4 AMPI is an implementation of the MPI standard written on top of Charm++, meant
5 to give MPI applications access to high-level, application-independent features
6 such as overdecomposition (processor virtualization), dynamic load balancing,
7 automatic fault tolerance, and overlap of computation and communication. For
8 more information on all topics related to AMPI, consult the AMPI manual here:
10 http://charm.cs.illinois.edu/manuals/html/ampi/manual.html
15 AMPI has its own target in the build system. You can run the top-level
16 build script interactively using "./build", or you can specify your
17 architecture, operating system, compilers, and other options directly.
20 ./build AMPI netlrts-linux-x86_64 gfortran gcc --with-production
23 Compiling and Linking AMPI Programs
24 -----------------------------------
25 AMPI source files can be compiled and linked with the wrappers found
26 in bin/, such as ampicc, ampicxx, ampif77, and ampif90, or with
27 "charmc -language ampi". For example:
29 ampif90 pgm.f90 -o pgm
31 To enable transparent migration of user heap data, link with
32 "-memory isomalloc". To perform dynamic load balancing, link in a Charm++
33 load balancer (suite) using "-module <LB>". For example:
35 ampicc pgm.c -o pgm -memory isomalloc -module CommonLBs
37 Note that you need to specify a Fortran compiler when building Charm++/AMPI
38 for Fortran compilation to work.
43 AMPI programs can be run with charmrun like any other Charm++ program. In
44 addition to the number of processes, specified with "+p n", AMPI programs
45 also take the total number of virtual processors (VPs) to run with as "+vp n".
46 For example, to run an AMPI program 'pgm' on 4 processors using 32 ranks, do:
48 ./charmrun +p 4 ./pgm +vp 32
50 To run with dynamic load balancing, add "+balancer <LB>":
52 ./charmrun +p 4 ./pgm +vp 32 +balancer RefineLB
57 Global and static variables are unsafe for use in virtualized AMPI programs.
58 This is because globals are defined at the process level, and AMPI ranks are
59 implemented as user-level threads, which may share a process with other ranks
60 Therefore, to run with more than 1 VP per processor, all globals and statics
61 that are non-readonly and whose value does not depend on rank must be modified
62 to use local storage. Consult the AMPI manual for more information on global
63 variable privatization and automated approaches to privatization.
65 AMPI programs must have the following main function signatures, so that AMPI
66 can bootstrap before invoking the user's main function:
67 * C/C++ programs should use "int main(int argc, char **argv)"
68 * Fortran programs should use "Subroutine MPI_Main" instead of
72 Incompatibilities and Extensions
73 --------------------------------
74 AMPI has some known flaws and incompatibilities with other MPI implementations:
75 * MPI_Cancel does not actually cancel pending communication.
76 * MPI_Sendrecv_replace gives incorrect results.
77 * Persistent sends with Irsend don't work.
78 * Isend/Irecv do not work when using MPI_LONG_DOUBLE.
79 * MPI_Get_elements returns the expected number of elements instead of the
80 actual number received.
81 * MPI_Unpack gives incorrect results.
82 * Data alignment in user defined types does not match the MPI standard.
83 * Scatter/gather using noncontiguous types gives incorrect results.
84 * Datatypes are not reused, freed, or reference counted.
85 * The PMPI profiling interface is not implemented in AMPI.
87 AMPI also has extensions to the MPI standard to enable use of the high-level
88 features provided by the Charm++ adaptive runtime system. All extensions are
90 * AMPI_Migrate tells the runtime system that the application has reached a
91 point at which the runtime system may serialize and migrate ranks.
92 It is used for dynamic load balancing and fault tolerance. See the AMPI
93 manual for more information on how to use it.
94 * AMPI_Register_pup is used to register PUP routines and user data.
95 * AMPI_Get_pup_data returns a pointer to user data managed by the runtime.
96 * AMPI_Register_main is used to register multiple AMPI modules.
97 * AMPI_Load_set_value sets the calling rank's load to the given user value.
98 * AMPI_Load_start_measure starts load balance information collection.
99 * AMPI_Load_stop_measure stops load balance information collection.
100 * AMPI_Migrate_to_pe migrates the calling rank to the given PE.
101 * AMPI_Comm_set_migratable sets the migratability of the given communicator.
102 * AMPI_Command_argument_count returns the number of command line arguments
103 given to a Fortran AMPI program excluding charmrun and AMPI parameters.
104 * AMPI_Get_command_argument returns an argument from the command line
105 to a Fortran AMPI program.
107 Note that AMPI defines a preprocessor symbol "AMPI" so that user codes can
108 check for AMPI's presence at compile time using "#ifdef AMPI".