7 ms·

Python newthreading - simplified threads, path to eliminate GIL

80 points by timf 16y ago

makmanalp 16y ago
I know it's not the same thing, but I quite liked multiprocessing. Although it's severely limited in terms of passing objects around and there is always the overhead of spawning new processes. As a sidenote, the John Nagle there is the same one in Nagle's algorithm: http://en.wikipedia.org/wiki/Nagle%27s_algorithm http://en.wikipedia.org/wiki/Nagle%27s_algorithm
- jnoller 16y ago
  Good to hear (thanks)! Alas, multiprocessing will always be slagged by the serialization and deserialization of objects. Note, I'm the maintainer, so I'm not trolling :)
  - makmanalp 16y ago
    Sheesh, all these years you think people who write core modules are robots and then one day one comes up and replies to you on a message board! :) What you say is true, but in any case, my congratulations on a very pretty API.
    - eru 16y ago
      He might still be a robot. :o)
      - ableal 16y ago
        Especially in the original sense of hard-working "biological entities that can be mistaken for humans" ;-) http://en.wikipedia.org/wiki/R.U.R._(Rossum%27s_Universal_Robots) http://en.wikipedia.org/wiki/R.U.R._(Rossum%27s_Universal_Ro... Incidentally, another story by Karel Čapek - http://en.wikipedia.org/wiki/The_Absolute_at_Large http://en.wikipedia.org/wiki/The_Absolute_at_Large - features a "reactor that can annihilate matter to produce cheap and abundant energy [and] as a by-product, the absolute. The absolute is a spiritual essence".
        eru 16y ago
        I immensely enjoy "War with the Newts" (http://en.wikipedia.org/wiki/War_with_the_Newts http://en.wikipedia.org/wiki/War_with_the_Newts) and only realized after reading the Wikipedia page, that it is by the same author, Karel Čapek.
TimothyFitz 16y ago
There are dozens of things that would have to be broken to make this work, for instance sys.set_trace. Either tracing is broken, or tracing is rewritten with a new threading strategy (that's not this strategy) or you're back to GIL-land.
- jnoller 16y ago
  Nothing is impossible - not even GIL removal. Just very hard.
  - eru 16y ago
    Some things may be as hard as a complete rewrite, though.
    - jnoller 16y ago
      When talking about an interpreter; I find those terms completely acceptable. So long as the language itself continues to work, and I can still use C extensions.
      - eru 16y ago
        The C extension are one reason why some features are so hard to implement. E.g. reference counting will always be with us, because the current interface to the C extensions rely on it. And you had to change all of them, if you wanted to get rid of it.
jnoller 16y ago
Very, very interesting. I'm going to have to suck it up and pull it from sourceforge - I'll be very interested to see how they implemented the message passing and "safe object" stuff. With time this could change things quite a bit for python.
cool-RR 16y ago
Looks very cool. Of course, I am already fantasizing about it being merged to the main branch. Assuming it will work well, is this something that can get into 3.2? I saw this release schedule for 3.2: http://www.python.org/dev/peps/pep-0392/ http://www.python.org/dev/peps/pep-0392/ I see that 3.2 alpha 1 is due today.
- jnoller 16y ago
  No, it would not get into 3.2, and it doesn't offer speed improvements at this time. It's a great experiment right now. It's probably a year+ off from ever seeing core, if ever.
jrockway 16y ago
The path from a pure python user-level threading library to eliminating the GIL in CPython is longer than the subject implies. Something to keep in mind with Python/Perl/Ruby anyway: Making your app run on 2 cores makes it 2x faster. Implenting your algorithm in Haskell instead makes it 50x faster. Cooperative userspace threads work very well for the kinds of applications you write in Python.
po 16y ago
What is with developers naming their projects newsomething? Django had newforms, Java had nio and Guido van Rossum just posted about the multiple inheritence method resolution order (MRO) which had a python 2.2 new-style and python 2.3 new-style. I could probably come up with a huge list of newwhatever projects. It's such a lazy, uninformative, short-sighted way of naming your work.
- abdulhaq 16y ago
  Naming things 'new-' something is not IT developer specific. New College, Oxford was founded in 1379.
  - po 16y ago
    That's true, there's also The New School in NYC. Someone voted me down; I suppose I'm am being a little too aggressive about it but I do think it's a stupid way to name a project. :-) Most of these projects represent a lot of work and a huge revolutionary leap over the "old" projects. It seems like that much work deserves a more fitting name. "There are only two hard things in Computer Science: cache invalidation and naming things." -Phil Karlton

80 points by timf 16y ago

I know it's not the same thing, but I quite liked multiprocessing. Although it's severely limited in terms of passing objects around and there is always the overhead of spawning new processes. As a sidenote, the John Nagle there is the same one in Nagle's algorithm: http://en.wikipedia.org/wiki/Nagle%27s_algorithm http://en.wikipedia.org/wiki/Nagle%27s_algorithm

jnoller 16y ago
Good to hear (thanks)! Alas, multiprocessing will always be slagged by the serialization and deserialization of objects. Note, I'm the maintainer, so I'm not trolling :)
- makmanalp 16y ago
  Sheesh, all these years you think people who write core modules are robots and then one day one comes up and replies to you on a message board! :) What you say is true, but in any case, my congratulations on a very pretty API.
  - eru 16y ago
    He might still be a robot. :o)
    - ableal 16y ago
      Especially in the original sense of hard-working "biological entities that can be mistaken for humans" ;-) http://en.wikipedia.org/wiki/R.U.R._(Rossum%27s_Universal_Robots) http://en.wikipedia.org/wiki/R.U.R._(Rossum%27s_Universal_Ro... Incidentally, another story by Karel Čapek - http://en.wikipedia.org/wiki/The_Absolute_at_Large http://en.wikipedia.org/wiki/The_Absolute_at_Large - features a "reactor that can annihilate matter to produce cheap and abundant energy [and] as a by-product, the absolute. The absolute is a spiritual essence".
      - eru 16y ago
        I immensely enjoy "War with the Newts" (http://en.wikipedia.org/wiki/War_with_the_Newts http://en.wikipedia.org/wiki/War_with_the_Newts) and only realized after reading the Wikipedia page, that it is by the same author, Karel Čapek.

TimothyFitz 16y ago

There are dozens of things that would have to be broken to make this work, for instance sys.set_trace. Either tracing is broken, or tracing is rewritten with a new threading strategy (that's not this strategy) or you're back to GIL-land.

jnoller 16y ago
Nothing is impossible - not even GIL removal. Just very hard.
- eru 16y ago
  Some things may be as hard as a complete rewrite, though.
  - jnoller 16y ago
    When talking about an interpreter; I find those terms completely acceptable. So long as the language itself continues to work, and I can still use C extensions.
    - eru 16y ago
      The C extension are one reason why some features are so hard to implement. E.g. reference counting will always be with us, because the current interface to the C extensions rely on it. And you had to change all of them, if you wanted to get rid of it.

jnoller 16y ago

Very, very interesting. I'm going to have to suck it up and pull it from sourceforge - I'll be very interested to see how they implemented the message passing and "safe object" stuff. With time this could change things quite a bit for python.

cool-RR 16y ago

Looks very cool. Of course, I am already fantasizing about it being merged to the main branch. Assuming it will work well, is this something that can get into 3.2? I saw this release schedule for 3.2: http://www.python.org/dev/peps/pep-0392/ http://www.python.org/dev/peps/pep-0392/ I see that 3.2 alpha 1 is due today.

jnoller 16y ago
No, it would not get into 3.2, and it doesn't offer speed improvements at this time. It's a great experiment right now. It's probably a year+ off from ever seeing core, if ever.

jrockway 16y ago

The path from a pure python user-level threading library to eliminating the GIL in CPython is longer than the subject implies. Something to keep in mind with Python/Perl/Ruby anyway: Making your app run on 2 cores makes it 2x faster. Implenting your algorithm in Haskell instead makes it 50x faster. Cooperative userspace threads work very well for the kinds of applications you write in Python.

po 16y ago

What is with developers naming their projects newsomething? Django had newforms, Java had nio and Guido van Rossum just posted about the multiple inheritence method resolution order (MRO) which had a python 2.2 new-style and python 2.3 new-style. I could probably come up with a huge list of newwhatever projects. It's such a lazy, uninformative, short-sighted way of naming your work.

abdulhaq 16y ago
Naming things 'new-' something is not IT developer specific. New College, Oxford was founded in 1379.
- po 16y ago
  That's true, there's also The New School in NYC. Someone voted me down; I suppose I'm am being a little too aggressive about it but I do think it's a stupid way to name a project. :-) Most of these projects represent a lot of work and a huge revolutionary leap over the "old" projects. It seems like that much work deserves a more fitting name. "There are only two hard things in Computer Science: cache invalidation and naming things." -Phil Karlton