Improve speed via Stream.resource and faster encoding

[sax]

Collapsing Stream.flat_map + Stream.transform into a single Stream.resource improves the throughput of XML generation by about ~35% in my benchmarks. Optimizing the escaping, encoding, and validation of tags to optimize for ASCII, pass-through of safe input without additional allocation, and use of binary accumulators improves my benchmark by another ~15%.

[ananthakumaran]

Could you rebase with master, I just pushed a fix for the ubuntu 20 runner issue

[sax]

Just rebased.

[sax]

Just in case it's helpful, here is the benchmark I used: https://github.com/synchronal/exceed/blob/main/benchmark/exceed.exs. So my benchmarks include zipping (because of generating Excel files) as well as stream generation.

[ananthakumaran]

1. The first change I do see significant improvement for your benchmark, I am assuming this has to do with how Stream.flat_map works? it suspends too much? I will spend some more time here.

2. I don't see any perf difference when I apply the printer.ex changes (without the flat_map change). The first 2 runs are using the existing printer.ex (master), second 2 runs are from your branch (only printer.ex file change)

[sax]

Sorry for the delay, was just finishing my (US Pacific Time) day when you messaged.

On my M2 Mac Mini I'm seeing the following with Elixir 1.19.5-otp-28 and OTP 28.5:

• HEAD of master: 32k - 34k rows/sec
• Collapsing the two stream operations into one: 46k - 48k rows/sec
• Including the printer.ex changes: 59k - 65k rows/sec

When I switch to Elixir 1.19.5-otp-27 and OTP 27.3.4.11 things do change:

• HEAD of master: 32k - 34k rows/sec
• Collapsing the two stream operations into one: 52k - 54k rows/sec
• Including the printer.ex changes: 52k - 56k rows/sec

I'm seeing this pretty consistently.

[sax]

Oh! I've pulled some overrides of escape_binary into Exceed.Xml, since I was finding performance improvements over the one in XmlStream a while ago. In order to see a real benchmark of XmlStream changes, we need to swap out escape_binary with calling into the library.

[ananthakumaran]

I am OK with merging some of the changes that are mostly straightforward transformations in printer.ex file. But then, there couple of bigger changes that take simple, straightforward code and introduce a bunch of new code, which I am having a very hard time understanding what it does and why the performance increases.

escape_binary, the current code seems to be already binary optimized. Running ERL_COMPILER_OPTIONS=bin_opt_info mix compile --force doesn't point out any issues. I am OK to convert the return value to binary instead of IO data, if it provides any measurable performance. The proposed changes are much more complicated. I am assuming this is some pattern used by other libraries like Jason? Though, I don't understand why and how it improves (if at all)

I do see the changes related to Stream.flat_map results in about 2x performance, but again, what used to be a simple Stream library usage is now converted in something that's much more difficult to read and understand. I can't find any public documentation for :suspended, etc., and it seems like it depends on private/undocumented part of Stream/Enumearable.

[sax]

Ok, I'll see if I can get another version together with just the stream changes, in a way that is more readable. I think the benefit is in the combination of reducing stream operations and utilizing pattern matching in function heads.

I also find it strange how dependent these are on specific Elixir/Erlang versions. I would like to figure out the reason behind that.