PARALLEL.RU - Информационно-аналитический центр по параллельным вычислениям

Model implementation of buffered mode

The model implementation uses the packing and unpacking functions described in Section Pack and unpack and the nonblocking communication functions described in Section Nonblocking communication .

We assume that a circular queue of pending message entries (PME) is maintained. Each entry contains a communication request handle that identifies a pending nonblocking send, a pointer to the next entry and the packed message data. The entries are stored in successive locations in the buffer. Free space is available between the queue tail and the queue head.

A buffered send call results in the execution of the following code.

• Traverse sequentially the PME queue from head towards the tail, deleting all entries for communications that have completed, up to the first entry with an uncompleted request; update queue head to point to that entry.
• Compute the number, n, of bytes needed to store an entry for the new message. An upper bound on n can be computed as follows: A call to the function
  MPI_PACK_SIZE(count, datatype, comm, size), with the count, datatype and comm arguments used in the MPI_BSEND call, returns an upper bound on the amount of space needed to buffer the message data (see Section Pack and unpack ). The MPI constant MPI_BSEND_OVERHEAD provides an upper bound on the additional space consumed by the entry (e.g., for pointers or envelope information).
• Find the next contiguous empty space of n bytes in buffer (space following queue tail, or space at start of buffer if queue tail is too close to end of buffer). If space is not found then raise buffer overflow error.
• Append to end of PME queue in contiguous space the new entry that contains request handle, next pointer and packed message data; MPI_PACK is used to pack data.
• Post nonblocking send (standard mode) for packed data.
• Return