Process-based Asynchronous Progress Model for MPI Communication
Switch branches/tags
Nothing to show
Clone or download
Latest commit a75cd3f Nov 14, 2017

README

                                 Casper Notes

====================================
Introduction
====================================

This project provides a portable and flexible process-based asynchronous 
progress model for MPI remote memory access (RMA) communication and 
two-sided (point-to-point) communication.


====================================
Getting Started
====================================

1. Installation

    The simplest way to configure Casper is using the following
    command-line:

        $ ./configure --prefix=/your/casper/installation/dir CC=/path/to/mpicc

    If you have an MPI installation in your path (`which mpicc`), you
    do not need to explicitly provide the CC variable:

        $ ./configure --prefix=/your/casper/installation/dir

    Once you are done configuring, you can build the source and
    install it using:

        $ make
        $ make install

2. To check if Casper built correctly, you can do the following:

        $ make check

3. To start using Casper, you need to let Casper intercept your calls
   to MPI before they reach the MPI library.

     - If your application is using a dynamic MPI library (common
       case), you can simply preload the Casper library while
       executing the application:

           $ mpiexec -genv LD_PRELOAD=/your/casper/installation/dir/lib/libcasper.so -n 4 ./test

     - If your application is using a static MPI library, you'll need
       to recompile (actually just relink) your application by placing
       the Casper symbols before the MPI symbols.  You can do so using:

           $ mpicc -o test test.c -lcasper -lmpi

       Once you have linked with Casper, you can run as normal:

           $ mpiexec -n 4 ./test

4. Execution

     - You can set the number of ghost processes per node through the
       environment variable CSP_NG, it is set to 1 by default.

       [Example 1]
       Running on 2 node, 2 ghost processes and 6 user processes on eacn node.

           $ mpiexec -genv CSP_NG=2 -np 16 -ppn 8 ./test

       [Example 2]
       Running on 4 node, 4 ghost processes and 20 user processes on eacn node.

           $ mpiexec -genv CSP_NG=4 -np 96 -ppn 24 ./test


====================================
Support
====================================
If you have problems or need any assistance about the Casper
installation and usage, please contact casper-users@lists.mpich.org
mailing list.


====================================
Bug Reporting
====================================
If you have found a bug in Casper, please contact
casper-users@lists.mpich.org mailing list.  If possible, please try to
reproduce the error with a smaller application or benchmark and send
that along in your bug report.


====================================
Casper Test Suite
====================================
Capser includes a set of testing programs located under test/. These programs
can be compiled and run via:

    $ cd test
    $ make
    $ make testing MPIEXEC="<your mpiexec> -n"

For more options to run the test suite, please check:

    $ ./test/runtest -h

====================================
Environment Variables
====================================
1. Basic Variables
    CSP_NG (integer)
    Specify the number of ghost processes per node, 1 by default.

2. Advanced Variables (Only specify them if you know what they mean)
    CSP_RUMTIME_LOAD_OPT (random|op|byte, default random)
    Specify how to distribute operations when runtime load balancing enabled,
    random by default.

    CSP_RUNTIME_LOAD_LOCK (nature|force, default nature)
    Specify how to grant lock when runtime load balancing enabled, nature
    by default.


====================================
Debugging Options
====================================
Enable debugging messages by compiling with "-DCSP_DEBUG -DCSPG_DEBUG".
    ./configure --prefix=<your installation dir> CC=<your mpicc> \
                CFLAGS="-DCSP_DEBUG -DCSPG_DEBUG"
    make && make install


====================================
Known Issues
====================================
1. Some MPI features are not or partially supported in Casper.
    a. MPI Fortran Binding
    Some MPI Fortran routines are not translated into MPI standard
    functions, thus cannot be correctly wrapped in Casper. For example,
    MPICH uses MPIR_* functions to implement following Fortran routines:
        mpi_comm_get_attr_
        mpi_comm_set_attr_
        mpi_type_get_attr_
        mpi_type_set_attr_
        mpi_win_get_attr_
        mpi_win_set_attr_
        mpi_attr_get_
        mpi_attr_put_

    b. Dynamic Process Routine

    c. No support for derived datatype in point-to-point message offloading
       routines. User must explicitly set "datatype_used=predefined"
       info at communicator creation time to enable offloading.

    d. No support for special wildcard message (Use MPI_ANY_SOURCE with
       distinct tag matching) in point-to-point message offloading routines.
       User must explicitly set "wildcard_used=none" or
       "wildcard_used=any_src|any_tag_same_tag" to enable offloading.

2. Casper currently disables user info "alloc_shared_noncontig=true" in
   MPI_Win_allocate, to avoid complex management of shared segments
   displacement on ghost processes.

3. Casper does not support MPI Error Handler Callback in C++ and Fortran
   Binding.

4. The Point-to-Point asynchronous progress routines are not thread-safe.