Intro to MPI

Message Passing Interface
- This is a library definition.
- It is language independent
  - C, C++, FORTRAN
  - Javascript?, python, java
  - But interfaces for other languages as well.
- Early draft for MPI1 in 1992, official release 1994.
MPI is a communication protocol
- For parallel computers.
- Consider Bridges-2
  - 488 nodes
    - 2x 64 core CPUs
    - 256 GB of memory each.
    - 3.7 TB SSD
    - 16 more of these with (Memory Nodes)
      - 512 GB memory
    - 4 more with
      - 4x 24 core cpu
      - 4 TB main memory
      - 7TB SSD
    - 34 GPU nodes
      - 24 with 8 Nvidia Tesla
      - 9 with 8 Nvidia V100
      - 1 with 16 Nvidial Volta
      - With various configurations
  - This machine supports MPI
  - Let's quickly check out the Top 500 list
    - What is a Flop/s ?
    - What is a TFlop?
    - What is an EFlop ? Note the current performance record is 7.84 EFlops/s
- MPI has been through a number of revisions. Currently MPI-3
- There are several MPI implimentations
  - mpich is from Argonne National Lab and Mississippi State.
  - Fronteer runs Cray MPI, a version of MPICH
  - OPEN MPI is the merger of several MPI projects.
- I spent way too much time looking at the above.
- For me
  - I am using mpich.
  - Just pop this open in another window"
  - We are in Argonne's "zone".
  - And The Ohio State University
  - I may go to one of the C++ interfaces, but I still sort of know the C interface.
  - I will start with c.
  - It has been a while, so I have stolen these first two examples.
  - Also note, I might be a bit old-fashion in my usage.
- A MPI Implementation gives you
  - A set of header files (/usr/include/mpich-x86_64/)
  - A set of libraries (/usr/lib64/mpich/lib/)
  - A bunch of executables (/usr/lib64/mpich/bin)
    - Compilers
      - mpicc
      - mpic++
      - mpif77, mpif90, ...
    - Other utilities
      - mpirun
  - Man Pages (/usr/lib64/mpich/share/man)
- Look at my Makefile.
  - I add some flags (but I don't think this is necessary, but I want to show it)
  - I change the default compiler. (This is necessary)
- To run the process
  - I need all machines upon which I wish to run this computatin configured.
    - Running the MPI daemon
    - Probably Private/Public or better key exchanges done.
    - Or possibly on a completely trusted network.
  - I might need to submit to a batch job runner
  - Or I need to run with
    - mpiexec args executable pgmargs [ : args executable pgmargs ... ]
    - ```
     mpiexec -n  10 hello 
```
- Let's take a look at hello.cpp
  - int MPI_Init(int *argc, char ***argv)
    - ARG, a triple pointer reference.
    - These can be nullptr.
    - You should do as little as possible before you do this.
  - This was designed in the heyday of POSIX stuff so the error is similar.
    - The call returns an integer which indicates
    - Success: MPI_SUCCESS
    - otherwise.
    - int MPI_Error_string(int errorcode, char *string, int *resultlen)
  - int MPI_Comm_size(MPI_Comm comm, int *size)
    - This returns the number of proceses in this group.
    - Think of the group as a Orginizational unit.
    - For now MPI_COMM_WORLD is the collection of all processes.
      - We may build other groups later.
    - This will tell us the number of processes in the computation.
  - int MPI_Comm_rank(MPI_Comm comm, int *rank)
    - Each process has a unique "id" or rank in each of its communication groups.
    - This is used for many things.
  - int MPI_Get_processor_name(char *name, int *resultlen)
  - MPI_Finalize(void)
- Look at first.cpp
  - int MPI_Send(const void *buf, int count, MPI_Datatype datatype, int dest, int tag, MPI_Comm comm)
  - int MPI_Recv(void *buf, int count, MPI_Datatype datatype, int source, int tag, MPI_Comm comm, MPI_Status *status)
    - MPI_ANY_SOURCE
    - MPI_ANY_TAG
  - The tag is sort of like the name on an evelope.
    - It was sent to this address but why.
    - This will possibly allow me to customize when I receive a message.