Re: [RFC PATCH v14] sys_membarrier(): system/process-wide memory barrier (x86)

[Mathieu Desnoyers <mathieu.desnoyers@efficios.com>]

----- Original Message -----
> On Wed, Mar 18, 2015 at 04:52:14PM +0000, Mathieu Desnoyers wrote:
> > ----- Original Message -----
> > > On Wed, Mar 18, 2015 at 12:23:02PM -0400, Mathieu Desnoyers wrote:
> > > > memory barriers in reader:    1701557485 reads, 3129842 writes
> > > > signal-based scheme:          9825306874 reads,    5386 writes
> > > > sys_membarrier:               7992076602 reads,     220 writes
> > > > 
> > > > The dynamic sys_membarrier availability check adds some overhead to
> > > > the read-side compared to the signal-based scheme, but besides that,
> > > > with the expedited scheme, we can see that we are close to the
> > > > read-side
> > > > performance of the signal-based scheme. However, this non-expedited
> > > > sys_membarrier implementation has a much slower grace period than
> > > > signal
> > > > and memory barrier schemes.
> > > 
> > > Doesn't the query flag allow you to find out in advance rather than
> > > dynamically within the reader?  What's the reader performance if you
> > > hardcode availability of membarrier?
> > 
> > What I am currently doing is to use sys_membarrier with a query
> > flag within a lib constructor, and cache the result in a global
> > variable. In the reader, I just test the variable, and thus detect
> > whether I can use sys_membarrier, or if I need to fallback to
> > barriers on both reader and writer.
> > 
> > Are you suggesting I try removing the global variable load+test
> > from the reader fast path ?
> 
> Right.  You said that "The dynamic sys_membarrier availability check
> adds some overhead to the read-side compared to the signal-based
> scheme"; I wondered how much.

With 8 reader threads in parallel, no writer (workload found
in userspace RCU tests/benchmark/test_urcu*.c):

* memory barriers in read-side

                         307.4 million reads/s

* sys_membarrier read-side

  With dynamic check:   1142.0 million reads/s
  Hardcoded barrier():  1453.2 million reads/s  (For a 27% speedup over dynamic check.)

* QSBR (quiescent-state based) read-side

                        2276.9 million reads/s

It might start being worthwhile to consider turning memory barriers
into no-op within lib constructors at some point. Remember that
rcu_read_lock/unlock can be inlined into applications, which may
add to the challenge.

Thanks,

Mathieu

-- 
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com