wip: crypto: use cppgc to manage Hash #51017
Conversation
|
Review requested:
|
nodejs-github-bot
added
c++
|
|
||
| namespace node { | ||
|
|
||
| #define ASSIGN_OR_RETURN_UNWRAP_CPPGC(ptr, obj, ...) \ |
There was a problem hiding this comment.
Would you mind adding a small documentation/comment around what cppgc is, some important links etc to this document?
There was a problem hiding this comment.
Some documentation should probably go into src/README.md.
src/cppgc_helpers.h
Outdated
| Environment* env_; \ | ||
| v8::TracedReference<v8::Object> traced_reference_; | ||
|
|
||
| class CppgcMixin { |
There was a problem hiding this comment.
Can you add a comment for the purpose of this class?
| @@ -103,7 +109,8 @@ void Hash::New(const FunctionCallbackInfo<Value>& args) { | |||
| xof_md_len = Just<unsigned int>(args[1].As<Uint32>()->Value()); | |||
| } | |||
|
|
|||
| Hash* hash = new Hash(env, args.This()); | |||
| Hash* hash = cppgc::MakeGarbageCollected<Hash>( | |||
There was a problem hiding this comment.
Would you mind adding a comment to here why we don't call new Hash but prefer the recommended change to here.
| ASSIGN_OR_RETURN_UNWRAP(&ctx, args.Holder()); | ||
| } else { | ||
| ctx = CppgcMixin::Unwrap<T>(args.Holder()); | ||
| if (ctx == nullptr) return; |
There was a problem hiding this comment.
In what scenario can ctx be nullptr?
I am mostly opening it to run the benchmark CI. |
|
@joyeecheung I ran local benchmarks earlier today (using a commit you had added) and it did solve the performance issues for me. However, I did not investigate further and I wasn’t very careful. However I got interested in your commit because it is attacking a performance bottleneck. |
|
Results from the CI https://ci.nodejs.org/view/Node.js%20benchmark/job/benchmark-node-micro-benchmarks/1476/console And https://ci.nodejs.org/view/Node.js%20benchmark/job/benchmark-node-micro-benchmarks/1475/console - looks like n is too small for the CI machine and needs to be increased. It seems interesting how in the CI machine the digest benchmarks get a more stable speedup (though we also know that this largely depends on the CPU and how big of a bottleneck OpenSSL is on that CPU). |
|
Updated a bit to use cppgc::GarbageCollectedMixin and cppgc::NameProvider. Startup snapshot integration is still unsolved but I don't think it's needed for Hash. Though it would be nice to be able to customize the memory tracking so we can track OpenSSL memory in the heap snapshot. Still looking into it. Somehow numbers are even better now on macOS + M2 Max |
|
From discussions in https://docs.google.com/document/d/1ny2Qz_EsUnXGKJRGxoA-FXIE2xpLgaMAN6jD7eAkqFQ/edit looks like we need to upstream an API to V8 to track external memory in the heap snapshot. Looking into an API - it probably wouldn't be too different from what we already have in |
|
Started working on the external memory tracking API: https://chromium-review.googlesource.com/c/v8/v8/+/5630497 it may need to take care of the GC tuning part too, from the discussions in the doc |
Original commit message:
[cppgc] add cppgc::Visitor::TraceExternal()
This allows tracking externally-managed memory owned by cppgc-managed
objects in the heap snapshots.
To accompany this API, a class cppgc::External is added to abstract
over externally-managed structures.
Design doc: https://docs.google.com/document/d/1-kHbj9SNL3wMXZz_GZiDw6kHUJ1IvicFzEmqC_um0Wg/edit#heading=h.n1atlriavj6v
Change-Id: I7a27eaa45e29d3c14936997d115a112a6e54458c
Refs: v8/v8@fa7509f
joyeecheung
force-pushed
the
cppgc-hash
branch
from
86862d9
to
0067a0b
Compare
|
Rebased after v8 12.8 landed on main, and used updated version of https://chromium-review.googlesource.com/c/v8/v8/+/5630497 as well as the helpers in #52295 - the other PR will be ready after I finish the docs, whereas this PR depends on the mentioned V8 CL that needs to be hashed out (especially the GC scheduling integration part) since unlike ContextifyScript, crypto::Hash does have external memory, so this won't be unblocked until the V8 CL lands. With the WIP V8 CL the heap snapshot after migration looks like this: 
(Currently cppgc-managed objects cannot specify edge names in the heap snapshot, that is being worked on in https://docs.google.com/document/d/1PQQHhT0MLlStoiqNmji2-GcX62xtAsXPrihXi403ib4/edit#heading=h.n1atlriavj6v) |
While working on #40786 I happened to notice that the current BaseObject management comes with a significant overhead from global handle creation and incidentally migrating these objects to Oilpan gives us faster creation of objects (in my local testing, creating a cppgc-managed object is about 2.5x faster than creating a weak BaseObject).
This is only a WIP and not yet ready for review. It's mostly open as a reference for nodejs/performance#136 and for running benchmark CI. The primary blocker is that I'll need to figure out how to support embedder object book-keeping in the heap snapshot with cppgc (right now it doesn't work and embedder objects become missing in the heap snapshots) - EDIT: this is currently being worked on in https://chromium-review.googlesource.com/c/v8/v8/+/5630497
Refs: nodejs/performance#136
Refs: #40786
On macOS + M2 where OpenSSL 3 performs much better
On Ubuntu + Intel(R) Xeon(R) Platinum 8280 CPU @ 2.70GHz, where OpenSSL 3 does not perform as well and becomes the bottleneck in digest computation: