It seems that if a CUDA program breaks down and exits abnormally, the part of host memory implicitly allocated by cudaMallocManaged wouldn’t be returned to system.
I encounter this problem in Windows 7.
Does anyone has idea how to released this part of host memory? Thanks.