Faiss tuning issue

[JinHai-CN]

Dear guys, I am tuning faiss performance and meet following issues:

1. distance_compute_bias_threshold default is 20 which will influence the L2 precision. According to WIKI and negative distance returned in IndexFlatL2 search query #297, it only works on IndexL2. But when I check the code, it seems Inner Product calculate also impacted by the parameter. Why?

2. I test the FAISS with MKL and OpenBlas in intel x86 platform. Obviously, FAISS with MKL will be more fast than FAISS with OpenBlas, when index is FLAT L2/IP. Question, if the index is IVF, will FAISS with MKL still be faster?

Thank you for help!

[mdouze]

1. right, inner product also switches from the Faiss implementation to BLAS depending on this threshold. However, for inner product there should be no impact on accuracy because the subtraction trick is not used
   https://github.com/facebookresearch/faiss/wiki/Implementation-notes#matrix-multiplication-to-do-many-l2-distance-computations

2. for IVF there are two factors that impact speed: the coarse-level quantization and the actual search. The coarse-level quantization's speed can be impacted by MKL.

[XuPeng-SH]

2. for IVF there are two factors that impact speed: the coarse-level quantization and the actual search. The coarse-level quantization's speed can be impacted by MKL.

Does Openblas affect the actual search time? I tested with Openblas and SSE respectively and I see the actual search time from "indexIVF_stats.search_time" are different. It can be reproduced repeatedly.

Thanks!

[mdouze]

Yes the search time changes depends on openblas vs. SSE.

[mdouze]

no activity, closing.