non-breaking Introduces a non-breaking change C++ and removed cpp CMake labels Jun 19, 2025
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers
No reviews
Assignees
No one assigned
Labels
C++ improvement Improves an existing functionality non-breaking Introduces a non-breaking change
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
1 participant
@bkarsin
0