The current Parla updates and tracks memory consumption when a task is mapped and completes.
This is a loigcal memory trace and it would be better to exploit actual device memory consumption.
Cupy uses its own memory pool and provides APIs for this on here: link.