Subj : Re: NPTL rwlock is really suboptimal
To   : comp.programming.threads
From : Joseph Seigh
Date : Thu Jan 27 2005 11:09 pm

On 27 Jan 2005 18:08:22 -0800, <roger.faulkner@sun.com> wrote:

>
> Joseph Seigh wrote:
>> It appears to give reader preference so if you have a lot of reader
> thread
>> activity, the writer threads starve, i.e. hang indefinitely.  rwlocks
> are
>> a strategy to reduce contention but if you have any kind of
> contention it
>> starves the writers.  You're better off using a mutex.
>>
>> I'll have to see how Solaris rwlock handles this if I get around to
> porting
>> the atomic_ptr stuff to Solaris again.
>
> In Solaris 9 and 10, preference is given to writers.
> In Solaris 10, the priority of the threads is taken into account,
> as specified for UNIX03:
>
> The pthread_rwlock_rdlock() function applies a read lock  to
> the  read-write  lock  referenced  by  rwlock.   The calling
> thread acquires the read lock if a writer does not hold  the
> lock and there are no writers blocked on the lock.
>
> The calling thread does not acquire the  lock  if  a  writer
> holds the lock or if writers of higher or equal priority are
> blocked on the lock; otherwise, the calling thread  acquires
> the  lock.  If  the  read  lock is not acquired, the calling
> thread blocks until it can acquire the lock.

Yeah, I was playing with my sb100 with the orginal solaris 8.
rwlocks don't hang.  They don't do any better than mutexes
with the work load I put on it.

Sytem was wonky if I increased the threads dramatically.  I should
upgrade to 9 or 10 sometime if I have the time and I can find where
Sun hides the system iso's and supplemental tools iso.

-- 
Joe Seigh

.