The user must specify a memory buffer to use when posting send or receive operations. This buffer may be a normal memory region, or it may be a buffer that was allocated using BMI memory management functions. If the user elects to allocate the memory using the BMI facilities, then BMI has the opportunity to optimize the buffer for the type of network being used. This mode of operation is preferred for achieving optimal performance. However, normal memory buffers are also allowed in order to better support certain scenarios common to file system operations. Some file system operations act upon existing memory regions (for example, the client side Unix read() system call). In these situations, we would like to avoid imposing a buffer copy, and instead give the BMI layer the flexibility to handle the buffer at a lower level if possible.
If a memory buffer is allocated using BMI function calls, then it must also be deallocated using BMI. These buffers are not guaranteed to be manageable by standard operating system libraries.