Optimize dumpCoroutinesInfoAsJsonAndReferences

[mvicsokolova]

I've removed trimIndent and toStringRepr calls, made during the composition of coroutine dump json.

The motivation: Intellij IDEA Debugger supports the Coroutine View, which can show the hierarchy of all the available coroutines. This View struggled of sever performance issues (IDEA-335303). There were reasons for that on the debugger side, like fetching additional information for each coroutine. But other than that, the invocation of dumpCoroutinesInfoAsJsonAndReferences took significant amount of time.

I made some dumb performance measurements on the dumps of ~5000 coroutines, the absolute time is irrelevant ofc, just % of time of the parent call are:

1. Original version, most of time is taken by trimIndent:

2. Without trimIndent invocation:

3. Without toStringRepr invocation as well, it finally makes sense and shows dumpCoroutineInfos as the next significant call:

It's not necessary to call trimIndent on every piece of this JSON, so that it could be parsed on the debugger side. So, I suggest to optimize this code.

[fzhinkin]

@mvicsokolova, is there a way to improve trimIndent performance? It might be fine to abstain from calling it here, but I believe that it will be beneficial for all other projects relying on that function.

[mvicsokolova]

Yes, I'll create a ticket for trimIndent optimization in kotlin-stdlib 👍🏼

[mvicsokolova]

I'm not sure for now though, what's the better way to deal with special symbols in the coroutine name, than the current toStringRepr.

[dkhalanskyjb]

Yes, repr() is needed to ensure that JSON is valid. Not all hope is lost, though. I'm not a performance optimization expert, but there are a few ideas I'd try.

• Usually, there are very few dispatchers in the program. This presents an optimization opportunity. If you cache the string representations of dispatchers in an IdentityHashMap<CoroutineDispatcher, String>, the number of repr() calls should drastically decrease (at least by about a half for a large number of named coroutines, but realistically, even more).
• If this doesn't help, maybe you could first check if the string contains any special symbols (99% of them won't), and if not, skip char-by-char string copying entirely.
• Do we need a pretty JSON at all? If we're not trimming indents, why not write the JSON string in a single line? The size of the resulting string would be literally half of what it is.

[dkhalanskyjb]

If we are going to merge this, at least trimIndent's functionality should be recreated by deindenting the string (with an additional comment explaining that the abscence of trimIndent is intentional); otherwise, we get redundant whitespace in the result.