Erik Bernhardsson

multiprocessing should add no CPU overhead since GIL doesn't matter any more. The only issue is since it's using fork(), the memory usage might go up (potentially a lot, if...

I'm skeptical about batch queries, would rather not let algorithms use that. I can switch back to single threads if you can give me some pointers about how to implement...

@yurymalkov I noticed nmslib doesn't use the Py_BEGIN_ALLOW_THREADS macro. Why not? If you enable that macro (or maybe it's called something else when using pybind11) then you nmslib should be...

Sounds good. When was this released? I just ran a benchmark for cosine distance and I'm running one for Euclidean right now (see #50 for preliminary stats) but let me...

ok weird. i'll keep the benchmarks for now then. nmslib is still like 3x faster than anything else :)

The version in ann-benchmarks is the latest version of git which is also the latest version in PyPI: 1.11.1 (with AVX optimizations added)

Let me see what happens if I swap out the thread pool for a process pool instead. Might also help with L1/L2 thrashing.

ok – i can also try to increase the number of blobs? i think everything else equals, it's good to have a reproducible pipeline. if we end up using your...

thanks for running this experiment btw!!

The real utility here is converting a pfx to a pem file – not the temporary file context manager. Arguably there's nothing requests-specific about handling a pfx file, so I'm...