Ramine
3/19/2015 1:33:00 AM
Hello,
In this post i will speak about a very important subject...
In my previous posts i have spook about my scalable SeqlockX and spook
about Transactional memory that are optimistic synchronisation
mechanisms, and i have also spook about scalable distributed
reader-writer lock.. and you have seen explaining to you important
subjects.. hope you have understood my previous posts, but there is
still a big problem ! when you use concurrent AVL trees or other
concurrent datastructures that look like concurrent AVL trees you have
to allocate memory for the "nodes" of your AVL tree, those memory nodes
must be accessed when we do a search() in for example a time complexity
of log(n), but to be able to scale the concurrent AVL tree you have to
allocate those memory nodes on different NUMA nodes , but here comes the
problem, when you will use those NUMA nodes on a concurrent AVL tree you
can not always get a good scalability on those NUMA nodes because
accessing the concurrent AVL tree it is like a probabilistic mechanism ,
i mean you can still get "contention" on NUMA nodes, and
since you still can get contention that's not what we call a good
scalability, and this is what i wanted to explain.
Thank you,
Amine Moulay Ramdane.