GitLab

As noted in blake3 documentation
> To get any performance benefit from multi-threading, the input buffer
size needs to be very large. As a rule of thumb on x86_64, there is no
benefit to multi-threading inputs less than 128 KiB. Other platforms
have different thresholds, and in general you need to benchmark your
specific use case. Where possible, memory mapping an entire input file
is recommended, to take maximum advantage of multi-threading without
needing to tune a specific buffer size. Where memory mapping is not
possible, good multi-threading performance requires doing IO on a
background thread, to avoid sleeping all your worker threads while the
input buffer is (serially) refilled. This is quite complicated compared
to memory mapping.