Motivation:

Right now we store our selectable event loop reference in a NIO ThreadSpecificVariable. That's nice enough, but this has become a hot code path in cases where users are using the complete concurrency override. In those cases, the ThreadSpecificVariable forces us to both call pthread_getspecific and to end up needing to perform a dynamic cast to our expected type.

These dynamic casts are costly: in one sample project where we spent fully 7.84% of the runtime in dispatching tasks onto the EL, the cost of that dynamic cast alone was 0.5% of the runtime, with pthread_getspecific being another 0.2%. This is 10% of the cost of dispatching a job, entirely unnecessarily.

Modifications:

• Use some fancy C thread-locals, which have much better performance: see improve inEventLoop checks #3302 for more on that.
• Use Unamanged to pass the self reference into and out of the thread-local, avoiding dynamic casts.

Result:

Further improved performance for concurrency takeover.

Before this change:

any() from event loop thread (fast path): 30.654833 ns/iteration
any() from outside (fallback path): 17.881084 ns/iteration

After this change:

any() from event loop thread (fast path): 11.571958 ns/iteration
any() from outside (fallback path): 17.52175 ns/iteration

Nearly a 3x speedup on that particular operation. Not so bad.

Note the irony that before this change it was actually more expensive to call any from on an EL thread than off it. Both calls had to call pthread_getspecific, which means all of this cost was the runtime type conversion.