~skeeto/public-inbox

1

Time taken to find the best hash functions

Tim McNamara
Details
Message ID
<25fbd369-c280-4d67-9003-053600afa722@www.fastmail.com>
DKIM signature
pass
Download raw message
Chris -

I was wondering how many hours you spent hunting for hash functions to produce the results in your blog post Prospecting Hash Functions. I'm trying to reproduce your results but my scores seem to be nowhere close to yours.

Well done for this work, by the way. It's excellent.

Tim

Re: Time taken to find the best hash functions

Details
Message ID
<20210917140825.vmeq5u5s72qibovx@nullprogram.com>
In-Reply-To
<25fbd369-c280-4d67-9003-053600afa722@www.fastmail.com> (view parent)
DKIM signature
missing
Download raw message
The primary "hash prospector" tool described in my article, with the x86 
JIT compiler, will practically never find anything nearly as good as 
functions listed in the README. Its search is too random. The main result 
was steering me towards multiply-xorshift hashes, essentially re-inventing 
the concept. Run it long enough and it eventually discovers this pattern, 
never reporting anything else thereafter.

All my good results come from my "hillclimb" tool which explores and 
refines the multiply-xorshift parameter space. Notably it doesn't JIT, and 
so I can run it on a Raspberry Pis and such. Just pipe it into tee inside 
a tmux session, and check on it a week later.

I've run "hillclimb" for months on end on various spare computers, 
including a couple of Raspberry Pis. It probably amounts to a decade of 
single-core-years at this point. I'm not running any instances at this 
moment.

> Well done for this work, by the way. It's excellent.

Thanks!
Reply to thread Export thread (mbox)