From: Ryan Davis <ryand-ruby@...>
Date: 2012-10-12T07:12:10+09:00
Subject: Re: [ANN] google_hash v 0.7.0 released. Faster hashes for MRI.


On Oct 11, 2012, at 14:40 , Eric Wong <normalperson@yhbt.net> wrote:

> Roger Pack <rogerdpack2@gmail.com> wrote:
>>> From the readme:
>>> 
>>>> These also use significantly less memory, because (if you specify IntToInt, it stores only 4 bytes per int, instead of Ruby's usual 20 bytes).  This also frees up Ruby so it doesn't hvae to garbage collect as much.  Yea!
>>> 
>>> 20 bytes?? What exactly is this referring to?
> 
> I thought this was referring to st_table_entry size in st.c, but that's
> 6 words (24 bytes in 32-bit, 48 bytes in 64-bit) on MRI 1.9 with ordered
> hashes (unpacked).
> 
>> Normal size (or at least used to be) of a "ruby object" in 32 bit MRI.
>> I believe it's RObject:
>> http://fossies.org/dox/ruby-1.9.3-p194/structRObject.html
>> The size may be bigger in 64 bit MRI, I'm not sure.
> 
> RObject is 40 bytes on 64-bit MRI.  On the plus side with 64-bit,
> embedded strings can be up to 23 bytes (vs 11 bytes for 32-bit) so
> there's a better chance of avoiding malloc() overhead with strings.

But we're talking ints here... so they just take up the space of VALUE. I don't know why that's being compared against RObject.