On Tue, 25 Sep 2007, John Baldwin wrote:
On Monday 24 September 2007 04:57:06 pm Jeff Roberson wrote:
On Mon, 24 Sep 2007, John Baldwin wrote:
On Saturday 22 September 2007 10:32:06 pm Attilio Rao wrote:
Recently several people have reported problems of starvation with
rwlocks.
In particular, users which tried to use rwlock on big SMP environment
(16+ CPUs) found them rather subjected to poor performances and to
starvation of waiters.
Inspecting the code, something strange about adaptive spinning popped
up: basically, for rwlocks, adaptive spinning stubs seem to be
customed too down in the decisioning-loop.
The desposition of the stub will let the thread that would adaptively
spin, to set the respecitve (both read or write) waiters flag on,
which means that the owner of the lock will go down in the hard path
of locking functions and will performe a full wakeup even if the
waiters queues can result empty. This is a big penalty for adaptive
spinning which can make it completely useless.
In addiction to this, adaptive spinning only runs in the turnstile
spinlock path which is not ideal.
This patch ports the approach alredy used for adaptive spinning in sx
locks to rwlocks:
http://users.gufi.org/~rookie/works/patches/kern_rwlock.diff
In sx it is unlikely to see big benefits because they are held for too
long times, but for rwlocks situation is rather different.
I would like to see if people can do benchmarks with this patch (maybe
in private environments?) as I'm not able to do them in short times.
Adaptive spinning in rwlocks can be improved further with other tricks
(like adding a backoff counter, for example, or trying to spin with
the lock held in read mode too), but we first should be sure to start
with a solid base.
I did this for mutexes and rwlocks over a year ago and Kris found it was
slower in benchmarks. www.freebsd.org/~jhb/patches/lock_adapt.patch is
the
last thing I sent kris@ to test (it only has the mutex changes). This
might
be more optimal post-thread_lock since thread_lock seems to have heavily
pessimized adaptive spinning because it now enqueues the thread and then
dequeues it again before doing the adaptive spin. I liked the approach
orginially because it simplifies the code a lot. A separate issue is that
writers don't spin at all if a reader holds the lock, and I think one
thing
to test for that would be an adaptive spin with a static timeout.
We don't enqueue the thread until the same place. We just acquire an
extra spinlock. The thread is not enqueued until turnstile_wait() as
before.
So I looked at this some more and now I don't understand why the trywait() and
cancel() changes were done. Some background:
I wanted the turnstile chain lock to protect only the association between
the turnstile and lock. This way it further reduces the contention domain
for the hard lock/unlock cases. We sometimes see a lot of contention on
the turnstile chain that may be due to hash collisions. If this reduced
contention doesn't help I think it would be possible to revert this
change.
I originally had more of the code running without the turnstile chain
held but while trying to reduce diffs to understand the problem better it
eventually returned to cover almost all of the hard paths.
When the initial turnstile code was split out of the the mutex code, the main
loop in mtx_lock_sleep() looked like this:
while (!_obtain_lock) {
ts = turnstile_lookup(); /* (A) locked ts chain
and did O(n) lookup
to find existing ts
*/
I'm not sure why you're calling this O(n). It's O(n) with the number of
hash collisions, but it's a hash lookup. We should consider using a real
hash algorithm here if the distribution is bad.
Thanks,
Jeff
if (lock unlocked) { (B)
turnstile_release();
continue;
}
if (failed to set contested flag) { (C)
turnstile_release();
continue;
}
if (owner is running) { (D)
turnstile_release();
... spin ...
continue;
}
turnstile_wait(ts, ...); (E)
}
So one thing I noticed after a while is that every iteration of this loop was
doing the O(n) lookup to find the turnstile even though we only used that in
(E). The lookup was wasted overhead for the (B), (C), and (D) cases. Hence
this commit a while later:
jhb 2004-10-12 18:36:20 UTC
FreeBSD src repository
Modified files:
sys/kern kern_condvar.c kern_mutex.c kern_synch.c
subr_sleepqueue.c subr_turnstile.c
sys/sys sleepqueue.h turnstile.h
Log:
Refine the turnstile and sleep queue interfaces just a bit:
- Add a new _lock() call to each API that locks the associated chain lock
for a lock_object pointer or wait channel. The _lookup() functions now
require that the chain lock be locked via _lock() when they are called.
- Change sleepq_add(), turnstile_wait() and turnstile_claim() to lookup
the associated queue structure internally via _lookup() rather than
accepting a pointer from the caller. For turnstiles, this means that
the actual lookup of the turnstile in the hash table is only done when
the thread actually blocks rather than being done on each loop iteration
in _mtx_lock_sleep(). For sleep queues, this means that sleepq_lookup()
is no longer used outside of the sleep queue code except to implement an
assertion in cv_destroy().
- Change sleepq_broadcast() and sleepq_signal() to require that the chain
lock is already required. For condition variables, this lets the
cv_broadcast() and cv_signal() functions lock the sleep queue chain lock
while testing the waiters count. This means that the waiters count
internal to condition variables is no longer protected by the interlock
mutex and cv_broadcast() and cv_signal() now no longer require that the
interlock be held when they are called. This lets consumers of condition
variables drop the lock before waking other threads which can result in
fewer context switches.
MFC after: 1 month
Revision Changes Path
1.52 +16 -15 src/sys/kern/kern_condvar.c
1.151 +4 -5 src/sys/kern/kern_mutex.c
1.263 +4 -3 src/sys/kern/kern_synch.c
1.13 +31 -14 src/sys/kern/subr_sleepqueue.c
1.150 +34 -12 src/sys/kern/subr_turnstile.c
1.5 +11 -12 src/sys/sys/sleepqueue.h
1.5 +17 -16 src/sys/sys/turnstile.h
After that commit the mutex loop looked like this:
while (!_obtain_lock) {
turnstile_lock(); /* (A) locked ts chain
only */
if (lock unlocked) { (B)
turnstile_release();
continue;
}
if (failed to set contested flag) { (C)
turnstile_release();
continue;
}
if (owner is running) { (D)
turnstile_release();
... spin ...
continue;
}
turnstile_wait(ts, ...); /* (E) Does O(n) lookup
*/
}
That is, after the above commit, we only did the O(n) lookup if we needed the
actual turnstile object in (E). This decreased the overhead for the (B),
(C), and (D) cases.
The changes to add trywait() and cancel() basically seem to have reverted the
above commit by going back to doing the O(n) lookup in (A) thus re-adding the
overhead to the (B), (C), and (D) cases. In addition trywait() / cancel()
actually maintain some additional state and acquire another lock so that (B),
(C), and (D) have even more overhead than the first cut of the turnstile code
including having to back out some of the changes in cancel() rather than just
a single spinlock unlock.
This is actually why I had assumed you had enqueued the turnstile in trywait()
as I thought you had dropped the chain lock in trywait() (which would require
enqueing the thread in trywait(), but would also greatly complicate
cancel()). Otherwise, the current trywait() / cancel() changes seem like a
step backwards to me.
Do you really think it is necessary to do the O(n) lookup even when we don't
need it (B, C, D) rather than just doing it in turnstile_wait() (E)?
--
John Baldwin
_______________________________________________
[email protected] mailing list
http://lists.freebsd.org/mailman/listinfo/freebsd-smp
To unsubscribe, send any mail to "[EMAIL PROTECTED]"
