Rajith Attapattu wrote:
On Mon, Feb 2, 2009 at 6:05 PM, Carl Trieloff <[email protected]> wrote:
Some of the work being done has been around performance. At this point we
have 2 locks
that still need some work.
One of these locks are in the OS memory allocator, to this end I have
spoken to the glic/gcc
maintainers and they are working on a better malloc/free for us. However
the other side of the
equation is to do some optional memory caching.
The biggest area where this will help us is around frame caching. An
experiment in this area
has been done with TLS caching (did not work out as the caches need to be
cross thread) and
a global pool test has determined that we can reduce the number of
allocations by 25% with
such a cache.
However to do this we need a lock free bucket. So, if someone has interest
to create a lock free bucket.
I.e. a lock free structure that you can push pointers into and pop them
back off from many threads at
the same time. If the bucket is empty return NULL. No ordering is required.
What is interesting in that all the cost in a lock free fifo is the
corrections to maintain order which we
don't need. so the CAS or DCAS impls of fifo without order can be used. For
algorithms see work
from Maged M. Michael & Michael L. Scott -- or google a bit
I had a stab at this sometime back. Started to experiment with the lock free
version first.
I think you would need DCAS as you need to swap both the pointer and the
counter atomically.
Not sure if this can be achieve with CAS only, especially when u move to 64
bit machines.
Also not all processors support DCAS either. So I gave up this route.
I think a more feasible approach might be the two-lock queue . I had some
code around this.
Unfortunately haven't had much time to tinker with it any further. I believe
Andrew/Gordon had done a simillar execersie.
Andrew had some thoughts around this, perhaps he could chime in.
we don't need fifo for this case, so it can be done with atomics, cas or
dcas - no locks needed.
Carl.
