[BUG] Performance issue on nq=1 topk=1 on ivf index

[JinHai-CN]

Describe the bug
Too much overhead cost on the query

Steps/Code to reproduce behavior
Use sq8h to build a SIFT1B table and do the query with NQ=1 and TOPK=1

Expected behavior
When we check the index search cost, each file will cost about 0.68ms. There are 239 files in total. So the query time should be closed to 163ms. But query time in the log file tell us it is 297ms except the network IO cost.

Environment details
server_version: 0.5.3-release
search_resources: gpu0, gpu1

[JinHai-CN]

All IVF cpu index is not fast enough. Here are some hints to improve the nq=1, topk=1 query performance:

• Profiling the query bottle neck besides the index engine. And improve the whole process.
• Solve the sync query issue between gRPC and DBEngine.
• In faiss, apply multithread in one query vector across different nprobe, even in one nprobe.
• Check why faiss version 1.2 is faster than version 1.6.
• Replace KMeans algorithm with better one.