1. The OSCAR Cluster System Tarik Booker CS 370

2. Topics Introduction OSCAR Basics Introduction to LAM LAM Commands MPI Basics Timing Examples

3. Welcome to OSCAR! Welcome to the free Linux-based clustered system Use multiple computers to create one powerful multi-processor system

4. Account Setup Fill in the sign-in sheet Receive account password (paper slip) Log in: – Use SSH Only to log into: oscar.calstatela.edu

5. SSH (Secure Shell) Log In Use cs370studentxx as your account (where xx = your account number) student30 example:

6. Environment (LAM) Setup LAM (Local Access Minicomputer) is the implementation for MPI (Message Passing Interface). To run your parallel programs, you need to have this running.

7. Environment (LAM) Setup (2) After logging in to your account, type (assume ‘>’ is prompt): >ls You should have two files: hello.c and hosts We need to run LAM. Do this by typing: >lamboot hosts Note: to see more in-depth loading, type: >lamboot –v hosts Both methods are perfectly fine.

8. Environment (LAM) Setup (3) LAM should have taken a while to load. (We are starting a LAM process daemon on each node) After done, verify LAM is running by typing: >ps –ux This is merely a list of the processes running on your account. LAM is now setup and running on your account. (running lam process)

9. LAM Troubleshooting If anything happens with your LAM process (i.e. LAM no longer shows up on your process list) use the previous steps to start your LAM process again. If something is wrong with your LAM process (i.e. LAM is loaded, in the process list, but refuses to run, or runs indefinitely), use the “lamhalt” command, simply: >lamhalt

10. Compiling a Parallel Program Included in your account is the ‘hello.c’ program. We’ll use this as a test program for LAM/MPI. We will be using the MPI C compiler. Use the command: >mpicc hello.c This will compile your parallel program. To specify the output file, type: >mpicc hello.c -o hello This compiles hello.c into the executable called ‘hello.’

11. Running Your Parallel Program Running a program through MPI is a bit different than other interfaces. You must use the ‘mpirun’ command and specify the number of nodes used. The typical usage is: >mpirun N hello ‘hello’ is the previous executable from the last slide. The ‘N’ (UPPERCASE!) says to use all nodes. Note that we don’t have to use all nodes. Try typing: >mpirun n0-5 hello (this uses only the nodes between 0 and 5) (Also try >mpirun n4,5,6 hello)

12. The MPI Program Let’s look at hello.c The two most important functions are: – MPI_Init(MPI_COMM_WORLD); – MPI_Finalize(); These functions initialize and close the parallel environment (respectively).

13. LAM Commands LAM is our specific implementation of MPI LAM comes with additional non-MPI commands (for node management) Most not necessary, but useful

14. lamboot lamboot(hostfile) Starts LAM Environment Use –v flag for verbose boot >lamboot –v hosts

15. lamhalt Shuts down lam environment >lamhalt

16. mpirun Runs an mpi program >mpirun N hello

17. lamclean If your program terminates “badly,” use lamclean to delete old processes and allocated resources. >lamclean

18. wipe Stronger version of lamhalt that kills every node on lam >wipe

19. laminfo Detailed information list for LAM environment >laminfo

20. lamnodes List all nodes in the LAM environment >lamnodes

21. lamshrink Remove a node from the LAM environment (without rebooting) Ex: >lamshrink n3 (Note: This also invalidates node n3, or leaves an empty slot in its place)

22. lamgrow Add a node to the LAM environment (without rebooting) >lamgrow oscarnode3 Also: >lamgrow –n 3 oscarnode3 – (Adds oscarnode3 to the previously empty n3 slot. Note the space between n and 3!)

23. lamexec Run a non-MPI program in the LAM environment >lamexec {non-MPI program}

24. Termination Order of Bad Programs In the event of a bad termination (program takes up too much memory, doesn’t stop, etc.) use this order of termination: >lamclean(good) >lamhalt(better) >wipe(severe) >kill –9 [process_number](nuclear)

25. Basic MPI Functions We Covered MPI_Init and MPI_Finalize MPI_Send MPI_Recv MPI_Bcast MPI_Reduce MPI_Barrier

26. Note: MPI is a Message Passing Interface Don’t necessarily use Shared Memory Instead, information is passed around nodes

27. MPI_Send Send a variable to another node MPI_Send(variable, number of variables to send, MPI Data type, node that receives message, MPI Tag, group communicator) Ex: MPI_Send(&value, 1, MPI_INT, 2, 0, MPI_COMM_WORLD)

28. MPI_Recv Receive a variable from another node MPI_Recv(variable, number of variables to send, MPI Data type, node sending message, message tag, group communicator, MPI status indicator) Ex: – MPI_Recv(&value, 1, MPI_INT, 0, 0,MPI_COMM_WORLD, &status – (Note: You must create an MPI_Status variable when using this function)

29. MPI_Bcast Broadcasts a variable to all nodes MPI_Bcast(variable, number of variables, Data type of variable, node that sent broadcast, nodes to send messages to) Ex: – MPI_Bcast(&value, 1, MPI_INT, 0, MPI_COMM_WORLD)

30. MPI_Reduce Collect data at a node Converge information with a specific operation MPI_Reduce(variable to send, variable that receives, number of values to receive, MPI Data type, reduction operation, node receiving data, communicator to use) Ex: – MPI_Reduce(&nodePi, &pi, 1, MPI_FLOAT, MPI_SUM, 0, MPI_COMM_WORLD) There are many types of reduction operators (not only summation); you can even create your own

31. MPI_Barrier Use a barrier in MPI MPI_Barrier(communicator) Ex: – MPI_Barrier(MPI_COMM_WORLD)

32. Timing in MPI Introduction Timing functions What to do

33. Timing Intro MPI has timing features Not computational time but “Wall time” Ticks

34. Timing Functions Wall time function – double MPI_Wtime(void) Clock Tick Function – double MPI_Wtick(void)

35. What to do with timing Select starting point and store time Select end point and store time Subtract start from end, and multiply by tick – Use “%.30lf” in printf to display time instance

36. Code Example start_time = MPI_Wtime(); /* Code that does something */ end_time = (MPI_Wtime() - start_time)*tick);

37. Programming Examples Ring Arctan (Using Gregory’s formula) Pi (Using Euler’s formula) Music Program Mandelbrot Program

38. The Ring Pass a variable, one at a time, to each node in the universe (environment) Value

39. Code Example int main(int argc, char** argv) { int size, node; int value; MPI_Status status; MPI_Init(&argc, &argv); MPI_Comm_size(MPI_COMM_WORLD, &size); MPI_Comm_rank(MPI_COMM_WORLD, &node); if(node == 0) { printf("Value:"); scanf("%d", &value); MPI_Send(&value, 1, MPI_INT, node+1, 0, MPI_COMM_WORLD); } else { MPI_Recv(&value, 1, MPI_INT, node-1, 0, MPI_COMM_WORLD, &status); if(node < size - 1) { MPI_Send(&value, 1, MPI_INT, node+1, 0, MPI_COMM_WORLD); } printf("Node %d has %d in value. \n", node, value); MPI_Finalize(); return 0; }

40. Ring Code Example (2) #include int main(int argc, char** argv) { int size, node; int value; MPI_Status status; MPI_Init(&argc, &argv); MPI_Comm_size(MPI_COMM_WORLD, &size); MPI_Comm_rank(MPI_COMM_WORLD, &node);

41. Ring Code Example (3) if(node == 0) { printf("Value:"); scanf("%d", &value); MPI_Send(&value, 1, MPI_INT, node+1, 0, MPI_COMM_WORLD); } else{ MPI_Recv(&value, 1, MPI_INT, node-1, 0, MPI_COMM_WORLD, &status); if(node < size - 1) MPI_Send(&value, 1, MPI_INT, node+1, 0, MPI_COMM_WORLD); } printf("Node %d has %d in value. \n", node, value); MPI_Finalize(); return 0; } Parent node Everyone else receives All nodes but parent and last node send

42. Let’s Run Ring example…

43. Computing arctan (tan -1 ) of x Using Gregory’s Formula – arctan(x) = x - x 3 /3 + x 5 /5 - x 7 /7 + x 9 /9 - … Let’s use MPI to program this formula

44. Arctan code int main(int argc, char** argv) { int size, node; //MPI variable placeholders int i, j,x; // Loop counters double init_value; double angle = 0.0; double sum = 0.0; int terms; // Number of terms processed double finished_sum = 0.0; MPI_Status status; MPI_Init(&argc, &argv); // Start MPI environment MPI_Comm_size(MPI_COMM_WORLD, &size); //Get MPI size MPI_Comm_rank(MPI_COMM_WORLD, &node); //Get this node number

45. Arctan code (2) if(node == 0) { printf("Angle:"); scanf("%lf", &angle); printf("Number of arctan terms:"); scanf("%d", &terms); } MPI_Bcast(&terms, 1, MPI_INT, 0, MPI_COMM_WORLD); MPI_Bcast(&angle, 1, MPI_DOUBLE, 0, MPI_COMM_WORLD);

46. Arctan code (3) // Start processing arctan init_value = angle; double middle_sum; for(x=node; x<terms; x=x+size-1) { middle_sum = 0.0; double index = (double)x - 1.0; index = index + x; double temp = init_value; for(i = 0; i<(int)index - 1; ++i) temp = temp * init_value; middle_sum = temp / index; if(x % 2 == 0) middle_sum = middle_sum * -1.0; sum = sum + middle_sum; } if(node==0) sum = 0.0; MPI_Reduce(&sum, &finished_sum, 1, MPI_DOUBLE, MPI_SUM, 0, MPI_COMM_WORLD);

47. Arctan code (4) MPI_Barrier(MPI_COMM_WORLD); // Wait for all processes if(node == 0) printf(" Arctan of %lf = %.20lf\n",angle, finished_sum); MPI_Finalize(); return 0; }

48. Let’s run arctan example

49. Computing Pi Using Euler’s formula – Pi/4 = arctan(1/2) + arctan(1/3) Let’s use MPI to compute this value

50. Computing Pi (2) Arctan Code is the same – Run twice Set barrier, then compute 4 * [arctan(1/2) + arctan(1/3)]

51. Additional OSCAR Programs Created for directed studies and other classes Outside scope of this class Music Program Mandelbrot Program

52. OSCAR Music Program Uses each node’s internal speaker to play a choreographed song.

53. OSCAR Fractals Fractals use heavy computations to generate images Clustered computers are perfect for dealing with fractal generation LAM comes with XMTV, a graphics server (that runs on top of LAM)

54. The Mandelbrot Set A subset of the complex plane consisting of parameters for which the Julia set is connected

55. The Mandelbrot Set can be computed using OSCAR

56. OSCAR Cluster Has Countless Uses Any intense computation (mathematical, musical, graphical) can be solved in no time with OSCAR Cluster Check webpage for any information (questions, announcements, etc.)