A working experiment at stackful coroutines that can be spawned, suspended and resumed by any thread; without sacrificing much performance.
Each thread has its own scheduler that first try to exhaust its own queue, then try to steal from a random scheduler. Schedulers' queue follow the "Scheduling Multithreaded Computations by Work Stealing" (1999) paper, with tweaks from the follow-up "Thread Scheduling for Multiprogrammed Multiprocessors" (2001) paper.
Most recently enqueued fibers are resumed sooner by the schedulers (to improve cache reuses) while the least recently enqueued fibers will be stolen by empty schedulers (to avoid starvation).
Thread-safe and fiber-aware synchronization primitives such as mutexes and monitors (condition variables) are available. An example channel implementation is also available, but limited to pass pointers and the overall performance is still quite poor.
See the samples and benchmarks directories for usage examples.
-
Being multithreaded doesn't mean that an application performance will be improved.
An application that switches contexts a lot, or doesn't need to communicate much between coroutines, or is CPU-bound will have better performance with many threads, because many fibers can be advanced in parallel.
But an application that communicates a lot between coroutines and is mostly IO-bound can perform poorly with many threads, because the threads will have to synchronize a lot, which can significantly hinder performance.
-
Starting more threads than available CPU cores/threads will perform better than starting less threads.
I'm not sure about the reasons. Maybe this delegates more to the kernel, which does a better job at scheduling threads.
Distributed under the Apache 2.0 license.
Papers on-non blocking structures, shared-memory multiprocessor schedulers and synchronization primitives:
- "Empirical Studies of Competitive Spinning for a Shared-Memory Multiprocessor" (1991)
- "Simple, Fast, and Practical Non-Blocking and Blocking Concurrent Queue Algorithms" (1996)
- "Scheduling Multithreaded Computations by Work Stealing" (1999)
- "Verification of a Concurrent Deque Implementation" (1999)
- "Thread Scheduling for Multiprogrammed Multiprocessors" (2001)
- "An optimistic approach to lock-free FIFO queues" (2008)
Helpful code samples & thread synchronisation informations: