Mirror of forum.nim-lang.org

1214 :: thread usage scenarios and long term goals

• jlp765 (orginal) [2015-05-12T01:55:19+02:00] view original

  Andreas, et. al.

  can you please advise on the goals/scenarios for using threading in Nim, ie when best to use it and when not to use it.

  Currently, using parallel/spawn in loops to take advantage of multi-core CPUs is a lost cause, because it is way too slow compared to just doing a normal loop (but will this change in the future?)

  Also, what are the long term goals?

---

• filwit (orginal) [2015-05-12T03:22:19+02:00] view original

  Threading in Nim is a hot-topic with many potential future solutions. I certainly can't answer those questions, but I'm not sure why you think parallel/spawn is slow. I use it in a ray-tracer here to effective result (~4x speed up on 4-core CPU). Perhaps your loop or data is too small to take advantage or the cost of spawning threads? Or perhaps you're not spawning the correct number of threads for your CPU (I spawn 16 I think, which keeps the the 4 cores occupied best for the specific workload).

---

• kirbyfan64sos (orginal) [2015-05-12T03:33:41+02:00] view original

  If your loop is too short, the overhead of creating a new thread will outweigh the benefits of parallel processing. That's a language-generic issue that happens everywhere, including C++; Stroustrup covered it in the 4th edition of The C++ Programming Language.

---

• Araq (orginal) [2015-05-12T10:39:21+02:00] view original

  In addition to what filwit said, spawn does not create a thread, it passes a "task" to the thread pool, but it's true parallel is not for micro-threading, that's better left to the modern out-of-order CPUs, perhaps helped by explicit vector operations etc.

---

• Jehan (orginal) [2015-05-12T20:43:14+02:00] view original

  filwit: I certainly can't answer those questions, but I'm not sure why you think parallel/spawn is slow.

  The task creation overhead will generally dominate for short fragments of code. This is really unavoidable; modern processors are inherently limited in how fine-grained their parallelism can be and still be efficient.

  That said, the parallel/spawn framework is very limited in that it can only handle a subset of interesting parallel algorithms well. This affects (among other things) anything involving a shared cache or other shared data structure, such as typical parallel implementations of A* or other search algorithms. In the long term, Nim definitely needs other frameworks or more general language features, too.

---