【Why is the throughput of testing the same algorithm with raft ann benchmark one order of magnitude higher than using ann benchmark】

[zjx1230]

Hello, Everyone! Recently, I tested the performance of hnswlib using raft-ann-benchmark and ann-benchmarks, and found that under the same recall rate, the QPS of raft-ann-benchmark is about an order of magnitude higher than that of ann-benchmarks. Is this due to the overhead of Python calling the cpp library when calculating time in ann-benchmarks?

environment: x86_64,CentOS7,96core raft-ann-benchmark command python -m raft-ann-bench.run --force --dataset glove-100-inner --algorithms hnswlib -m throughput

ann-benchmarks command python run.py --local --algorithm hnswlib --batch