1. Parallel Programming C. Ferner & B. Wilkinson, 2014 Introduction to Message Passing Interface (MPI) Introduction 9/4/2012 1

2. Parallel Programming C. Ferner & B. Wilkinson, 2014 Shared-Memory Systems Processor Bus interface Processor Bus interface Processor Bus interface Processor Bus interface Memory controller Memory Processor/ memory us Shared memory All processors can access all of the shared memory 9/4/2012 2

3. Parallel Programming C. Ferner & B. Wilkinson, 2014 For processors that share the same memory, if one processor computes a value or values needed other processors, they need to synchronize Processors that need the data must wait for the processor that computes it However, all processors can access the same memory so nothing else is required. 9/4/2012 3

4. Parallel Programming C. Ferner & B. Wilkinson, 2014 Distributed-Memory Systems Processor Interconnection network Local Computers Messages memory For clusters each processor cannot access the memory of other processors. Memory is private. 9/4/2012 4

5. Parallel Programming C. Ferner & B. Wilkinson, 2014 If one processor computes a value or values needed other processors, it has to transmit that data This requires message-passing All data is private 9/4/2012 5

6. Parallel Programming C. Ferner & B. Wilkinson, 2014 MPI (Message Passing Interface) Widely adopted message passing library standard. MPI-1 finalized in 1994, MPI-2 in 1996, MPI-3 in 2012 Process-based -- processes communicate between themselves with messages. Point-to-point and collectively. A specification, not an implementation. Several free implementations exist, OpenMPI, MPICH, Large number of routines: MPI-1 128 routines, MPI-2 287 routines, MPI-3 440+ routines, but typically only a few used. C and Fortran bindings (C++ removed from MPI-3) Originally for distributed systems but now used for all types, clusters, shared memory, hybrid. 6

7. Parallel Programming C. Ferner & B. Wilkinson, 2014 Message-Passing Sockets (very low level) Parallel Virtual Machine (created by Oak Ridge National Lab) MPI (Message Passing Interface) - standard  LAM  MPICH  OpenMPI (we use this one) All of these are libraries that can be used from within a C program 9/4/2012 7

8. Parallel Programming C. Ferner & B. Wilkinson, 2014 To begin MPI_Init() – Initializes the processes and gets them ready to run. Should be the first executable statement in the program MPI_Finalize() – cleans up after the parallel program is done. Should be the last executable statement in the program Although all of the processes exist before and after the Init and Finalize, it is convenient to think of MPI_Init() as when all the processors are created and MPI_Finialize() is when they are destroyed 9/4/2012 8

9. Parallel Programming C. Ferner & B. Wilkinson, 2014 To begin Process 0 Process 1Process 2Process 3Process 0 MPI_Init() MPI_Finalize() Time 9/4/2012 9

10. Parallel Programming C. Ferner & B. Wilkinson, 2014 To begin int main(int argc, char **argv) { MPI_Init(&argc, &argv); MPI_Finalize(); } 9/4/2012 10

11. Parallel Programming C. Ferner & B. Wilkinson, 2014 To begin MPI_Init takes two arguments &argc and &argv, which are the arguments of main. MPI_Finalize takes no arguments. Each processor is assigned a rank in the range 0 ≤ rank < NP, where NP is the number of processes being used. 9/4/2012 11

12. Parallel Programming C. Ferner & B. Wilkinson, 2014 Other Useful Functions To determine how many processes there are: MPI_Comm_size(MPI_COMM_WORLD, &NP); To determine the current process’ rank among all the processes: MPI_Comm_rank(MPI_COMM_WORLD, &myrank); Each processor is given a unique rank in the range 0 ≤ rank < NP 9/4/2012 12

13. Parallel Programming C. Ferner & B. Wilkinson, 2014 How is the number of processes determined? When you run your MPI program, you can specify how many processes you want: $ mpirun –np 8 The –np option tells mpirun to run your parallel program using the specified number of processes. OR $ mpiexec –n 8 The –n option tells mpiexec to run your parallel program using the specified number of processes. 9/4/2012 13

14. Parallel Programming C. Ferner & B. Wilkinson, 2014 How can these be used? #include int main(int argc, char **argv) { int NP, myrank; MPI_Init(&argc, &argv); MPI_Comm_size(MPI_COMM_WORLD, &NP); MPI_Comm_rank(MPI_COMM_WORLD, &myrank); printf("Hello world from rank %d of %d.\n", myrank, NP); MPI_Finalize(); } 9/4/2012 14

15. Parallel Programming C. Ferner & B. Wilkinson, 2014 Compiling and Running $ mpicc hello.c –o hello $ mpirun –np 5 hello Hello world from rank 0 of 5. Hello world from rank 3 of 5. Hello world from rank 4 of 5. Hello world from rank 1 of 5. Hello world from rank 2 of 5. $ mpicc is essentially gcc but makes sure that the MPI libraries are included Why are the statements not in order of rank? 9/4/2012 15

16. Parallel Programming C. Ferner & B. Wilkinson, 2014 All processes are executing the same code (although asynchronously) How can one have them execute separate code? Or how can one have a section of code executed by only one process? 9/4/2012 16

17. Parallel Programming C. Ferner & B. Wilkinson, 2014 Use their rank if (myrank >=0 && myrank < x) { … // Code executed by a subset of processes } OR if (myrank == 0) { … // Code executed by only one process } else { … // Code executed by all other processes } 9/4/2012 17 Client/Server model

18. Parallel Programming C. Ferner & B. Wilkinson, 2014 Sending messages Messages can be sent between processes using the MPI_Send() and MPI_Recv() functions. These are one-to-one communication MPI_Recv is blocking, meaning that execution will stop until the appropriate message is received. There are other non-blocking forms of communication as well as one-to-many and many-to-many. 9/4/2012 18

19. Parallel Programming C. Ferner & B. Wilkinson, 2014 2.19 Message passing concept using library routines Note each computer executes its own program

20. Parallel Programming C. Ferner & B. Wilkinson, 2014 Sending messages int MPI_Send(void *buf, int count, MPI_Datatype datatype, int dest, int tag, MPI_Comm comm) buf is the address of the data to send count is the number of elements (1 if scalar, N if an array, or strlen+1 if a string) datatype is the type of elements dest is the rank of the destination 9/4/2012 20

21. Parallel Programming C. Ferner & B. Wilkinson, 2014 Sending messages int MPI_Send(void *buf, int count, MPI_Datatype datatype, int dest, int tag, MPI_Comm comm) tag is user-defined (allows you to mark different message with your own tag). This is useful when two processors are sending multiple messages between each other. Comm is what is know as a Communicator. Basically, it is a subset of processors. MPI_COMM_WORLD is used for all processors. 9/4/2012 21

22. Parallel Programming C. Ferner & B. Wilkinson, 2014 Receiving messages int MPI_Recv(void *buf, int count, MPI_Datatype datatype, int source, int tag, MPI_Comm comm, MPI_Status *status) buf is the address in which to store the message count is the size of the buf. Can be bigger than the actual message. datatype is the type of elements source is the rank of the sender 9/4/2012 22

23. Parallel Programming C. Ferner & B. Wilkinson, 2014 Receiving messages int MPI_Recv(void *buf, int count, MPI_Datatype datatype, int source, int tag, MPI_Comm comm, MPI_Status *status) tag is user-defined Comm is the Communicator. MPI_COMM_WORLD is used for all processors. status is a structure that contains information about the transmission 9/4/2012 23

24. 24 Parameters of blocking send MPI_Send(buf, count, datatype, dest, tag, comm) Address of send buffer Number of items to send Datatype of each item Rank of destination process Message tag Communicator Notice a pointer Parameters of blocking receive Status after operation MPI_Recv(buf, count, datatype, src, tag, comm, status) Address of receive buffer Maximum number of items to receive Datatype of each item Rank of source process Message tag Communicator In our code we do not check status but good programming practice to do so. Usually send and recv counts are the same.

25. Parallel Programming C. Ferner & B. Wilkinson, 2014 MPI Datatypes (defined in mpi.h) MPI datatypes MPI_BYTE MPI_PACKED MPI_CHAR MPI_SHORT MPI_INT MPI_LONG MPI_FLOAT MPI_DOUBLE MPI_LONG_DOUBLE MPI_UNSIGNED_CHAR

26. Parallel Programming C. Ferner & B. Wilkinson, 2014 Example (Hello World 2) #include main(int argc, char **argv ) { char message[256]; int i,rank, NP, tag=99; char machine_name[256]; MPI_Status status; 9/4/2012 26

27. Parallel Programming C. Ferner & B. Wilkinson, 2014 Example (Hello World 2) MPI_Init(&argc, &argv); MPI_Comm_size(MPI_COMM_WORLD, &NP); MPI_Comm_rank(MPI_COMM_WORLD, &rank); gethostname(machine_name, 255); 9/4/2012 27

28. Parallel Programming C. Ferner & B. Wilkinson, 2014 Example (Hello World 2) if(rank == 0) { printf ("Hello world from master process %d running on %s\n", rank, machine_name); for (i = 1; i < NP; i++) { MPI_Recv(message, 256, MPI_CHAR, i, tag, MPI_COMM_WORLD, &status); printf("Message from process = %d : %s\n", i, message); } 9/4/2012 28

29. Parallel Programming C. Ferner & B. Wilkinson, 2014 Example (Hello World 2) else { sprintf(message, "Hello world from process %d running on %s", rank, machine_name); // The destination is the master process (rank 0) MPI_Send(message, strlen(message) + 1, MPI_CHAR, 0, tag, MPI_COMM_WORLD); } MPI_Finalize(); } 9/4/2012 29

30. Parallel Programming C. Ferner & B. Wilkinson, 2014 Result $ mpirun –np 9./hello Hello world from master process 0 running on compute-0-0.local Message from process = 1 : Hello world from process 1 running on compute-0-0.local Message from process = 2 : Hello world from process 2 running on compute-0-0.local Message from process = 3 : Hello world from process 3 running on compute-0-0.local Message from process = 4 : Hello world from process 4 running on compute-0-0.local Message from process = 5 : Hello world from process 5 running on compute-0-0.local Message from process = 6 : Hello world from process 6 running on compute-0-0.local Message from process = 7 : Hello world from process 7 running on compute-0-0.local Message from process = 8 : Hello world from process 8 running on compute-0-1.local $ 9/4/2012 30

31. Parallel Programming C. Ferner & B. Wilkinson, 2014 Any source or tag In the MPI_Recv, the source can be MPI_ANY_SOURCE The tag can be MPI_ANY_TAG These cause the Recv to take any message destined for the current process regardless of the source and/or regardless of the tag Ex. MPI_Recv(message, 256, MPI_CHAR, MPI_ANY_SOURCE, MPI_ANY_TAG, MPI_COMM_WORLD, &status); 9/4/2012 31

32. Parallel Programming C. Ferner & B. Wilkinson, 2014 Another Example (array) int array[100]; … // rank 0 fills the array with data if (rank == 0) MPI_Send (array, 100, MPI_INT, 1, 0, MPI_COMM_WORLD); else if (rank == 1) MPI_Recv(array, 100, MPI_INT, 0, 0, MPI_COMM_WORLD, &status); destination source tag Number of Elements 9/4/2012 32

33. Parallel Programming C. Ferner & B. Wilkinson, 2014 Another Example (scalar) int N; … // rank 2 assigns a value to N, which is needed by // rank 3 if (rank == 2) MPI_Send (&N, 1, MPI_INT, 3, 5, MPI_COMM_WORLD); else if (rank == 3) MPI_Recv(&N, 1, MPI_INT, 2, 5, MPI_COMM_WORLD, &status); dest source tag Number of elements for a scalar is only 1 9/4/2012 33

34. Parallel Programming C. Ferner & B. Wilkinson, 2014 Another Example (Ring) In the ring example, each process (except the master) receives a token from the process with rank 1 less than its own rank Then each process increments the token and sends it to the next process (with rank 1 more than its own) The last process sends the token to the master 9/4/2012 34

35. Parallel Programming C. Ferner & B. Wilkinson, 2014 Another Example (Ring) Each process (excepts the master) receives a token from the process with rank 1 less than its own rank. Then each process increments the token by 2 and sends it to the next process (with rank 1 more than its own). The last process sends the token to the master 35 Slide based upon slides from C. Ferner, UNC-W 0 17 26 35 4 Question: Do we have pattern for this?

36. Parallel Programming C. Ferner & B. Wilkinson, 2014 Another Example (Ring) #include int main (int argc, char *argv[]) { int token, NP, myrank; MPI_Status status; MPI_Init (&argc, &argv); MPI_Comm_size(MPI_COMM_WORLD, &NP); MPI_Comm_rank(MPI_COMM_WORLD, &myrank); 9/4/2012 36

37. Parallel Programming C. Ferner & B. Wilkinson, 2014 Another Example (Ring) if (myrank != 0) { // Everyone except the master receives from the processor 1 less // than its own rank. MPI_Recv(&token, 1, MPI_INT, myrank - 1, 0, MPI_COMM_WORLD, &status); printf("Process %d received token %d from process %d\n", myrank, token, myrank - 1); 9/4/2012 37

38. Parallel Programming C. Ferner & B. Wilkinson, 2014 Another Example (Ring) } else { // The master sets the initial value before sending. token = -1; } token += 2; MPI_Send(&token, 1, MPI_INT, (myrank + 1) % NP, 0, MPI_COMM_WORLD); 9/4/2012 38

39. Parallel Programming C. Ferner & B. Wilkinson, 2014 Another Example (Ring) // Now process 0 can receive from the last process. if (myrank == 0) { MPI_Recv(&token, 1, MPI_INT, NP - 1, 0, MPI_COMM_WORLD, &status); printf("Process %d received token %d from process %d\n", myrank, token, NP - 1); } MPI_Finalize(); } 9/4/2012 39

40. Parallel Programming C. Ferner & B. Wilkinson, 2014 Results (Ring) Process 1 received token 1 from process 0 Process 2 received token 3 from process 1 Process 3 received token 5 from process 2 Process 4 received token 7 from process 3 Process 5 received token 9 from process 4 Process 6 received token 11 from process 5 Process 7 received token 13 from process 6 Process 0 received token 15 from process 7 9/4/2012 40

41. Parallel Programming C. Ferner & B. Wilkinson, 2014 Send and Receive Semantics The MPI_Send and MPI_Recv are locally blocking They return after local actions are complete  MPI_Send returns after data has been copied into a buffer, message prepared and on its way  MPI_Recv returns after data has been received and buffer copied into user data (blocks until message arrives) There are other variations of send and receive (more on this later) 41 9/4/2012

42. Parallel Programming C. Ferner & B. Wilkinson, 2014 Matching up sends and recvs Notice in code how you have to be very careful matching up send’s and recv’s. Every send must have matching recv. The sends return after local actions complete but the recv will wait for the message so easy to get deadlock if written wrong Pre-implemented patterns are designed to avoid deadlock. We will look at deadlock again 2.42

43. Parallel Programming C. Ferner & B. Wilkinson, 2014 Measuring the Execution Time double MPI_Wtime( void ) Returns a double that is the number of seconds since epoch date (e.g. January 1, 1970). double start_time, end_time, elapsed_time; … start_time = MPI_Wtime(); … end_time = MPI_Wtime(); elapsed_time = end_time – start_time; Measure time to execute this section 9/4/2012 43

44. Parallel Programming C. Ferner & B. Wilkinson, 2014 Measuring the Execution Time Alternatively, one can use gettimeofday (a system call to the operating system) #include double elapsed_time; struct timeval tv1, tv2; gettimeofday(&tv1, NULL); … gettimeofday(&tv2, NULL); elapsed_time = (tv2.tv_sec - tv1.tv_sec) + ((tv2.tv_usec - tv1.tv_usec) / 1000000.0); Measure time to execute this section This is useful if you are creating a sequential (non-MPI) version with which to compare. 9/4/2012 44

45. Parallel Programming C. Ferner & B. Wilkinson, 2014 Executing program on multiple computers Usually computers specified in a file containing names of computers and possibly number of processes that should run on each computer. Then specify file with –machines option with mpiexec (or –hostfile or –f options). Implementation-specific algorithm selects computers from list to run user processes. Typically MPI would cycle through list in round robin fashion. If a machines file not specified, a default machines file used or it may only run on a single computer. 2.45

46. Parallel Programming C. Ferner & B. Wilkinson, 2014 Cluster at UNCW User Master node Compute nodes Dedicated Cluster Ethernet interface Switch Computers Compute Nodes: compute-0-0, compute-0- 1, compute-0-2, … Head Node: harpua Submit Host: babbage 9/4/2012 46

47. Parallel Programming C. Ferner & B. Wilkinson, 2014 Cluster at UNCW We use the Sun Grid Engine (SGE) to schedule jobs on the cluster This is to allow users to have exclusive use of the compute nodes so that users’ applications don’t interfere with the performance of others The scheduler (SGE) is responsible for allocating compute nodes to jobs exclusively Compile as normal: $ mpicc hello.c –o hello 9/4/2012 47

48. Parallel Programming C. Ferner & B. Wilkinson, 2014 SGE But running is done through a job submission file (or job description file) Some SGE commands:  qsub – submits a job to the schedule to run  qstat – see the status of submitted jobs (waiting, queued, running, terminated, etc.)  qdel - deletes a job (by number) from the system  qhost – see a list of hosts 9/4/2012 48

49. Parallel Programming C. Ferner & B. Wilkinson, 2014 SGE Example job submission file (hello.sge): #!/bin/sh # Usage: qsub hello.sge #$ -S /bin/sh #$ -pe orte 16 # Specify how many processors we want # -- our name --- #$ -N Hello # Name for the job #$ -l h_rt=00:01:00 # Request 1 minute to execute #$ -cwd # Make sure that the.e and.o file arrive in the working directory #$ -j y # Merge the standard out and standard error to one file mpirun -np $NSLOTS./hello 9/4/2012 49

50. Parallel Programming C. Ferner & B. Wilkinson, 2014 SGE Example job submission file (hello.sge): #!/bin/sh # Usage: qsub hello.sge #$ -S /bin/sh #$ -pe orte 16 # Specify how many processors we want 9/4/2012 50

51. Parallel Programming C. Ferner & B. Wilkinson, 2014 SGE Example job submission file (hello.sge): # -- our name --- #$ -N Hello # Name for the job #$ -l h_rt=00:01:00 # Request 1 minute to execute The name of the job plus the name of the output files: Hello.o### and Hello.op### Indicates that the job will need only a minute. This is important so that SGE will clean up if the program hangs or terminates incorrectly. May need to increase the time for longer programs or it will terminate the program before it has completed. 9/4/2012 51

52. Parallel Programming C. Ferner & B. Wilkinson, 2014 SGE Example job submission file (hello.sge): #$ -cwd # Make sure that the.e and.o file arrive in the working directory #$ -j y # Merge the standard out and standard error to one file Do the job in the current directory SGE will create 3 files: Hello.o##, Hello.e##, and Hello.op##. The –j y command will merge the Hello.o and Hello.e files (std out and error). 9/4/2012 52

53. Parallel Programming C. Ferner & B. Wilkinson, 2014 SGE Example job submission file (hello.sge): mpirun -np $NSLOTS./hello And finally the command to run the MPI program. $NSLOTS is the same number given with the #$ -pe orte 16 line. 9/4/2012 53

54. Parallel Programming C. Ferner & B. Wilkinson, 2014 SGE Example $ qstat $ qsub hello.sge Your job 106 ("Hello") has been submitted $ qstat job-ID prior name user state submit/start at queue slots ja-task-ID --------------------------------------------------------------------------------------------- -------------------- 106 0.00000 Hello cferner qw 09/04/2012 09:08:38 16 $ The state of “qw” means queued and waiting. 9/4/2012 54

55. Parallel Programming C. Ferner & B. Wilkinson, 2014 SGE Example $ qstat job-ID prior name user state submit/start at queue slots ja-task-ID --------------------------------------------------------------------------------------------- -------------------- 106 0.55500 Hello cferner r 09/04/2012 09:11:43 all.q@compute-0-0.local 16 [cferner@babbage mpi_assign]$ The state of “r” means running 9/4/2012 55

56. Parallel Programming C. Ferner & B. Wilkinson, 2014 SGE Example $ ls hello hello.c Hello.o106 Hello.po106 hello.sge ring ring.c ring.sge test test.c test.sge $ cat Hello.o106 Hello world from master process 0 running on compute-0-2.local Message from process = 1 : Hello world from process 1 running on compute-0-2.local Message from process = 2 : Hello world from process 2 running on compute-0-2.local … You will want to clean up the output files when you are done with them or you will end up with a bunch of clutter. 9/4/2012 56

57. Parallel Programming C. Ferner & B. Wilkinson, 2014 Deleting a job $ qstat job-ID prior name user state submit/start at queue slots ja-task-ID --------------------------------------------------------------------------------------------- -------------------- 108 0.00000 Hello cferner qw 09/04/2012 09:18:20 16 $ qdel 108 cferner has registered the job 108 for deletion $ qstat $ 9/4/2012 57

58. 2.58 Executing program on UNCC cluster On UNCC cci-gridgw.uncc.edu cluster, mpiexec command is mpiexec.hydra. Internal compute nodes have names used just internally. For example, a machines file to use nodes 5, 7 and 8 and the front node of the cci-grid0x cluster would be: cci-grid05 cci-grid07 cci-grid08 cci-gridgw.uncc.edu Then: mpiexec.hydra -machinefile machines -n 4./prog would run prog with four processes, one on cci-grid05, one on cci-grid07, one on cci-grid08, and one on cci- gridgw.uncc.edu.

59. 2.59 Specifying number of processes to execute on each computer Machines file can include how many processes to execute on each computer. For example: # a comment cci-grid05:2# first 2 processes on 05 cci-grid07:3# next 3 processes on 07 cci-grid08:4# next 4 processes on 08 cci-gridgw.uncc.edu:1# Last process on gridgw (09) 10 processes in total. Then: mpiexec.hydra -machinefile machines -n 10./prog If more processes were specified, they would be scheduled in round robin fashion.

60. 2.60 Eclipse IDE PTP Parallel Tools Platform plug-in Supports development of parallel programs (MPI, OpenMP). Possible to edit and execute MPI program on client or a remote machine. http://download.eclipse.org/tools/ptp/docs/ptp-sc11-slides-final.pdf Eclipse-PTP installed on the course virtual machine. Hope to explore Eclipse-PTP in assignments.

61. 2.61 Visualization Tools Programs can be watched as they are executed in a space-time diagram (or process-time diagram): Process 1 Process 2 Process 3 Time Computing Waiting Message-passing system routine Message Visualization tools available for MPI, e.g., Upshot.

62. Parallel Programming C. Ferner & B. Wilkinson, 2014 Questions? 9/4/2012 62