V3 perfomance charts

[oscardssmith]

Currently the Readme https://github.com/microsoft/mimalloc#performance has performance numbers from 2021 which means that it is now substantially out of date with V3. Can the benchmarks be re-run to show current performance?

[daanx]

Yeah, it has been too long. The good news is that I am working on updating this but it all takes time to do it right, no eta yet. Also, the benchmark suite https://github.com/daanx/mimalloc-bench has been in active development all this time (shout out to Julien Voisin and Matthew Parkinson for all their help on this), and now has over 20 allocators in it, so you can run that for more recent info (some allocators have improved quite a bit since 2021 (maybe partially due to mimalloc-bench existing :-)), like jemalloc, tcmalloc, and hoard).

I urge careful consideration though -- the mimalloc-bench suite tries to probe difficult situations and one cannot distill it to a single number for the "best" allocator. In particular, the (current) suite doesn't measure well long-running services with large memory footprints where with mimalloc we have a lot of experience now. For micro bench marks for example, you can always be a tad quicker if you never purge OS pages, but any industrial strength allocator needs to handle that in the real world. In my recent blog post https://www.microsoft.com/en-us/research/blog/mimalloc-a-high-performance-scalable-memory-allocator-for-the-modern-era/ I show in the end an example of the tension between scalability (every thread owns its own memory) and memory usage (all memory is maximally shared between threads) -- where an otherwise very fast allocator uses significantly more memory on the Windows thread pool.