tor/changes/better_workqueues

11 lines
560 B
Plaintext
Raw Normal View History

o Major features:
- Refactor the CPU worker implementation for better performance by
avoiding the kernel and lengthening pipelines. The original
implementation used sockets to transfer data from the main thread
to the worker threads, and didn't allow any thread to be assigned
more than a single piece of work at once. The new implementation
avoids communications overhead by making requests in shared
memory, avoiding kernel IO where possible, and keeping more
request in flight at once. Resolves issue #9682.