Project

General

Profile

Actions

Feature #3620

closed

Add Queue, SIzedQueue and ConditionVariable implementations in C in addition to ruby ones

Added by panaggio (Ricardo Panaggio) over 13 years ago. Updated over 10 years ago.

Status:
Closed
Target version:
[ruby-core:31513]

Description

=begin
Queue, SizedQueue and ConditionVariable are important synchronization primitives and are nowadays implemented in Ruby.

Attached patch (initiated by myself and heavily enriched by Nobu) contains these sync primitives implemented in C, which makes them faster (see [1] for the benchmark's code):

Rehearsal -------------------------------------------------
Q#push 1.590000 0.010000 1.600000 ( 1.605502)
T#push 0.600000 0.010000 0.610000 ( 0.630444)
Q#pop 4.390000 0.000000 4.390000 ( 4.389781)
T#pop 0.580000 0.000000 0.580000 ( 0.578918)
Q#empty? 0.480000 0.000000 0.480000 ( 0.484305)
T#empty? 0.360000 0.000000 0.360000 ( 0.358559)
Q#clear 1.210000 0.000000 1.210000 ( 1.214494)
T#clear 0.600000 0.000000 0.600000 ( 0.588611)
Q#size 0.370000 0.000000 0.370000 ( 0.365587)
T#size 0.350000 0.000000 0.350000 ( 0.356985)
Q#num_waiting 0.380000 0.000000 0.380000 ( 0.379199)
T#num_waiting 0.370000 0.000000 0.370000 ( 0.368075)
--------------------------------------- total: 11.300000sec

It has already been discussed on ruby-core (see ruby-core:31100).

This patch is one of the deliverables of my RubySoC project (slot #17): "Improving Ruby's Synchronization Primitives and Core Libraries" [2,3]

[1] http://github.com/panaggio/rubysoc-2010/blob/master/benchmarks/queue.rb
[2] http://pastebin.com/viSnfqe6
[3] http://rubysoc.org/projects
=end


Files

final-queue.patch (18 KB) final-queue.patch patch with proposed alterations: Queue, SizedQueue and ConditionVariable in C panaggio (Ricardo Panaggio), 07/28/2010 02:51 AM
final_queue_without_mutex.diff (16.7 KB) final_queue_without_mutex.diff final_queue patch without using mutex (relying on GVL) funny_falcon (Yura Sokolov), 06/26/2012 04:18 PM
patch.diff (24.3 KB) patch.diff bug fixed final_queue_without_mutex.diff Glass_saga (Masaki Matsushita), 05/17/2013 10:59 PM
patch2.diff (24.4 KB) patch2.diff using rb_funcall2() Glass_saga (Masaki Matsushita), 05/22/2013 09:07 PM
thread.c (12.4 KB) thread.c ko1 (Koichi Sasada), 08/29/2013 06:54 PM
patch3.diff (23 KB) patch3.diff Glass_saga (Masaki Matsushita), 09/05/2013 08:52 PM
Actions #1

Updated by akr (Akira Tanaka) almost 13 years ago

  • Project changed from Ruby to Ruby master
  • Category changed from ext to ext
Actions #2

Updated by normalperson (Eric Wong) over 12 years ago

ping? The subject of Queue performance came up again in:
http://blade.nagaokaut.ac.jp/cgi-bin/scat.rb/ruby/ruby-talk/391324

Updated by normalperson (Eric Wong) over 12 years ago

Issue #3620 has been updated by Eric Wong.

ping? The subject of Queue performance came up again in:
http://blade.nagaokaut.ac.jp/cgi-bin/scat.rb/ruby/ruby-talk/391324

Feature #3620: Add Queue, SIzedQueue and ConditionVariable implementations in C in addition to ruby ones
http://redmine.ruby-lang.org/issues/3620

Author: Ricardo Panaggio
Status: Open
Priority: Normal
Assignee:
Category: ext
Target version:

=begin
Queue, SizedQueue and ConditionVariable are important synchronization primitives and are nowadays implemented in Ruby.

Attached patch (initiated by myself and heavily enriched by Nobu) contains these sync primitives implemented in C, which makes them faster (see [1] for the benchmark's code):

Rehearsal -------------------------------------------------
Q#push 1.590000 0.010000 1.600000 ( 1.605502)
T#push 0.600000 0.010000 0.610000 ( 0.630444)
Q#pop 4.390000 0.000000 4.390000 ( 4.389781)
T#pop 0.580000 0.000000 0.580000 ( 0.578918)
Q#empty? 0.480000 0.000000 0.480000 ( 0.484305)
T#empty? 0.360000 0.000000 0.360000 ( 0.358559)
Q#clear 1.210000 0.000000 1.210000 ( 1.214494)
T#clear 0.600000 0.000000 0.600000 ( 0.588611)
Q#size 0.370000 0.000000 0.370000 ( 0.365587)
T#size 0.350000 0.000000 0.350000 ( 0.356985)
Q#num_waiting 0.380000 0.000000 0.380000 ( 0.379199)
T#num_waiting 0.370000 0.000000 0.370000 ( 0.368075)
--------------------------------------- total: 11.300000sec

It has already been discussed on ruby-core (see ruby-core:31100).

This patch is one of the deliverables of my RubySoC project (slot #17): "Improving Ruby's Synchronization Primitives and Core Libraries" [2,3]

[1] http://github.com/panaggio/rubysoc-2010/blob/master/benchmarks/queue.rb
[2] http://pastebin.com/viSnfqe6
[3] http://rubysoc.org/projects
=end

--
http://redmine.ruby-lang.org

Updated by nahi (Hiroshi Nakamura) about 12 years ago

  • Description updated (diff)
  • Assignee set to ko1 (Koichi Sasada)
Actions #5

Updated by shyouhei (Shyouhei Urabe) about 12 years ago

  • Status changed from Open to Assigned

Updated by funny_falcon (Yura Sokolov) about 12 years ago

Related issue (but about plain ruby realization)
https://bugs.ruby-lang.org/issues/6174

Updated by ko1 (Koichi Sasada) almost 12 years ago

Sorry for absent from this request.

Eric, can this patch work fine on current trunk?
... which patch should I use?

Updated by normalperson (Eric Wong) almost 12 years ago

"ko1 (Koichi Sasada)" wrote:

Eric, can this patch work fine on current trunk?
... which patch should I use?

Yes, I just tested final-queue.patch with current trunk (r36217)
and it works.

The .gitignore hunk in the patch can be ignored, everything
else creates a new file and doesn't conflict.

This is Richard's patch, btw. I'm not sure if Richard still
pays attention to Ruby (haven't seen any activity from him in
a while).

push/pop are still significantly faster on current trunk with the
queue.rb benchmark:

                  user     system      total        real

Q#push 0.640000 0.000000 0.640000 ( 0.642801)
T#push 0.200000 0.000000 0.200000 ( 0.197539)
Q#pop 1.390000 0.000000 1.390000 ( 1.393139)
T#pop 0.230000 0.000000 0.230000 ( 0.228802)

(https://raw.github.com/panaggio/rubysoc-2010/master/benchmarks/queue.rb)

Updated by funny_falcon (Yura Sokolov) almost 12 years ago

It seems that there is no need for mutex in a native queue implementation (considering we have GVL),
and so that queue_synchronized wrapper:
rb_mutex_sleep could be replaces with rb_thread_sleep_forever without semantic change.

Without mutex, native queue becomes 2 times faster.
Modified patch is attached.

Updated by ko1 (Koichi Sasada) almost 12 years ago

Hi,

(2012/06/26 16:18), funny_falcon (Yura Sokolov) wrote:

It seems that there is no need for mutex in a native queue implementation (considering we have GVL),
and so that queue_synchronized wrapper:
rb_mutex_sleep could be replaces with rb_thread_sleep_forever without semantic change.

Without mutex, native queue becomes 2 times faster.
Modified patch is attached.

I found a bug.

begin sample code

q = Thread::Queue.new
#q = Queue.new
th1 = Thread.new{
begin
p [:th1, q.pop]
rescue RuntimeError => e
sleep
p e
end
}
th2 = Thread.new{
sleep 0.1
p [:th2, q.pop]
}
p [th1, th2]
sleep 0.5
th1.raise "async interrupt!"
sleep 0.5
q << :s

th1.join

p [th1, th2]
th2.join # BLOCK forever!

end sample code

When th1 escapes from blocking by "pop" by exception, then waiting list
of Queue should be maintained (remove th1 from waiting list).

Other comment:

  • "extthread" is good name?

  • variable name

static void
wakeup_all_threads(VALUE list)
{
VALUE thread, list0 = list;
long i;

 list = rb_ary_subseq(list, 0, LONG_MAX);
 rb_ary_clear(list0);
 for (i = 0; i < RARRAY_LEN(list); ++i) {
thread = RARRAY_PTR(list)[i];
rb_thread_wakeup_alive(thread);
 }
 RB_GC_GUARD(list);

}

I prefer:

static void
wakeup_all_threads(VALUE list0)
{
VALUE thread, list = rb_ary_subseq(list0, 0, LONG_MAX);
long i;

 rb_ary_clear(list0);
 for (i = 0; i < RARRAY_LEN(list); ++i) {
thread = RARRAY_PTR(list)[i];
rb_thread_wakeup_alive(thread);
 }
 RB_GC_GUARD(list);

}

But I prefer more:

static void
wakeup_all_threads(VALUE list)
{
VALUE thread;
long i;

 for (i = 0; i < RARRAY_LEN(list); ++i) {
thread = RARRAY_PTR(list)[i];
rb_thread_wakeup_alive(thread);
 }
 rb_ary_clear(list);

}

Any reason to dup array before iteration?

  • T_DATA -> T_STRUCT or T_OBJECT

In this case, you don't need to use T_DATA. You can only use T_STRUCT
or T_OBJECT (with hidden attr). Maybe it will be simple.

Thanks,
Koichi


Feature #3620: Add Queue, SIzedQueue and ConditionVariable implementations in C in addition to ruby ones
https://bugs.ruby-lang.org/issues/3620#change-27471

--
// SASADA Koichi at atdot dot net

Updated by ko1 (Koichi Sasada) over 11 years ago

  • Target version set to 2.6

ping.
I think it can be introduce into Ruby 2.0 if there is a nice implementation.

Updated by Glass_saga (Masaki Matsushita) almost 11 years ago

I fixed some bugs:
(1) blocking forever bug pointed out by ko1 in [ruby-core:45950]
(2) SEGV in do_sleep()
(3) SizedQueue's bug which is similar to (1)

Now, the C implementation passes test-all.
Following diff is from final_queue_without_mutex.diff to my patch.

diff --git a/ext/thread/thread.c b/ext/thread/thread.c
index 9aff5e5..33365c1 100644
--- a/ext/thread/thread.c
+++ b/ext/thread/thread.c
@@ -150,7 +150,7 @@ static VALUE
do_sleep(VALUE args)
{
struct sleep_call *p = (struct sleep_call *)args;

  • return rb_funcall(p->argv[0], rb_intern("sleep"), p->argc-1, p->argv+1);
  • return rb_funcall(p->argv[0], rb_intern("sleep"), p->argc-1, p->argv[1]); /* (2) */
    }

static VALUE
@@ -339,15 +339,31 @@ rb_queue_push(VALUE self, VALUE obj)
return self;
}

+struct waiting_delete {

  • VALUE waiting;
  • VALUE th;
    +};

+static VALUE
+queue_delete_from_waiting(struct waiting_delete *p)
+{

  • rb_ary_delete(p->waiting, p->th);
  • return Qnil;
    +}

static VALUE
queue_do_pop(Queue *queue, VALUE should_block)
{

  • while (!RARRAY_LEN(queue->que)) {
  • struct waiting_delete args;
  • while (RARRAY_LEN(queue->que) == 0) {
    if (!(int)should_block) {
    rb_raise(rb_eThreadError, "queue empty");
    }
  •   rb_ary_push(queue->waiting, rb_thread_current());
    
  •   rb_thread_sleep_forever();
    
  •   args.waiting = queue->waiting;
    
  •   args.th      = rb_thread_current();
    
  •   rb_ary_push(args.waiting, args.th);
    
  •   rb_ensure((VALUE (*)())rb_thread_sleep_forever, (VALUE)0, queue_delete_from_waiting, (VALUE)&args); /* (1) */
    

    }

    return rb_ary_shift(queue->que);
    @@ -590,9 +606,14 @@ rb_szqueue_max_set(VALUE self, VALUE vmax)
    static VALUE
    szqueue_do_push(SizedQueue *szqueue, VALUE obj)
    {

  • struct waiting_delete args;

  • VALUE thread;

  • while (queue_length(&szqueue->queue_) >= szqueue->max) {

  •   rb_ary_push(szqueue->queue_wait, rb_thread_current());
    
  •   rb_thread_sleep_forever();
    
  •   args.waiting = szqueue->queue_wait;
    
  •   args.th      = rb_thread_current();
    
  •   rb_ary_push(args.waiting, args.th);
    
  •   rb_ensure((VALUE (*)())rb_thread_sleep_forever, (VALUE)0, queue_delete_from_waiting, (VALUE)&args); /* (3) */
    
    }
    return queue_do_push(&szqueue->queue_, obj);
    }

Updated by Glass_saga (Masaki Matsushita) almost 11 years ago

Here is benchmark results.
I used the code same as [ruby-core:45871].
( https://raw.github.com/panaggio/rubysoc-2010/master/benchmarks/queue.rb )

trunk (r40799):
user system total real
Q#push 3.290000 0.000000 3.290000 ( 3.292121)
Q#pop 1.900000 0.000000 1.900000 ( 1.898899)
Q#empty? 0.120000 0.000000 0.120000 ( 0.122000)
Q#clear 1.700000 0.010000 1.710000 ( 1.706695)
Q#size 0.130000 0.000000 0.130000 ( 0.128248)
Q#num_waiting 0.120000 0.000000 0.120000 ( 0.117081)

proposed:
user system total real
T#push 0.130000 0.000000 0.130000 ( 0.138489)
T#pop 0.120000 0.000000 0.120000 ( 0.119966)
T#empty? 0.100000 0.000000 0.100000 ( 0.104266)
T#clear 0.130000 0.000000 0.130000 ( 0.135279)
T#size 0.110000 0.000000 0.110000 ( 0.104839)
T#num_waiting 0.100000 0.000000 0.100000 ( 0.097017)

Updated by nobu (Nobuyoshi Nakada) almost 11 years ago

=begin
You can use (({rb_funcall2()})) instead here.

  • return rb_funcall(p->argv[0], rb_intern("sleep"), p->argc-1, p->argv+1);
  • return rb_funcall(p->argv[0], rb_intern("sleep"), p->argc-1, p->argv[1]); /* (2) */
    =end

Updated by Glass_saga (Masaki Matsushita) almost 11 years ago

2013/5/18 nobu (Nobuyoshi Nakada)

You can use rb_funcall2() instead here.

I made a patch which use rb_funcall2().

diff from patch.diff to patch2.diff:

diff --git a/ext/thread/thread.c b/ext/thread/thread.c
index 33365c1..f2d17e9 100644
--- a/ext/thread/thread.c
+++ b/ext/thread/thread.c
@@ -142,15 +142,17 @@ rb_condvar_initialize(VALUE self)
}

struct sleep_call {

  • int argc;
  • VALUE *argv;
  • VALUE mutex;
  • VALUE timeout;
    };

+static ID id_sleep;
+
static VALUE
do_sleep(VALUE args)
{
struct sleep_call *p = (struct sleep_call *)args;

  • return rb_funcall(p->argv[0], rb_intern("sleep"), p->argc-1, p->argv[1]);
  • return rb_funcall2(p->mutex, id_sleep, 1, &p->timeout);
    }

static VALUE
@@ -173,12 +175,16 @@ static VALUE
rb_condvar_wait(int argc, VALUE *argv, VALUE self)
{
VALUE waiters = get_condvar_ptr(self)->waiters;

  • VALUE mutex, timeout;
    struct sleep_call args;
  • args.argc = argc;
  • args.argv = argv;
  • rb_scan_args(argc, argv, "11", &mutex, &timeout);
  • args.mutex = mutex;
  • args.timeout = timeout;
    rb_ary_push(waiters, rb_thread_current());
    rb_ensure(do_sleep, (VALUE)&args, delete_current_thread, waiters);
  • return self;
    }

@@ -607,7 +613,6 @@ static VALUE
szqueue_do_push(SizedQueue *szqueue, VALUE obj)
{
struct waiting_delete args;

  • VALUE thread;

    while (queue_length(&szqueue->queue_) >= szqueue->max) {
    args.waiting = szqueue->queue_wait;
    @@ -699,6 +704,8 @@ Init_thread(void)
    VALUE rb_cQueue = DEFINE_CLASS_UNDER_THREAD(Queue, rb_cObject);
    VALUE rb_cSizedQueue = DEFINE_CLASS_UNDER_THREAD(SizedQueue, rb_cQueue);

  • id_sleep = rb_intern("sleep");
  • rb_define_alloc_func(rb_cConditionVariable, condvar_alloc);
    rb_define_method(rb_cConditionVariable, "initialize", rb_condvar_initialize, 0);
    rb_define_method(rb_cConditionVariable, "wait", rb_condvar_wait, -1);

Updated by ko1 (Koichi Sasada) over 10 years ago

I rewrite to use T_STRUCT instead of T_DATA (attached).

Matsushita-san:
Could you check it?

Issue:
With ext/thraed/thread.c instead of lib/thread.rb, we can't install ruby itself.

(1) rbinstall.rb is invoked with miniruby
(2) rbinstall.rb requires rubygems.rb
(3) rubygems.rb requires lib/rubygems/core_ext/kernel_require.rb
(4) lib/rubygems/core_ext/kernel_require.rb requires lib/monitor.rb
(5) lib/monitor.rb requires thread.rb
Problems are
(a) miniruby can't require extension libraries
(b) rbinstall.rb is not invoked with -I option

Easy way to solve these problem is use rbinstall.rb by ./ruby instead of ./miniruby with -I option.

Another solution is to embed thread.rb features (CV and Queue) as embeded classes.

Updated by normalperson (Eric Wong) over 10 years ago

"ko1 (Koichi Sasada)" wrote:

Another solution is to embed thread.rb features (CV and Queue) as
embeded classes.

I prefer this with no extra .so. Too many .so files hurts load time.
Just leave an empty thread.rb for compatibility.

Updated by ko1 (Koichi Sasada) over 10 years ago

(2013/08/30 12:14), SASADA Koichi wrote:

A patch for this approach:
http://www.atdot.net/sp/raw/m5xbsm

Note that last patch does not care about deadlock detection.

--
// SASADA Koichi at atdot dot net

Updated by nobu (Nobuyoshi Nakada) over 10 years ago

(13/08/29 18:54), ko1 (Koichi Sasada) wrote:

Problems are
(a) miniruby can't require extension libraries
(b) rbinstall.rb is not invoked with -I option

Easy way to solve these problem is use rbinstall.rb by ./ruby instead of ./miniruby with -I option.

Use $(RUNRUBY) in $(INSTRUBY) instead of $(MINIRUBY).

Updated by nobu (Nobuyoshi Nakada) over 10 years ago

(13/08/30 15:26), Nobuyoshi Nakada wrote:

(13/08/29 18:54), ko1 (Koichi Sasada) wrote:

Problems are
(a) miniruby can't require extension libraries
(b) rbinstall.rb is not invoked with -I option

Easy way to solve these problem is use rbinstall.rb by ./ruby instead of ./miniruby with -I option.

Use $(RUNRUBY) in $(INSTRUBY) instead of $(MINIRUBY).

Since tool/rbinstall.rb replaces $:, it doesn't work as-is.


diff --git a/common.mk b/common.mk
index f295fb5..64ff5cc 100644
--- a/common.mk
+++ b/common.mk
@@ -123,7 +123,7 @@ SCRIPT_ARGS = --dest-dir="$(DESTDIR)"
--make-flags="$(MAKEFLAGS)"
EXTMK_ARGS = $(SCRIPT_ARGS) --extension $(EXTS) --extstatic $(EXTSTATIC)
--make-flags="V=$(V) MINIRUBY='$(MINIRUBY)'" --
-INSTRUBY = $(SUDO) $(MINIRUBY) $(srcdir)/tool/rbinstall.rb
+INSTRUBY = $(SUDO) $(RUNRUBY) -r./$(arch)-fake $(srcdir)/tool/rbinstall.rb
INSTRUBY_ARGS = $(SCRIPT_ARGS)
--data-mode=$(INSTALL_DATA_MODE)
--prog-mode=$(INSTALL_PROG_MODE)
@@ -449,7 +449,7 @@ post-no-install-doc::

CLEAR_INSTALLED_LIST = clear-installed-list

-install-prereq: $(CLEAR_INSTALLED_LIST) PHONY
+install-prereq: $(CLEAR_INSTALLED_LIST) yes-fake PHONY

clear-installed-list: PHONY
@> $(INSTALLED_LIST) set MAKE="$(MAKE)"

--
Nobu Nakada

Actions #22

Updated by Glass_saga (Masaki Matsushita) over 10 years ago

Sorry for my late response.
I fixed a bug in ext/thread.c ([ruby-core:56861]).
It was not compatible with Objects extended by Mutex_m and test/test_mutex_m.rb failed.
Moreover, I use rb_thread_sleep_deadly() to make it get along with deadlock detection.

Attached patch includes:

  • updated ext/thread.c
  • added test from [ruby-core:45950] to test/thread/test_queue.rb
  • common.mk rewrited by Nakada-san

Updated by ko1 (Koichi Sasada) over 10 years ago

(2013/09/05 20:52), Glass_saga (Masaki Matsushita) wrote:

I fixed a bug in ext/thread.c ([ruby-core:56861]).
It was not compatible with Objects extended by Mutex_m and test/test_mutex_m.rb failed.
Moreover, I use rb_thread_sleep_deadly() to make it get along with deadlock detection.

Attached patch includes:

  • updated ext/thread.c
  • added test from [ruby-core:45950] to test/thread/test_queue.rb
  • common.mk rewrited by Nakada-san

Go ahead.

I'll make another ticket to embed Queue in srcdir/thread.c.

--
// SASADA Koichi at atdot dot net

Actions #24

Updated by Anonymous over 10 years ago

  • Status changed from Assigned to Closed
  • % Done changed from 0 to 100

This issue was solved with changeset r42862.
Ricardo, thank you for reporting this issue.
Your contribution to Ruby is greatly appreciated.
May Ruby be with you.


  • common.mk: use RUNRUBY instead of MINIRUBY because MINIRUBY can't
    require extension libraries. The patch is from nobu
    (Nobuyoshi Nakada).

  • ext/thread/extconf.rb: for build ext/thread/thread.c.

  • include/ruby/intern.h: ditto.

  • thread.c: ditto.

  • lib/thread.rb: removed and replaced by ext/thread/thread.c.

  • ext/thread/thread.c: Queue, SizedQueue and ConditionVariable
    implementations in C. This patch is based on patches from panaggio
    (Ricardo Panaggio) and funny_falcon (Yura Sokolov) and ko1
    (Koichi Sasada). [ruby-core:31513] [Feature #3620]

  • test/thread/test_queue.rb (test_queue_thread_raise): add a test for
    ensuring that killed thread should be removed from waiting threads.
    It is based on a code by ko1 (Koichi Sasada). [ruby-core:45950]

Updated by kosaki (Motohiro KOSAKI) over 10 years ago

(9/5/13 8:55 AM), SASADA Koichi wrote:

(2013/09/05 20:52), Glass_saga (Masaki Matsushita) wrote:

I fixed a bug in ext/thread.c ([ruby-core:56861]).
It was not compatible with Objects extended by Mutex_m and test/test_mutex_m.rb failed.
Moreover, I use rb_thread_sleep_deadly() to make it get along with deadlock detection.

Attached patch includes:

  • updated ext/thread.c
  • added test from [ruby-core:45950] to test/thread/test_queue.rb
  • common.mk rewrited by Nakada-san

Go ahead.

I'll make another ticket to embed Queue in srcdir/thread.c.

We already have.

http://bugs.ruby-lang.org/issues/7923

Updated by ko1 (Koichi Sasada) over 10 years ago

(2013/09/13 3:03), KOSAKI Motohiro wrote:

We already have.

http://bugs.ruby-lang.org/issues/7923

I feel "making trap safe Queue (Feature #7923)" and embedding Queue
class is not same.

--
// SASADA Koichi at atdot dot net

Actions

Also available in: Atom PDF

Like0
Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0