In the original implementation, all user messages are buffered by PVM. The user must pack the data into a PVM buffer before sending it, and unpack the data after it has been received into an internal buffer. This approach works well on systems with relatively high communication latency, such as the Ethernet. On MPP systems the packing and unpacking introduce substantial overhead. To solve this problem we added two new PVM functions, namely pvm_psend() and pvm_precv(). These functions combine packing/unpacking and sending/receiving into one single step. They could be mapped directly into the native message passing primitives available on the system, doing away with internal buffers altogether. On the Paragon these new functions give almost the same performance as the native ones.
Although the user can use both pvm_psend() and pvm_send() in the same program, on MPP the pvm_psend() must be matched with pvm_precv(), and pvm_send() with pvm_recv().