Re: [PATCH v3] libgfortran: Replace mutex with rwlock

[Bernhard Reutner-Fischer <rep.dot.nop@gmail.com>]

Hi!

[please do not top-post]

On Thu, 20 Apr 2023 21:13:08 +0800
"Zhu, Lipeng" <lipeng.zhu@intel.com> wrote:

> Hi Bernhard,
> 
> Thanks for your questions and suggestions.
> The rwlock could allow multiple threads to have concurrent read-only 
> access to the cache/unit list, only a single writer is allowed.

right.

> Write lock will not be acquired until all read lock are released.

So i must have confused rwlock with something else, something that
allows self to hold a read-lock and upgrade that to a write-lock,
purposely starving all successive incoming readers. I.e. just toggle
your RD_TO_WRLOCK impl, here, atomically. This proved to be benefical in
some situations in the past. Doesn't seem to work with your rwlock,
does it

> And I didn't change the mutex scope when refactor the code, only make a 
> more fine-grained distinction for the read/write cache/unit list.

Yes of course, i can see you did that.

> I complete the comment according to your template, I will insert the 
> comment in the source code in next version patch with other refinement 
> by your suggestions.
> "
> Now we did not get a unit in cache and unit list, so we need to create a
> new unit, and update it to cache and unit list.

s/Now/By now/ or s/Now w/W/ and s/get/find/
"
We did not find a unit in the cache nor in the unit list, create a new
(locked) unit and insert into the unit list and cache.
Manipulating either or both the unit list and the unit cache requires to
hold a write-lock [for obvious reasons]"

Superfluous when talking about pthread_rwlock_wrlock since that
implies that even the process acquiring the wrlock has to first
release it's very own rdlock.

> Prior to update the cache and list, we need to release all read locks,
> and then immediately to acquire write lock, thus ensure the exclusive
> update to the cache and unit list.
> Either way, we will manipulate the cache and/or the unit list so we must
> take a write lock now.
> We don't take the write bit in *addition* to the read lock because:
> 1. It will needlessly complicate releasing the respective lock;

Under pthread_rwlock_wrlock it will deadlock, so that's wrong?
Drop that point then? If not, what's your reasoning / observation?

Under my lock, you hold the R, additionally take the W and then
immediately release the R because you yourself won't read, just write.
But mine's not the pthread_rwlock you talk about, admittedly.

> 2. By separate the read/write lock, it will greatly reduce the
> contention at the read part, while write part is not always necessary or
> most unlikely once the unit hit in cache;

We know that.

> 3. We try to balance the implementation complexity and the performance
> gains that fit into current cases we observed.

.. by just using a pthread_rwlock. And that's the top point iff you
keep it at that. That's a fair step, sure. BTW, did you look at the
RELEASE semantics, respectively the note that one day (and now is that
very day), we might improve on the release semantics? Can of course be
incremental AFAIC

> "

If folks agree on this first step then you have my OK with a catchy
malloc and the discussion recorded here on the list. A second step would
be RELEASE.
And, depending on the underlying capabilities of available locks,
further tweaks, obviously.

PS: and, please, don't top-post

thanks,

> 
> Best Regards,
> Zhu, Lipeng
> 
> On 1/1/1970 8:00 AM, Bernhard Reutner-Fischer wrote:
> > On 19 April 2023 09:06:28 CEST, Lipeng Zhu via Fortran <fortran@gcc.gnu.org> wrote:  
> >> This patch try to introduce the rwlock and split the read/write to
> >> unit_root tree and unit_cache with rwlock instead of the mutex to
> >> increase CPU efficiency. In the get_gfc_unit function, the percentage
> >> to step into the insert_unit function is around 30%, in most instances,
> >> we can get the unit in the phase of reading the unit_cache or unit_root
> >> tree. So split the read/write phase by rwlock would be an approach to
> >> make it more parallel.
> >>
> >> BTW, the IPC metrics can gain around 9x in our test server with 220
> >> cores. The benchmark we used is https://github.com/rwesson/NEAT
> >>  
> >   
> >> +#define RD_TO_WRLOCK(rwlock) \
> >> +  RWUNLOCK (rwlock);\
> >> +  WRLOCK (rwlock);
> >> +#endif
> >> +  
> > 
> >   
> >> diff --git a/libgfortran/io/unit.c b/libgfortran/io/unit.c index
> >> 82664dc5f98..4312c5f36de 100644
> >> --- a/libgfortran/io/unit.c
> >> +++ b/libgfortran/io/unit.c  
> >   
> >> @@ -329,7 +335,7 @@ get_gfc_unit (int n, int do_create)
> >>    int c, created = 0;
> >>
> >>    NOTE ("Unit n=%d, do_create = %d", n, do_create);
> >> -  LOCK (&unit_lock);
> >> +  RDLOCK (&unit_rwlock);
> >>
> >> retry:
> >>    for (c = 0; c < CACHE_SIZE; c++)
> >> @@ -350,6 +356,7 @@ retry:
> >>        if (c == 0)
> >> 	break;
> >>      }
> >> +  RD_TO_WRLOCK (&unit_rwlock);  
> > 
> > So I'm trying to convince myself why it's safe to unlock and only then take the write lock.
> > 
> > Can you please elaborate/confirm why that's ok?
> > 
> > I wouldn't mind a comment like
> > We can release the unit and cache read lock now. We might have to allocate a (locked) unit, below in
> > do_create.
> > Either way, we will manipulate the cache and/or the unit list so we have to take a write lock now.
> > 
> > We don't take the write bit in *addition* to the read lock because..
> > 
> > (that needlessly complicates releasing the respective locks / it triggers too much contention when we..
> > / ...?)
> > 
> > thanks,
> >   
> >>
> >>    if (p == NULL && do_create)
> >>      {
> >> @@ -368,8 +375,8 @@ retry:
> >>    if (created)
> >>      {
> >>        /* Newly created units have their lock held already
> >> -	 from insert_unit.  Just unlock UNIT_LOCK and return.  */
> >> -      UNLOCK (&unit_lock);
> >> +	 from insert_unit.  Just unlock UNIT_RWLOCK and return.  */
> >> +      RWUNLOCK (&unit_rwlock);
> >>        return p;
> >>      }
> >>
> >> @@ -380,7 +387,7 @@ found:
> >>        if (! TRYLOCK (&p->lock))
> >> 	{
> >> 	  /* assert (p->closed == 0); */
> >> -	  UNLOCK (&unit_lock);
> >> +	  RWUNLOCK (&unit_rwlock);
> >> 	  return p;
> >> 	}
> >>
> >> @@ -388,14 +395,14 @@ found:
> >>      }
> >>
> >>
> >> -  UNLOCK (&unit_lock);
> >> +  RWUNLOCK (&unit_rwlock);
> >>
> >>    if (p != NULL && (p->child_dtio == 0))
> >>      {
> >>        LOCK (&p->lock);
> >>        if (p->closed)
> >> 	{
> >> -	  LOCK (&unit_lock);
> >> +	  WRLOCK (&unit_rwlock);
> >> 	  UNLOCK (&p->lock);
> >> 	  if (predec_waiting_locked (p) == 0)
> >> 	    destroy_unit_mutex (p);
> >> @@ -593,8 +600,8 @@ init_units (void)
> >> #endif
> >> #endif
> >>
> >> -#ifndef __GTHREAD_MUTEX_INIT
> >> -  __GTHREAD_MUTEX_INIT_FUNCTION (&unit_lock);
> >> +#if (!defined(__GTHREAD_RWLOCK_INIT) &&
> >> +!defined(__GTHREAD_MUTEX_INIT))
> >> +  __GTHREAD_MUTEX_INIT_FUNCTION (&unit_rwlock);
> >> #endif
> >>
> >>    if (sizeof (max_offset) == 8)  
> >