A little utility to measure latency at various levels of the memory hierarchy. Only READ/load latency right now, though.
This was basically inspired by the discussion here (forum.nim-lang.org seems to no longer render many historical posts - sorry - that is one reason why I am re-posting this code here):
https://forum.nim-lang.org/t/5734#35832
as well as some lame arguments about cache locality in the context of very hard to measure cold-cache (& cold branch predictor) hash tables where vanilla linear probing basically always wins (sometimes by a lot) in spite of the fact that hot everything L1 micro-benchmarks can make it seem like "pseudorandom probing" can have a (small, probably not CPU-portable) edge.
The --kind=ranElt and =truRan tests here basically emulate hash lookups
while the --kind=shuff emulates cold cache memory loads (but branch predictors
are still hot cache) or a load pattern more like hopping a long linked list or
a very deep tree. (truRan only works on Linux right now.)
This utility (in shuffle mode) is actually not so bad a way to measure memory systems against each other at various data scales. I see a great deal of variation in main memory/DIMM latencies which are not (often) covered in marketing speak like "DDR-N", but often very impactful on performance.
Usage:
lat [optional-params]
Time latency three ways. shuffle measures real latency.
-k=, --kind= Algo shuff shuff: chase ran perm
ranElt: access ran elt
truRan: pre-read getrandom
-s=, --sizeKiB= int 1048576 set sizeKiB
-n=, --nAcc= int 1000000 set nAcc
-a=, --avgN= int 4 set avgN
-m=, --minN= int 4 set minN
--seed= int 0 0=>random, else set