I am developing a NUMA application and I need to migrate big arrays from a node (the node where they are actually created) to another node.
I can't use the numa_alloc_onnode() function because I need to have shared memory between the process who allocates memory and its child.
The situation is the following:
I have this array (which dimension is equal to the number of NUMA nodes) of (BIG) node* array, where node* is a typedef of a struct:
node** overall_trees;
i-th node* array have to be in the i-th NUMA node.
What I have done is:
overall_trees[i] = mmap(NULL,(1+number_of_nodes)*sizeof(node), PROT_READ | PROT_WRITE, MAP_SHARED | MAP_ANONYMOUS, -1, 0);
numa_move_pages(0, (int) (1+number_of_nodes)*sizeof(node) / PAGE_SIZE, &overall_trees[i], &i, &status, MPOL_MF_MOVE);
but numa_move_pages() returns -1.
Actually this is the first time in which I am dealing with NUMA allocation and maybe what I am trying to do is totally wrong. If a completely different solution is possible, feel free to expose it.
Thank you