搭配阻塞接收的非阻塞发送仍需调用MPI_Wait吗？

Great question! Let's break this down clearly—because even though blocking receives guarantee their own completion, non-blocking sends have hard rules you can't ignore.

Short Answer #

Yes, you absolutely still need to call MPI_Wait (or MPI_Test/MPI_Waitall, etc.) for any non-blocking send request, regardless of whether the matching receive is blocking or not.

Why This Matters #

Let's dive into the key reasons:

• Resource management: MPI implementations allocate internal resources (like request handles, temporary buffer space) for every non-blocking operation. If you don't complete the request with MPI_Wait or similar functions, these resources will leak over time—leading to slowdowns, crashes, or unpredictable behavior.
• Buffer safety: A non-blocking send doesn't guarantee your send buffer is safe to modify until the operation fully completes. Even if the blocking receive on the other end has finished (meaning data has arrived), MPI might still be using your send buffer internally (e.g., for buffered sends where data is copied to an MPI-managed buffer in the background). Skipping MPI_Wait and modifying the buffer early could corrupt in-transit data or cause undefined behavior.
• MPI standard compliance: The MPI standard explicitly requires that all non-blocking requests are completed. Failing to do so violates the standard, and different implementations might handle this violation inconsistently (some might warn, others might crash silently).

Example to Illustrate #

Suppose you have this code snippet:

// Process 0: Non-blocking send
int send_data = 42;
MPI_Request send_req;
MPI_Isend(&send_data, 1, MPI_INT, 1, 0, MPI_COMM_WORLD, &send_req);

// Process 1: Blocking receive
int recv_data;
MPI_Recv(&recv_data, 1, MPI_INT, 0, 0, MPI_COMM_WORLD, MPI_STATUS_IGNORE);

Even though Process 1's MPI_Recv has returned (confirming data is received), Process 0's send_req is still in an incomplete state. You must add:

// Process 0: Complete the non-blocking send
MPI_Wait(&send_req, MPI_STATUS_IGNORE);

to clean up the request and ensure the send buffer is safe to reuse.

Edge Case: What if I use MPI_Request_free? #

You might wonder if MPI_Request_free can replace MPI_Wait—but no. MPI_Request_free only deallocates the request handle itself; it doesn't wait for the underlying send operation to finish. You still need to confirm the operation is complete before freeing the request (or let MPI_Wait handle both completion and resource cleanup implicitly).

Final Takeaway #

Blocking receives take care of their own completion, but they don't manage the non-blocking send's request on the sender side. Always complete your non-blocking operations with MPI_Wait or MPI_Test family functions—this keeps your code compliant, safe, and free of resource leaks.

内容的提问来源于stack exchange，提问作者Jennifer Q