Subj : RCU+SMR
To   : comp.programming.threads
From : Joe Seigh
Date : Thu Jul 07 2005 09:48 pm

There's a pre-alpha prototype on
http://atomic-ptr-plus.sourceforge.net/
package name fastsmr_0-0-1

It only supports 32 bit ppc OS X and 32bit intel Linux.
If you want it ported to other platforms, I'll send you
my shipping address.  I'm particularly interested in
Niagara 8 core processor or Opteron 2 core processor
machines. :)

I haven't finished refactoring it so the api isn't totally
stable.  It might be a while before another release as
I'm thinking of doing a project with it to get a better feel
for the issues.  And I still have no idea what the project
will be.

In the OS X version, I intially had the test for thread
run state backwards.  Only I didn't notice it at first
since all the threads were showing a status of not running,
which is what you'd expect on a uniprocessor.  It turns out
that for short preemption like sched_yield or nanosleep for
short periods of time, the thread is just thrown onto the
run queue with "running" status even though it may be a while
before the thread actually runs.  Naturally, this slowed
things down once I fixed the status check.   I speeded up
the testcase by having the threads wait longer which caused
the threads to actually get suspended with a "not running"
status.  Fun things, these scheduler artifacts.


-- 
Joe Seigh

When you get lemons, you make lemonade.
When you get hardware, you make software.

.