World: r3wp

R3 doesn't have anything like R2's system/user.


I don't think, anybody mentioned R2's system/user. Do you mean system/words 
?

Sorry, yes, that's what I meant.

Though technically all R3 objects are more like system/words than 
they are like R2's objects.

R3 has system/contexts/user , which seem to work like R2's system/words. 
Try

? system/contexts/user


before and after making some random words, e.g. in a block like [some 
random words]

No, it really doesn't. All loaded words are added to system/words. 
Only words that are referenced directly in user scripts, or added 
explicitly with INTERN, are added to system/contexts/user. I had 
to add a bit of really careful code to make sure that system/contexts/user 
doesn't get words added to it until absolutely necessary, because 
it is an isolated context from the runtime library system/contexts/lib.

Symbols in R3 are stored in an internal symbols table, a btree(?) 
or some other unknown data structure that you can't reference externally.

Is the internal symbol table there to save memory? Maybe like Lua's 
internal structure to hold strings?

The internal symbol table is there to make symbols work at all. In 
R2, system/words was the symbol table. However, it does save memory 
relative to strings because there are no duplicates, and because 
the symbol data for the words is stored in UTF8 instead of 16-bit 
characters.

They aren't added to any context until you add them explicitly. The 
R3 interpreter would not presume to know what a word would mean to 
you until you tell it what it means, by binding the word to a context 
or by just using the word as data.

bbl

>> repeat i to-integer #7fffffffffffffff [to word! ajoin ["a" i]]
** Internal error: not enough memory
** Where: to repeat
** Near: to word! ajoin ["a" i]

at ~1700MB resident

and it took 2h50m cpu time to get there :)

:) Cool test!

This code might be a better test: repeat i to-integer #7fffffffffffffff 
[if zero? i // 1'000'000 [recycle] to-hex i]

It should have less memory usage overall and if words are recycled 
then it won't run out. I'll run it now.

Where are the words coming into the picture?

TO-HEX generates an issue!, which is a word type in R3. Yes, you 
can even bind them.

Spooky! :)

I figure that not creating the temporary string, and running recycle 
every once in a while, might make the memory problems go away. So 
far the R3 process is staying at exactly the same memory, less than 
a gig. I also tossed an assignment in there so I can know which number 
it fails on.

the jump from ~900M to ~1.2G took ages

then another aeon fuer 1.2G to 1.4G

and hours for 1.4G to 1.7G and fail

(jfyi)

The time on mine won't be comparable because it's only running on 
one of 4 cores, and the others will be mildly occupied.

R3 won't utilise more than 1 core either way :)

Of course :)

Ooo, tasks! But error handling is still broken in tasks, so this 
test won't work :(

Ah, here's the jump :)

From 1.0G to 1.9G, after 14m cpu time

But it's still running

Mine just jumped up a hundred megs, still running :)

It failed a lot earlier for me (some time in the last half hour), 
never getting over 1.2GB, and now there isn't enough memory to figure 
out what the number it failed at was. Looks like you run out of memory 
before you run into the word limit, and that the symbol table isn't 
cleaned up by the recycler. Good to know.

these experiments remind me how painful was it to figure out why 
are we getting those funky out of memory messages, saying something 
about PermGen in java

The internal symbol table is there to make symbols work at all.


I don't think, I understand this fully. To me, it's like needing 
an internal string table to make strings work. Or an internal integer 
table to make integers work. Why not just have contexts, and don't 
put words in any context, if the word is just data? And if a word 
go from being data to hold more meaning (have a value attached), 
then put it in a context.

If you take a book written in finnish, you see a lot of words, but 
they have no meaning to you. When you close the book, the finnish 
words shouldn't take up any space in your brain.

Geomol, I think you are confusing design with implementation. The 
implementation can be improved; however, it's a compromise between 
the complexity of it and how common your "millions of words" scenario 
is in practice.

Makes sense.

Putting words in a context

 is binding. That's very different from the symbol table, which you 
 could say "binds" symbol strings to integer IDs

If you can't read Finnish, it means the Finnish symbols are not bound 
to values in contexts in your head. Closing the book and erasing 
the symbol table is equivalent to quitting the REBOL proces, which 
does release the memory

I'm sure some Finnish words are also words in Danish, so forgetting 
the symbol table wouldn't help your situation :-)

I'm sure because my name is one of them...

I prefer Finnish Vodka, and Finnish music :-)

John, you are missing some things others know and find obvious. For 
example, do you know the answer to the following question?

    What is the ratio between

        stringn: func [n] [head insert/dup copy "" #"a" n]


        word: to word! stringn 1 t1: time-block [equal? word word] 0,05

        word: to word! stringn 1000 t2: time-block [equal? word word] 0,05
        t2 / t1

compared to


        string: stringn 1 t1: time-block [equal? string string] 0,05

        string: stringn 1000 t2: time-block [equal? string string] 0,05
        t2 / t1
?

time-block? u mean delta-time?

(your naming is more intuitive though... it's also 'time in bash...)

Thanks, Ladislav. Good examples!


The thing, I missed, was that REBOL has this extra internal data 
structure to hold symbols (words), I though, just the contexts was 
used for that. So comparing words (that are not bound to any context) 
are much faster than comparing strings in REBOL. I see different 
possibilities, depending on implementation. If a word is changed 
to the result from a hash calculation, then two different words might 
give the same result, right? It's unlikely, but it could happen. 
That's why map datastructures are combined with lists, when two different 
hash calculations give same result. The other possibility is, that 
words are changed to pointers pointing to their entry in the map 
(hash table). Do you know, what of the two, REBOL implement? In other 
words, could two different words be equal in REBOL?


About the strings, then it's possible to do kind of the same with 
hashing. Lua does that. If you have two different but identical strings 
(same string content) in Lua, they share the same memory area. Hashing 
is involved, and I guess, comparing string would be equal fast as 
comparing words, if REBOL did the same. (unless words are exchanged 
with the result from the hash calculation, but then two words might 
be equal)

After a little testing, it seems, words are changed to a hash result, 
and REBOL gives an error, if two different words give same result:

>> s: "abcdefghijklmn"

>> forever [if equal? a: to word! random s b: to word! random s [print 
[a b]]]
** Internal Error: No more global variable space

The error comes really fast.

Petr, Nightwish? :-)

Above test is in R2.

R3 keeps going and can't be stopped with <Esc>.