[cpp-threads] RE: Ping on memory model and concurrency

Boehm, Hans hans.boehm at hp.com
Fri Aug 12 20:46:21 BST 2005


[I was on vacation ...]

Re: New paper:
I agree there needs to be one.  I've been trying to track the current
state of the proposal at
http://www.hpl.hp.com/personal/Hans_Boehm/c++mm/mm.html,
though there are some doubts as to whether major pieces of that
currently
make sense.

The main difficulties we seem to have are:

1) Simple formulations of the memory model can't easily describe
some of the primitives you need to get the last bit of performance,
e.g. to eliminate an extra barrier in the initialization code for
DCL, and for some of the fancier lock-free algorithms.  I think we
have a compromise:

http://jupiter.robustserver.com/pipermail/cpp-threads_decadentplace.org.
uk/2005-July/000556.html

but more work is needed.

2) I was hoping that someone more familiar with C++ library conventions
would make a pass at the atomic operations library interface.  (I'm
working
intermittently on a C version, which should allow a quick and reasonably
portable
prototype implementation once we have the interface.) So far I don't
think
this has happened, though I think we've made some progress on what
should and shouldn't
go into it.  If nobody volunteers, I may try to take a stab at this.  (I
think
Alexander and Peter had some assumptions about what this should look
like.
And I think their basic approach is probably fine, but they originally
included
finer control over memory ordering than we can really specify at this
level.)
I consider this important because it has turned out to be a major
constraint
on the memory model spec.

3) I haven't seen much progress with respect to a threads API (thread
creation,
locking, etc.)  I personally still think this is a difficult political
problem
whose solution is far less critical than the memory model issues.  If
the
memory model issues were solved, it would be easy to fix the
specifications
for pthreads, win32 threads, boost threads, etc.  But the committee
definitely
appeared to be interested in addressing it.

Re: Attendance

I'm currently planning on being there in person.  It would be great if
others
could make it, too.

Hans





More information about the cpp-threads mailing list