MPI

Question 1

What is the function used to initialise the MPI execution environment?

Answer 1

int MPI_Init(int *argc, char ***argv)

Question 2

What is the function used to finalise the MPI execution environment?

Answer 2

int MPI_Finalize()

Question 3

What does int MPI_Comm_size(MPI_Comm comm, int *size) do?

Answer 3

Reports the number of MPI processes in the specified communicator, into the size variable

Question 4

What does int MPI_Comm_rank(MPI_Comm comm, int *rank) do?

Answer 4

Reports the rank of the calling process in the specified communicator, into the rank variable

Question 5

What is the range of ranks for MPI processes?

Answer 5

From 0 to size-1

Question 6

What is the predefined communicator that refers to all concurrent processes in an MPI program?

Answer 6

MPI_COMM_WORLD

Question 7

What are the two basic functions for point-to-point communication in MPI?

Answer 7

MPI_Send and MPI_Recv

Question 8

What is the purpose of the MPI_Send function? What is the full function definition?

Answer 8

• Sends a message to another process
• int MPI_Send(void *buf, int count, MPI_Datatype datatype, int dest, int tag, MPI_Comm comm)

Question 9

What is the purpose of the MPI_Recv function? What is the full function definition?

Answer 9

• Receives a message from another process
• int MPI_Recv(void *buf, int count, MPI_Datatype datatype, int source, int tag, MPI_Comm comm, MPI_Status *status)

Question 10

List some MPI C data types.

Answer 10

• MPI_CHAR
• MPI_INT
• MPI_FLOAT
• MPI_DOUBLE
• etc

Question 11

What are collectives? Why use them?

Answer 11

• Operations for communicating within a group of processes specified by a communicator
• Collectives can make MPI programming more efficient

Question 12

What is the purpose of MPI_Bcast? What is the full function definition?

Answer 12

• Broadcasts data from one process to all others in a communicator (one-to-all)
• int MPI_Bcast(void *buffer, int count, MPI_Datatype datatype, int root, MPI_Comm comm)

Question 13

What is the purpose of MPI_Scatter? What is the full function definition?

Answer 13

• Scatters data from one process to all others in a communicator
• Each process receives a subset of the data
• int MPI_Scatter(const void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm)

Question 14

What is the purpose of MPI_Gather? What is the full function definition?

Answer 14

• Gathers data from all processes in a communicator to one process
• int MPI_Gather(const void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm)

Question 15

What is the purpose of MPI_Allgather? What is the full function definition?

Answer 15

• Gathers data from all processes in a communicator to all processes
• All processes recieve the result
• int MPI_Allgather(const void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int recvcount, MPI_Datatype recvtype, MPI_Comm comm)

Question 16

What is the purpose of MPI_Reduce? What is the full function definition?

Answer 16

• Reduces a variable and returns result to specified process
• int MPI_Reduce(const void *sendbuf, void *recvbuf, int count, MPI_Datatype type, MPI_Op op, int root, MPI_Comm comm)

Question 17

What is the purpose of MPI_Allreduce? What is the full function definition?

Answer 17

• Reduces a variable and returns result to all processes
• int MPI_Allreduce(const void *sendbuf, void *recvbuf, int count, MPI_Datatype type, MPI_Op op, MPI_Comm comm)

Question 18

What command is used to compile an MPI program in C?

Answer 18

mpicc -o program program.c

Question 19

What command is used to run an MPI process?

Answer 19

mpirun -np 5 program

5 is the number of processes

Question 20

What is blocking communication in MPI? What is a benefit?

Answer 20

• Communication where execution block until the message is received or safe to change buffers
• It helps avoid overwriting buffers before it is safe to do so

Question 21

What are non-blocking communications? Why can they be good? Why can they be bad?

Answer 21

• They allow simultaneous computation and communication
• Helps avoid deadlocks
• We need to ensure the program works correctly and does not overwrite buffers

Question 22

What is the purpose of MPI_Isend? What is the full function definition?

Answer 22

• Sends a message without blocking
• int MPI_Isend(void *buf, int count, MPI_Datatype datatype, int dest, int tag, MPI_Comm comm, MPI_Request *request)

Question 23

What is the purpose of MPI_Irecv? What is the full function definition?

Answer 23

• Receives a message without blocking
• int MPI_Irecv(void *buf, int count, MPI_Datatype datatype, int source, int tag, MPI_Comm comm, MPI_Request *request)

Question 24

What is the purpose of MPI_Wait? What is the full function definition?

Answer 24

• Waits for a non-blocking request to complete
• int MPI_Wait(MPI_Request *request, MPI_Status *status)

Question 25

What is the purpose of `MPI_Waitall`? What is the full function definition?

Answer 25

- Waits for all non-blocking requests to complete - `int MPI_Waitall(int count, MPI_Request array_of_requests[], MPI_Status array_of_statuses[])`

Question 26

How do collectives affect non-blocking communication?

Answer 26

A blocking collective synchronises all the processes in the communicator

Question 27

What is Domain Decomposition in MPI?

Answer 27

A method to distribute work by dividing the computational domain among different MPI processes

Question 28

How is a PDE solver typically set up using MPI collectives?

Answer 28

1. Rank zero reads input parameters and *broadcasts* them 2. Rank zero reads initial conditions and *scatters* them 3. Rank zero *gathers* results and writes output

Question 29

What is the significance of using the same timestep in MPI PDE solvers?

Answer 29

All processes must use the same timestep - the *minimum value* across the whole domain

Question 30

What are halos in the context of finite difference stencils?

Answer 30

Copies of rows/columns held by another MPI process to facilitate communication at domain boundaries

Question 31

How often do halos need to be updated?

Answer 31

Once per timestep

Question 32

What is a potential issue when using `MPI_Send` and `MPI_Recv` for halo communication?

Answer 32

It can incur significant overhead