Self-Assessment Quiz: MPI

Twenty questions to confirm the MPI model, the call arguments, and the halo-exchange pattern landed before you move on to GPUs in Chapter 35. Aim for 16 or more. Work the "what does this print?" and argument-order questions by hand; the answer key and a topic map are at the end.

Question 1

In the MPI model, the parallel workers are: - A. Threads sharing one address space - B. Processes, each with its own private memory - C. GPU cores - D. Coarray images sharing a global address space

Question 2

mpirun -np 4 ./prog launches: - A. One process that spawns four threads - B. Four copies of the same program, each with a different rank - C. Four different programs - D. One process, four times faster

Question 3

A rank is: - A. The speed of a process - B. A process's integer identity (0 … N−1) within a communicator - C. The number of processes - D. The amount of memory a process has

Question 4

Which call returns the total number of processes in MPI_COMM_WORLD? - A. mpi_comm_rank - B. mpi_comm_size - C. mpi_init - D. mpi_get_count

Question 5

What is wrong with call mpi_comm_rank(MPI_COMM_WORLD, rank) in the use mpi interface? - A. Nothing - B. It is missing the trailing ierr argument - C. MPI_COMM_WORLD is not defined - D. rank must be real

Question 6

The correct argument order for mpi_send is: - A. buf, dest, count, tag, datatype, comm, ierr - B. buf, count, datatype, dest, tag, comm, ierr - C. dest, buf, count, datatype, tag, comm, ierr - D. buf, count, dest, datatype, comm, tag, ierr

Question 7

Compared with mpi_send, an mpi_recv call additionally includes, before ierr: - A. A root argument - B. A status argument - C. An op argument - D. A second buffer

Question 8

You send an array of real(dp) but pass MPI_INTEGER as the datatype. The most likely result is: - A. A compile error - B. A runtime crash with a clear message - C. No error — the receiver silently gets wrong data - D. MPI converts the integers to reals automatically

Question 9

Two ranks each call mpi_send to the other first, then mpi_recv. This code: - A. Always works correctly - B. Can deadlock — it may pass for small messages and hang for large ones - C. Always deadlocks immediately - D. Is a compile error

Question 10

The cleanest single-call fix for the exchange in Question 9 is: - A. mpi_bcast - B. mpi_sendrecv - C. mpi_reduce - D. mpi_barrier

Question 11

mpi_bcast sends: - A. Different pieces of an array, one per process - B. One process's value to every process - C. Every process's value combined onto one process - D. A message between two named processes

Question 12

The difference between mpi_reduce and mpi_allreduce is: - A. allreduce is slower - B. reduce delivers the result to one root; allreduce delivers it to every process - C. reduce only sums; allreduce only takes the max - D. There is no difference

Question 13

Running example-03-collectives.f90 with -np 8, the "sum of (rank+1)" line prints: - A. 8 - B. 28 - C. 36 - D. 64

Question 14

A ghost cell (halo cell) is: - A. A cell whose value is always zero - B. A stored copy of a neighbouring subdomain's boundary data that the process does not own - C. A cell that has been deallocated - D. The corner cell of the global grid

Question 15

Once a process's halo is filled each step, the interior stencil update it runs is: - A. A special distributed version of the stencil - B. The unmodified serial Chapter 24 stencil - C. Different for each process - D. Skipped for the edge rows

Question 16

MPI_PROC_NULL is used at the top and bottom strips because: - A. It makes those processes run faster - B. A send/recv to it is a no-op, so the same exchange code runs everywhere with no boundary if - C. It fills the ghost cells with zeros - D. It is required by mpi_init

Question 17

The solver stores its strip as u(nx, 0:nloc+1) — full width on the first index, decomposed direction on the second — because: - A. It looks tidier - B. In column-major Fortran a whole row u(:,k) is then contiguous, so each halo message is a contiguous buffer - C. MPI requires the second index to be the ghost - D. It uses less memory

Question 18

True or false, with justification: a data race on the temperature field, the hazard you feared in OpenMP, is impossible between MPI processes.

Question 19

Non-blocking communication (mpi_isend/mpi_irecv + mpi_wait) is worth the extra complexity because it: - A. Uses less memory - B. Lets a process compute ghost-independent work while the exchange is in flight (overlap) - C. Cannot deadlock while mpi_sendrecv can - D. Automatically checks ierr

Question 20

What does this print, launched with -np 4? (Reason about the collective.)

mine = real(rank, dp)                              ! 0, 1, 2, 3
call mpi_allreduce(mine, s, 1, MPI_DOUBLE_PRECISION, MPI_SUM, MPI_COMM_WORLD, ierr)
if (rank == 0) print '(f4.1)', s
  • A. 3.0
  • B. 6.0
  • C. 4.0
  • D. 10.0

Answer Key

Q Ans Why
1 B MPI processes have private memory; they coordinate by messages, not shared variables.
2 B SPMD: -np 4 runs four copies of the one program, each with a distinct rank.
3 B A rank is a process's 0-based identity within a communicator.
4 B mpi_comm_size returns the process count; mpi_comm_rank returns the caller's own rank.
5 B Every mpi_* call in the use mpi interface needs the trailing ierr.
6 B buf, count, datatype, dest, tag, comm, ierr — memorize it.
7 B mpi_recv inserts a status argument before ierr (and names a source, not a dest).
8 C MPI trusts your datatype; it copies the wrong bytes with no error — the dangerous kind of bug.
9 B Standard-mode mpi_send may buffer small messages (works) but blocks on large ones (deadlock).
10 B mpi_sendrecv does both in one deadlock-free call.
11 B Broadcast: one process's value copied to all.
12 B reduce → one root; allreduce → every process gets the combined result.
13 C Sum of rank+1 for ranks 0..7 is $1+2+\dots+8 = N(N+1)/2 = 36$.
14 B A ghost cell is a stored, unowned copy of a neighbour's boundary data.
15 B With the halo filled, the stencil reads ghosts uniformly — the exact serial update runs.
16 B Send/recv to MPI_PROC_NULL is a no-op, removing every boundary special-case.
17 B Column-major makes u(:,k) contiguous, so the halo ships as a plain contiguous buffer.
18 True Each process updates only its own rows in private memory; nothing is shared to race over. What replaces the worry is communication correctness — exchanging the right halo at the right time.
19 B Non-blocking's payoff is overlap: compute the deep interior while the halo is in transit.
20 B Sum of ranks 0+1+2+3 = 6; allreduce gives it to all, rank 0 prints 6.0.

Topics to review by question

  • Q1–5 → §34.1 (the MPI model: processes, ranks, communicators, the skeleton and ierr).
  • Q6–10 → §34.2 (point-to-point: argument order, datatype, deadlock and mpi_sendrecv).
  • Q11–13, Q20 → §34.3 (collectives: bcast, reduce vs allreduce, and what they compute).
  • Q14–17 → §34.4 (domain decomposition, ghost cells, MPI_PROC_NULL, column-major layout).
  • Q18–19 → §34.4 and the Spaced Review (distributed vs shared memory; non-blocking overlap).

Scored below 16? The two areas most worth rereading are the mpi_send/mpi_recv argument order and the deadlock cure (Q6–10, §34.2), and the ghost-cell/halo-exchange pattern (Q14–17, §34.4) — they are the whole of writing a distributed stencil code.