Introduction to MPI

Question 1

Which two functions need to be called to initialise and finalise the MPI environment?

Answer 1

1. MPI_Init to initialise the environment
2. MPI_Finalize to cleanly shut down the MPI environment

Question 2

Which function is used to find the number of processes there are?

Answer 2

MPI_Comm_size returns the number of processes currently running

Question 3

Which function is used to find the current process rank?

Answer 3

MPI_Comm_rank returns the rank of the calling process

Question 4

What is a communicator?

Answer 4

A communicator is a grouping of processes. There is a predefined global communicator known as MPI_COMM_WORLD containing all processes.

Question 5

Which variable is accepted by MPI_Comm_size and MPI_Comm_rank?

Answer 5

comm - the communicator that is being queried.

Question 6

Which compile command is used for MPI?

Answer 6

mpicc

Question 7

Which command is used to run an MPI process?

Answer 7

mpirun - starts multiple instances of the executable

Question 8

How is the number of MPI processes specified?

Answer 8

A flag is passed to mpirun which specified the number of processes to start.

Question 9

How do processes communicate in MPI?

Answer 9

By passing messages

Question 10

What is point-to-point communication?

Answer 10

Message passing between a pair of processes in a communicator

Question 11

Which MPI functions are used for point-to-point messaging?

Answer 11

1. MPI_Send
2. MPI_Recv

Question 12

What does MPI_Send do?

Answer 12

It sends a message to another process describing the data to be sent and the connection to be made.

Question 13

What does MPI_Recv do?

Answer 13

It receives a message from another process describing the data to be received and the connection to be made.

Question 14

How are message lengths specified?

Answer 14

In terms of MPI data types rather than in bytes.

Question 15

What is collective communicaiton?

Answer 15

An alternative to point-to-point messaging that allows for communication to all processes in a communicator in one command.

Question 16

What are the 4 types of MPI collective functions?

Answer 16

1. MPI_Bcast
2. MPI_Scatter
3. MPI_Gather/ALLgather
4. MPI_Reduce/ALLreduce

Question 17

What does MPI_Bcast do?

Answer 17

Send data from one process to all the others in the specified communicator (one-to-all). All processes receive the same data.

Question 18

What does MPI_Scatter do?

Answer 18

Distributes data to all the others in a communicator. Each processes receives a subset of the data.

Question 19

What does MPI_Gather do?

Answer 19

Gather data from all processes in a communicator to one process (all-to-one). Each process contributes a subset of the data received.

Question 20

What does MPI_Allgather do?

Answer 20

Gathers data from all the processes in a communicator to all processes (all-to-all). Each process contributes a subset of the data received. All the processes receive the result.

Question 21

What does MPI_Reduce do?

Answer 21

Carries out a reduction and returns the result to the specified process.

Question 22

What does MPI_Allreduce do?

Answer 22

Carries out a reduction and returns the result to all processes

Question 23

What is a blocking communication?

Answer 23

MPI_Send and MPI_Recv will block until the message is received by the destination process or until it is safe to change the send buffer.

Question 24

Why are blocking communications used?

Answer 24

The behaviour is used to help write correct programs and avoid overwriting a send buffer before it is safe to do so.

Question 25

What is the issue with blocking communications?

Answer 25

If send and receive operations are not correctly matched, there may be a deadlock.

Question 26

What are the non-blocking versions of MPI_Send and MPI_Recv?

Answer 26

MPI_Isend and MPI_Irecv

Question 27

Why are non-blocking communications used?

Answer 27

They are used to avoid deadlocks as you don't have to match send and receives carefully.

Question 28

Why is non-blocking communication more efficient?

Answer 28

There can be an overlap between computation and communication as you can do something else whilst waiting for a message.

Question 29

How are non-blocking communications temporaily blocked?

Answer 29

Using the MPI_Wait or MPI_Waitall commands to wait for the communications to complete.

Question 30

What does MPI_Wait take as input?

Answer 30

The handle for the request to wait for

Question 31

What does MPI_Waitall take as input?

Answer 31

The number of requests to wait for and an array of request handles to wait for

Question 32

What type of communication is collective communication?

Answer 32

Blocking communication. It synchronises all the processes in the communicator.

Question 33

How is work distributed between MPI processed?

Answer 33

Domain Decomposition

Question 34

What do rank zero processes do?

Answer 34

They either read input parameters from files, initial conditions from files, or collect the results from multiple processes.

Question 35

How are differing time steps handled in PDEs?

Answer 35

If the time step is determined by a stability constraint and the velocity varies then the different parts of the domain will have a different maximum allowable time step. The time step needs to be the minimum value from the whole computational domain.

Question 36

How are neighbouring domains handled?

Answer 36

Halos of the values form neighbouring domains are stored and updated then passed when necessary between domains.