[NeurIPS 2023 OOD] Update of parameters by team UTokyo

[maronuu]

This is a PR that updates the parameter for epsearch.
The original PR is #197 .
If you encounter the same kind of issues described in #197 , please follow the comment in the PR.

Could you please re-run the alogrithm? Some part of construction can utilize cache in the previous evaluation.
Thank you for the patient evaluation and sorry for the multiple updates.

[maronuu]

I have no idea why the CI Billion-Scale ANN Benchmarks, NeurIPS 2023 / build (sustech-ood, random-xs, ood) (pull_request) Failing after 27m is failing...

[maumueller]

I have no idea why the CI Billion-Scale ANN Benchmarks, NeurIPS 2023 / build (sustech-ood, random-xs, ood) (pull_request) Failing after 27m is failing...

This is not related to your PR, so do not worry about it.

[maronuu]

Note

Here is a note that might be helpful for organizers to evaluate our submission.

Comment

Here are some notations that might be helpful to understand my source code

• Dockerfile is much longer compared to other submission, sorry, but it installs minimum requirements for our submission.
• ENV MKL_NUM_THREADS or similar envs are introduced to make sure that our algoirthm fully utilizes the cores.

Evaluation in my env

In my local environment, refining docker container using only 8 vCPU, and 16GB RAM,
we have:

• baseline diskann: ~1600 QPS
• our method: ~2400 QPS

And, in aws instance c6i.2xlarge, we have:

• baseline diskann: 3882 QPS
• our method: 4941 QPS