Re: [RFC PATCH] introduce sys_membarrier(): process-wide memory barrier

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Fri, Jan 08, 2010 at 09:38:42PM -0500, Mathieu Desnoyers wrote:
> * Paul E. McKenney (paulmck@linux.vnet.ibm.com) wrote:
> > On Fri, Jan 08, 2010 at 08:02:31PM -0500, Mathieu Desnoyers wrote:
> > > * Paul E. McKenney (paulmck@linux.vnet.ibm.com) wrote:
> > > > On Fri, Jan 08, 2010 at 06:53:38PM -0500, Mathieu Desnoyers wrote:
> > > > > * Steven Rostedt (rostedt@goodmis.org) wrote:
> > > > > > Well, if we just grab the task_rq(task)->lock here, then we should be
> > > > > > OK? We would guarantee that curr is either the task we want or not.
> > > > > 
> > > > > Hrm, I just tested it, and there seems to be a significant performance
> > > > > penality involved with taking these locks for each CPU, even with just 8
> > > > > cores. So if we can do without the locks, that would be preferred.
> > > > 
> > > > How significant?  Factor of two?  Two orders of magnitude?
> > > > 
> > > 
> > > On a 8-core Intel Xeon (T is the number of threads receiving the IPIs):
> > > 
> > > Without runqueue locks:
> > > 
> > > T=1: 0m13.911s
> > > T=2: 0m20.730s
> > > T=3: 0m21.474s
> > > T=4: 0m27.952s
> > > T=5: 0m26.286s
> > > T=6: 0m27.855s
> > > T=7: 0m29.695s
> > > 
> > > With runqueue locks:
> > > 
> > > T=1: 0m15.802s
> > > T=2: 0m22.484s
> > > T=3: 0m24.751s
> > > T=4: 0m29.134s
> > > T=5: 0m30.094s
> > > T=6: 0m33.090s
> > > T=7: 0m33.897s
> > > 
> > > So on 8 cores, taking spinlocks for each of the 8 runqueues adds about
> > > 15% overhead when doing an IPI to 1 thread. Therefore, that won't be
> > > pretty on 128+-core machines.
> > 
> > But isn't the bulk of the overhead the IPIs rather than the runqueue
> > locks?
> > 
> >      W/out RQ       W/RQ   % degradation
> fix:
>        W/out RQ       W/RQ   ratio
> > T=1:    13.91      15.8    1.14
> > T=2:    20.73      22.48   1.08
> > T=3:    21.47      24.75   1.15
> > T=4:    27.95      29.13   1.04
> > T=5:    26.29      30.09   1.14
> > T=6:    27.86      33.09   1.19
> > T=7:    29.7       33.9    1.14
> 
> These numbers tell you that the degradation is roughly constant as we
> add more threads (let's consider 1 thread per core, 1 IPI per thread,
> with active threads). It is all run on a 8-core system will all cpus
> active. As we increase the number of IPIs (e.g. T=2 -> T=7) we add 9s,
> for 1.8s/IPI (always for 10,000,000 sys_membarrier() calls), for an
> added 180 ns/core per call. (note: T=1 is a special-case, as I do not
> allocate any cpumask.)
> 
> Using the spinlocks adds about 3s for 10,000,000 sys_membarrier() calls
> or a 8-core system, for an added 300 ns/core per call.
> 
> So the overhead of taking the task lock is about twice higher, per core,
> than the overhead of the IPIs. This is understandable if the
> architecture does an IPI broadcast: the scalability problem then boils
> down to exchange cache-lines to inform the ipi sender that the other
> cpus have completed. An atomic operation exchanging a cache-line would
> be expected to be within the irqoff+spinlock+spinunlock+irqon overhead.

Let me rephrase the question...  Isn't the vast bulk of the overhead
something other than the runqueue spinlocks?

> > So if we had lots of CPUs, we might want to fan the IPIs out through
> > intermediate CPUs in a tree fashion, but the runqueue locks are not
> > causing excessive pain.
> 
> A tree hierarchy may not be useful for sending the IPIs (as, hopefully,
> they can be broadcasted pretty efficiciently), but however could be
> useful when waiting for the IPIs to complete efficiently.

OK, given that you precompute the CPU mask, you might be able to take
advantage of hardware broadcast, on architectures having it.

> > How does this compare to use of POSIX signals?  Never mind, POSIX
> > signals are arbitrarily bad if you have way more threads than are
> > actually running at the time...
> 
> POSIX signals to all threads are terrible in that they require to wake
> up all those threads. I have not even thought it useful to compare
> these two approaches with benchmarks yet (I'll do that when the
> sys_membarrier() support is implemented in liburcu).

It would be of some interest.  I bet that the runqueue spinlock overhead
is -way- down in the noise by comparison to POSIX signals, even when all
the threads are running.  ;-)

							Thanx, Paul