[Performance] Prefetch for on-disk inverted list

[pseudo-xqr]

Summary

Enabling prefetch with the default number of prefetch threads (32) for onDiskInvertedLists seems to degrade overall performance.

Platform

OS: Linux
Faiss version: v1.13.2-7-gc9bab48d1
Running on: CPU with 64 cores
Interface: C++ (Python should work as well)
Dataset used: hotpotqa

Reproduction instructions

I prebuilt the vector database on hotpotqa dataset with all-MiniLM-L6-v2 model, and save the trained vector database as the onDiskInvertedLists. The core code of running search is attached below:

vector<float> embed_dist(nq * k);
vector<idx_t> embed_idx(nq * k);
faiss::Index* index = faiss::read_index(index_path.c_str());
faiss::IndexIVFFlat* index_ivf = dynamic_cast<faiss::IndexIVFFlat*>(index);
faiss::IndexFlatL2* quantizer = dynamic_cast<faiss::IndexFlatL2*>(index_ivf->quantizer);
int nt = std::min(omp_get_max_threads(), (int)nq);
index_ivf->nprobe = nprobe;
index_ivf->search(nq, xq, k, embed_dist.data(), embed_idx.data());

In index_ivf->search function, if I comment out the function call prefetch_lists in the sub_search_func, the overall search latency is lower than without commenting out the function call.

Another thing that I have tried is to reduce the number of prefetch threads, but the performance is still worse than without prefetching. The experiments that

Config	Overall Latency	search_time captured by ivf_stats
No prefetch	10995ms	674501ms
Prefetch_nthread=32	11928ms	711488ms
Prefetch_nthread=8	11685ms	693821ms
Prefetch_nthread=4	11557ms	689031ms
Prefetch_nthread=1	11862ms	693961ms

Is this performance trend as expected to see performance degradation with the default 32 prefetch threads? Under what configuration does using 32 prefetch threads be performance-wise beneficial?

[mnorris11]

cc @pankajsingh88 ?