Optimize String#to_json #5456

larubujo · 2017-12-26T04:22:26Z

saw this: https://stackoverflow.com/questions/47952457/crystal-slow-json-serialization-of-structs-containing-large-strings

this pr still slower then go (crystal memory allocator slower, probably gc slower), but faster then before

old code: go char by char and append it
new code: go char by char, if escape then append all before that point, then escaped char

best case: string has no escape, append all in one go
worse case: all chars escapes, similar to old code

bench:

require "benchmark"
require "json"

record Page, uri : String, html : String do
  JSON.mapping(uri: String, html: String)
end

url = "not an uri has escape like \" \\ and also \n "
html = "a" * (128 * 1024)

page = Page.new(url, html)
io = IO::Memory.new

Benchmark.ips do |x|
  x.report("json") do
    page.to_json(io)
    io.clear
  end
end

old:

json 766.91  (   1.3ms) (± 3.00%) fastest

new:

json    1.7k (587.83µs) (± 8.62%) fastest

straight-shoota · 2017-12-26T11:25:44Z

src/json/builder.cr

+      pos = 0
+      reader = Char::Reader.new(string)
+
+      while reader.has_next?


You could use while char = reader.next_char (obviously modified to include the first char), eliminating an additional method call.

no. next_char raises on end

This can easily be handled with when Char::ZERO; break and the reader will never get to a point where it raises. This solution should be faster than calling has_next? and current_char on each iteration.

string can have zero byte in it. so condition not always work.

True, that's a valid character in a JS string.
Still, you could check has_next? only if char == Char::ZERO instead of every iteration.

search while reader.has_next? in std. 7 occurrences. nothing wrong with this code. plus i dont think its slower than your approach. this is idiomatic code.

straight-shoota · 2017-12-26T11:27:20Z

src/json/builder.cr

        when '\\'
-          io << "\\\\"
+          escape = "\\\\"


escape could just be a Char (the one being escaped).

io << '\\' << escaped_char. It would be more efficient than creating and inserting a new string

Strings are allocated on the heap, chars not. Here it is very easy to avoid heap allocations and use chars instead.

Constant strings are not allocated on the heap. They're pre-built in read only ELF data, so doing foo = "bar" is simply the cost of assigning an address to a variable.

True, forgot about that.

Still, using a Char should be slightly preferrable. It's just less data to store: no size field + backslash is only needed once for all escape sequences.

escape cant be char. for \b want "\b", not '\' and '\b'. escape must be string. or it must be 'b' in case of '\b', but then two writes, slower than one write. but also, nanosecond time. not relevant.

please send a different pr where performance is beat, then i can accept changes. otherwise this pr is good.

straight-shoota · 2017-12-26T11:29:14Z

src/json/builder.cr

-      string.each_char do |char|
-        case char
+
+      pos = 0


Please use a more descriptive name like last_start_pos.

too long... maybe start or start_pos... theres only one pos to track, why last? theres no first. theres no middle. algorithm is simple.

straight-shoota · 2017-12-26T11:33:18Z

src/json/builder.cr

        end
+
+        if escape


I'd suggest to put this behaviour directly in the when branches (maybe extract in a private method). This adds some code duplication but eliminates an additional condition check.

Or just skip to the next char with next from the .ascii_control? and else branches. Then you can keep it here but don't need the condition.

put directly in when: longer code.

i understand your concern, but these little details not improve performance, and make code longer, same to understand. so i prefer not change.

straight-shoota · 2017-12-26T11:34:06Z

src/json/builder.cr

          io << "\\u"
          ord = char.ord
          io << '0' if ord < 0x1000
          io << '0' if ord < 0x100
          io << '0' if ord < 0x10
          ord.to_s(16, io)
        else
-          io << char
+          set_pos = false


You could just insert next here, eliminating the need for a variable set_pos.

maybe, ends up being same longer

The code might be a bit longer, but I expect execution time to be lower.

i changed code, same time. so stay the same for me.

I would like to remove set_pos personally... it's not a big deal for me but the logic is slightly harder to follow. I want to see what the rest of the core team thinks.

larubujo · 2017-12-26T12:29:02Z

@straight-shoota sorry, but you not core team member, so why do i need apply suggestions?

better close this. you want to implement this, you do it.

RX14 · 2017-12-26T12:33:35Z

@larubujo all reviews are just suggestions, if you disagree with them feel free to not implement them, but I don't see why you'd close this PR.

straight-shoota · 2017-12-26T15:01:51Z

Yeah, you don't have to follow my suggestions. I was just giving you hints how I think you could improve your code. If you don't want to apply them, then don't. But I think they're reasonable to improve efficiency even further.

larubujo · 2017-12-29T20:06:08Z

i reopen. i can rename pos to start_pos, but other suggestions not apply for me.

RX14

This is fine to merge if the rest of the core team agrees.

We might need to change the scope of the escaping in the future to escape more characters but that's outside the scope of this PR.

larubujo · 2017-12-29T22:32:02Z

refactored. let me know what you think.

RX14 · 2017-12-29T22:40:20Z

I think I prefer the old version, it has less duplication and less to go wrong. Replacing the conditionals with next results in code which is harder to maintain.

larubujo · 2017-12-29T23:45:59Z

boring...

RX14 · 2017-12-30T00:22:35Z

@larubujo I meant your old version that you originally PRed, not the version before this PR! The version before you said "refactored. let me know what you think." Sorry for the misunderstanding.

larubujo · 2017-12-30T14:33:54Z

problem is, i lost that code. cant remember now. plus review is "do this, no, do that, revert this now". really boring and frustrating. better someone continue this, i got tired.

larubujo · 2017-12-31T00:44:18Z

reopen second time. please accept it. also note this is faster than original code, so better:

json   1.82k ( 550.2µs) (± 2.41%) fastest

slightly faster. but maybe moon closer to earth these days, who knows.

straight-shoota · 2017-12-31T11:01:30Z

I think it is good now.

However, it could improve the handling of special characters besides performance improvements. But this could be done in a subsequent PR. Namely, it misses named escapes for \0 and \v - they should still be escaped as \u0000 and \u000B but using the named escapes is more elegant and saves a little bitte of space. .ascii_control? should better be .control?. And it should fail for invalid unicode characters. Crystal strings can currently contain invalid UTF-8, JavaScript/JSON cannot.

bcardiff

I like the optimization, but since the flow is more complicated than linear iteration of chars I would prefer to have some specs to ensure corner cases of the code won't fail in the future due to further changes. Like: if the string starts with sequences to escape, if it has more than one sequences to escape together, if there is trailing sequences to escape. Those kind of cases might work right now, but some slightly changes could end in chars been lost.

larubujo · 2018-01-02T13:23:19Z

@bcardiff added more specs

RX14 · 2018-01-02T13:34:55Z

spec/std/json/serialization_spec.cr

@@ -215,6 +215,19 @@ describe "JSON serialization" do
      "\u{19}".to_json.should eq("\"\\u0019\"")
    end

+    it "does for String with control codes in a few places" do
+      "\fab".to_json.should eq("\"\\fab\"")


It's probably much more readable to use %q("\fab") here instead of "\"\\fab\"".

straight-shoota requested changes Dec 26, 2017

View reviewed changes

larubujo closed this Dec 26, 2017

larubujo reopened this Dec 29, 2017

RX14 approved these changes Dec 29, 2017

View reviewed changes

larubujo force-pushed the opt-string-to-json branch from c60de57 to 577d099 Compare December 29, 2017 22:31

larubujo force-pushed the opt-string-to-json branch from 577d099 to 5eb00cf Compare December 29, 2017 22:32

larubujo closed this Dec 29, 2017

larubujo reopened this Dec 31, 2017

straight-shoota approved these changes Dec 31, 2017

View reviewed changes

RX14 approved these changes Dec 31, 2017

View reviewed changes

bcardiff requested changes Dec 31, 2017

View reviewed changes

larubujo force-pushed the opt-string-to-json branch from 5eb00cf to 22a66cc Compare January 2, 2018 13:22

RX14 reviewed Jan 2, 2018

View reviewed changes

Optimize String#to_json

60bcc37

larubujo force-pushed the opt-string-to-json branch from 22a66cc to 60bcc37 Compare January 2, 2018 14:09

RX14 added kind:refactor performance topic:stdlib labels Jan 2, 2018

RX14 approved these changes Jan 2, 2018

View reviewed changes

asterite approved these changes Jan 2, 2018

View reviewed changes

RX14 added this to the Next milestone Jan 2, 2018

RX14 merged commit be9d9c7 into crystal-lang:master Jan 2, 2018

lukeasrodgers pushed a commit to lukeasrodgers/crystal that referenced this pull request Jan 7, 2018

Optimize String#to_json (crystal-lang#5456)

c1c0cb0

Optimize String#to_json #5456

Optimize String#to_json #5456

Conversation

larubujo commented Dec 26, 2017

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

RX14 Dec 29, 2017 • edited Loading

Choose a reason for hiding this comment

larubujo commented Dec 26, 2017

RX14 commented Dec 26, 2017

straight-shoota commented Dec 26, 2017

larubujo commented Dec 29, 2017

RX14 left a comment

Choose a reason for hiding this comment

larubujo commented Dec 29, 2017

RX14 commented Dec 29, 2017

larubujo commented Dec 29, 2017

RX14 commented Dec 30, 2017 • edited Loading

larubujo commented Dec 30, 2017

larubujo commented Dec 31, 2017

straight-shoota commented Dec 31, 2017

bcardiff left a comment

Choose a reason for hiding this comment

larubujo commented Jan 2, 2018

Choose a reason for hiding this comment

Choose a reason for hiding this comment

RX14 Dec 29, 2017 •

edited

Loading

RX14 commented Dec 30, 2017 •

edited

Loading