Issue #16375 has been updated by mame (Yusuke Endoh).


methodmissing (Lourens Naud=E9) wrote:
> Thanks for the consideration and discussing the proposal. Do you mean fil=
e a refactored version of this ticket which has drifted in scope on the Fea=
ture tracker with the memory improvements?

Sorry, I meant please choose "Feature" (not "Misc") as a tracker if you reg=
ister a ticket about improvements from the next time.

I've already moved this ticket to Feature tracker.  You have to do nothing =
about this ticket unless nobu requests.

----------------------------------------
Feature #16375: Right size regular expression compile buffers for literal r=
egexes and on Regexp#freeze
https://bugs.ruby-lang.org/issues/16375#change-83413

* Author: methodmissing (Lourens Naud=E9)
* Status: Open
* Priority: Normal
* Assignee: nobu (Nobuyoshi Nakada)
* Target version: 2.8
----------------------------------------
References PR https://github.com/ruby/ruby/pull/2696

As a continuation of [type specific resize on freeze implementations of Str=
ing and Array](https://bugs.ruby-lang.org/issues/16291) and looking into th=
e `Regexp` type I found these memory access patterns for regular expression=
 literals:

```
=3D=3D22079=3D=3D -------------------- 12 of 500 --------------------
=3D=3D22079=3D=3D max-live:    1,946,560 in 4,345 blocks
=3D=3D22079=3D=3D tot-alloc:   1,946,560 in 4,345 blocks (avg size 448.00)
=3D=3D22079=3D=3D deaths:      none (none of these blocks were freed)
=3D=3D22079=3D=3D acc-ratios:  1.36 rd, 0.98 wr  (2,651,994 b-read, 1,908,1=
58 b-written)
=3D=3D22079=3D=3D    at 0x4C2DECF: malloc (in /usr/lib/valgrind/vgpreload_e=
xp-dhat-amd64-linux.so)
=3D=3D22079=3D=3D    by 0x24C496: onig_new_with_source (re.c:844)
=3D=3D22079=3D=3D    by 0x24C496: make_regexp (re.c:874)
=3D=3D22079=3D=3D    by 0x24C496: rb_reg_initialize (re.c:2858)
=3D=3D22079=3D=3D    by 0x24C496: rb_reg_initialize_str (re.c:2892)
=3D=3D22079=3D=3D    by 0x24C496: rb_reg_compile (re.c:2982)
=3D=3D22079=3D=3D    by 0x12EB84: rb_parser_reg_compile (parse.y:12185)
=3D=3D22079=3D=3D    by 0x12EB84: parser_reg_compile (parse.y:12179)
=3D=3D22079=3D=3D    by 0x12EB84: reg_compile (parse.y:12195)
=3D=3D22079=3D=3D    by 0x2147E3: new_regexp (parse.y:10101)
=3D=3D22079=3D=3D    by 0x2147E3: ruby_yyparse (parse.y:4419)
=3D=3D22079=3D=3D    by 0x2161F7: yycompile0 (parse.y:5942)
=3D=3D22079=3D=3D    by 0x3241FF: rb_suppress_tracing (vm_trace.c:427)
=3D=3D22079=3D=3D    by 0x1FDBF6: yycompile (parse.y:5991)
=3D=3D22079=3D=3D    by 0x1FDBF6: rb_parser_compile_file_path (parse.y:6130)
=3D=3D22079=3D=3D    by 0x27AC96: load_file_internal (ruby.c:2034)
=3D=3D22079=3D=3D    by 0x137730: rb_ensure (eval.c:1129)
=3D=3D22079=3D=3D    by 0x27CEEA: load_file (ruby.c:2153)
=3D=3D22079=3D=3D    by 0x27CEEA: rb_parser_load_file (ruby.c:2175)
=3D=3D22079=3D=3D    by 0x1954CE: load_iseq_eval (load.c:587)
=3D=3D22079=3D=3D    by 0x1954CE: rb_load_internal (load.c:651)
=3D=3D22079=3D=3D    by 0x1954CE: rb_f_load (load.c:709)
=3D=3D22079=3D=3D    by 0x2FB957: vm_call_cfunc_with_frame (vm_insnhelper.c=
:2468)
=3D=3D22079=3D=3D    by 0x2FB957: vm_call_cfunc (vm_insnhelper.c:2493)
```

Digging a little further and remembering some context of previous oniguruma=
 memory investigation I remembered the pattern buffer struct has a compile =
buffer with a simple watermark for tracking used space. This changeset impl=
ements `reg_resize` (static as `ary_resize`) which attempts to right size t=
he compile buffer if over allocated at the following sites:

* After compiling a literal regular expression.
* Implement an explicit type specific `rb_reg_freeze` and point `Regexp#com=
pile` to it
* I also follow the `chain` member which points to another `regex_t` on the=
 struct if present, but have not been able to find references to it in the =
source tree other than for freeing a regex or inspecting it's memory footpr=
int.

I introduced 2 new debug counters, which yields the following results on bo=
oting Redmine on Rails 5:

```
[RUBY_DEBUG_COUNTER]    obj_regexp_lit_extracapa                    6319
[RUBY_DEBUG_COUNTER]    obj_regexp_lit_extracapa_bytes            301685
```

About 300kb reallocated across 6319 oversized instances.

An example of `Regexp#freeze`

```
irb(main):007:0> r =3D Regexp.compile("(?!%\h\h|[!$-&(-;=3D?-_a-~]).")
irb(main):008:0> ObjectSpace.memsize_of(r)
=3D> 588
irb(main):009:0> r.freeze
=3D> /(?!%hh|[!$-&(-;=3D?-_a-~])./
irb(main):010:0> ObjectSpace.memsize_of(r)
=3D> 543
```

There is likely more layers that can be peeled back here, but keeping it si=
mple and concise for review.

I think it's possible to get towards a state where 5 to 10MB RSS can be sha=
ved off a standard Rails process by:

* Being careful with buffer defaults for objects that have auxiliary buffers
* Identifying hooks where excess allocation can be reduced to current water=
mark if the object buffer does not need to grow anymore (literal object, fr=
ozen object etc.)
* A GC hook on promotion to OLD object to trim excessive capacity, which I'=
d consider as a special kind of garbage, as outlined in https://bugs.ruby-l=
ang.org/issues/15402

Would any further incremental work towards this be considered worthwhile fr=
om a ruby-core and community perspective?



-- =

https://bugs.ruby-lang.org/

Unsubscribe: <mailto:ruby-core-request / ruby-lang.org?subject=3Dunsubscribe>
<http://lists.ruby-lang.org/cgi-bin/mailman/options/ruby-core>