feat: parallelize downloads from TS compiler

[bartlomieju]

Fixes #2626

A small change to the compiler that requests all imports from a file at once instead of doing N round trips. Still needs a bit of cleanup.

master

$ time deno -r fetch https://deno.land/std/testing/runner.ts
Download https://deno.land/std/testing/runner.ts
Compile https://deno.land/std/testing/runner.ts
Download https://deno.land/std/flags/mod.ts
Download https://deno.land/std/fs/mod.ts
Download https://deno.land/std/testing/mod.ts
Download https://deno.land/std/fs/empty_dir.ts
Download https://deno.land/std/fs/ensure_dir.ts
Download https://deno.land/std/fs/ensure_file.ts
Download https://deno.land/std/fs/ensure_link.ts
Download https://deno.land/std/fs/ensure_symlink.ts
Download https://deno.land/std/fs/exists.ts
Download https://deno.land/std/fs/glob.ts
Download https://deno.land/std/fs/globrex.ts
Download https://deno.land/std/fs/move.ts
Download https://deno.land/std/fs/copy.ts
Download https://deno.land/std/fs/read_file_str.ts
Download https://deno.land/std/fs/write_file_str.ts
Download https://deno.land/std/fs/read_json.ts
Download https://deno.land/std/fs/write_json.ts
Download https://deno.land/std/fs/walk.ts
Download https://deno.land/std/fs/eol.ts
Download https://deno.land/std/fs/utils.ts
Download https://deno.land/std/fs/path/mod.ts
Download https://deno.land/std/fs/path/win32.ts
Download https://deno.land/std/fs/path/posix.ts
Download https://deno.land/std/fs/path/constants.ts
Download https://deno.land/std/fs/path/interface.ts
Download https://deno.land/std/fs/path/utils.ts
Download https://deno.land/std/testing/asserts.ts
Download https://deno.land/std/fmt/colors.ts
Download https://deno.land/std/testing/diff.ts
Download https://deno.land/std/testing/format.ts

real	0m4.273s
user	0m2.615s
sys	0m0.310s

this PR

$ time target/release/deno -r fetch https://deno.land/std/testing/runner.ts
Download https://deno.land/std/testing/runner.ts
Compile https://deno.land/std/testing/runner.ts
Download https://deno.land/std/flags/mod.ts
Download https://deno.land/std/fs/mod.ts
Download https://deno.land/std/testing/mod.ts
Download https://deno.land/std/fs/empty_dir.ts
Download https://deno.land/std/fs/ensure_dir.ts
Download https://deno.land/std/fs/ensure_file.ts
Download https://deno.land/std/fs/ensure_link.ts
Download https://deno.land/std/fs/ensure_symlink.ts
Download https://deno.land/std/fs/exists.ts
Download https://deno.land/std/fs/glob.ts
Download https://deno.land/std/fs/globrex.ts
Download https://deno.land/std/fs/move.ts
Download https://deno.land/std/fs/copy.ts
Download https://deno.land/std/fs/read_file_str.ts
Download https://deno.land/std/fs/write_file_str.ts
Download https://deno.land/std/fs/read_json.ts
Download https://deno.land/std/fs/write_json.ts
Download https://deno.land/std/fs/walk.ts
Download https://deno.land/std/fs/eol.ts
Download https://deno.land/std/fs/utils.ts
Download https://deno.land/std/fs/path/mod.ts
Download https://deno.land/std/fs/path/win32.ts
Download https://deno.land/std/fs/path/posix.ts
Download https://deno.land/std/fs/path/constants.ts
Download https://deno.land/std/fs/path/interface.ts
Download https://deno.land/std/fs/path/utils.ts
Download https://deno.land/std/testing/asserts.ts
Download https://deno.land/std/fmt/colors.ts
Download https://deno.land/std/testing/diff.ts
Download https://deno.land/std/testing/format.ts

real	0m2.259s
user	0m2.494s
sys	0m0.250s

CC @ry @kitsonk

[bartlomieju]

I must admit, doing this PR with JSON ops is a breeze... I couldn't figure it out with Flatbufs a couple times.

[ry]

👍 elegant

(at some point we need to get rid of this block_on stuff tho...)

[ry]

LGTM - no comments

[ry]

This caused the thread count to jump from 11 to 77 threads in cold_relative_import:

This is due to parallel usage of block_on... :(

[bartlomieju]

This is bad news. We really need to figure out what to do with block_on. Wouldn't small increase of default number of threads solve this issue?

[ry]

@bartlomieju The issue is that we're starting a whole new tokio runtime for each module resolved - it is wildly inefficient.

Wouldn't small increase of default number of threads solve this issue?

No

[ry]

@bartlomieju After some thought, I will not revert your patch. Your patch speeds up an important (unfortunately unmeasured) benchmark. I think what you're doing is more or less correct and just exposing a lower-level problem. I've added #2960 to 1.0 blockers to solve this most important design problem.

[bartlomieju]

@bartlomieju After some thought, I will not revert your patch. Your patch speeds up an important (unfortunately unmeasured) benchmark. I think what you're doing is more or less correct and just exposing a lower-level problem. I've added #2960 to 1.0 blockers to solve this most important design problem.

@ry alright, I will investigate alternate solutions as well