__posix_memalign taking time in ABT_eventual_create

[mdorier]

HPCtoolkit reports a significant overhead from __posix_memalign in ABT_eventual_create. Any reason for using that instead of just malloc?

[mdorier]

The same overhead caused by __posix_memalign happens in ABT_thread_get_attr. We have a code where each time a ULT is created, the first thing its function does is lookup its attribute, and we get hit by the overhead of this __posix_memalign. In the context of ABT_thread_get_attr I would have expected the resulting ABT_thread_attr object to either (1) have an internal reference count so it doesn't need to be copied entirely, or (2) be a reference to the thread's internal attr object and not needing to be freed. What we are looking for in our case is the stack pointer and size so a solution for us would be to have an ABT_thread_get_stack so we don't have to go through an ABT_thread_attr instance.