This change simplifies how we execute V8. Previously V8 Isolates jumped
around threads every time they were woken up. This was overly complex and
potentially hurting performance in a myriad ways. Now isolates run on
their own dedicated thread and never move.
- blocking_json spawns a thread and does not use a thread pool
- op_host_poll_worker and op_host_resume_worker are non-operational
- removes Worker::get_message and Worker::post_message
- ThreadSafeState::workers table contains WorkerChannel entries instead
of actual Worker instances.
- MainWorker and CompilerWorker are no longer Futures.
- The multi-threaded version of deno_core_http_bench was removed.
- AyncOps no longer need to be Send + Sync
This PR is very large and several tests were disabled to speed
integration:
- installer_test_local_module_run
- installer_test_remote_module_run
- _015_duplicate_parallel_import
- _026_workers
After landing #3358 the benchmarks exploded indicating problems with workers and deno_core_http_bench.
This PR dramatically fixes thread/syscall count that showed up on benchmarks. Thread count is not back to previous levels but difference went from hundreds/thousands to about ~50.
Towards simplifying (or better removing entirely) the CoreResource
trait. Resources should be any bit of privileged heap allocated memory
that needs to be referenced from JS, not very specific trait
implementations. Therefore CoreResource should be pushed towards being
as general as possible.
It's unnecessary indirection and is preventing the ability to easily
pass isolate references into the dispatch and dyn_import closures.
Note: this changes how StartupData::Script is executed. It's no longer done
during Isolate::new() but rather lazily on first poll or execution.
* In order to prevent ArrayBuffers from getting garbage collected by V8,
we used to store a v8::Persistent<ArrayBuffer> in a map. This patch
introduces a custom ArrayBuffer allocator which doesn't use Persistent
handles, but instead stores a pointer to the actual ArrayBuffer data
alongside with a reference count. Since creating Persistent handles
has quite a bit of overhead, this change significantly increases
performance. Various HTTP server benchmarks report about 5-10% more
requests per second than before.
* Previously the Persistent handle that prevented garbage collection had
to be released manually, and this wasn't always done, which was
causing memory leaks. This has been resolved by introducing a new
`PinnedBuf` type in both Rust and C++ that automatically re-enables
garbage collection when it goes out of scope.
* Zero-copy buffers are now correctly wrapped in an Option if there is a
possibility that they're not present. This clears up a correctness
issue where we were creating zero-length slices from a null pointer,
which is against the rules.
Op dispatch is now dynamically dispatched, so slightly less efficient.
The immeasurable perf hit is a reasonable trade for the API simplicity
that is gained here.
