Contents of the Package

Structure

dlb/
├── bin
├── include
├── lib
└── share
    ├── doc
    │   └── dlb
    │       └── examples
    │           ├── MPI+OMP
    │           ├── MPI+OmpSs
    │           └── statistics
    ├── man
    │   └── man3
    └── paraver_cfgs
        └── DLB

Binaries

dlb
Basic info, help and version
dlb_shm
Utility to manage shared memory
dlb_taskset
Utility to change the process mask of DLB processes
dlb_cpu_usage
Python viewer if using stats

Libraries

DLB installs different versions of the library for different situations, but in general you should only focus on these ones:

libdlb.so
Link against this library only if your application needs to call some DLB API function.
libdlb_mpi.so
Link in the correct order or just preload it using LD_PRELOAD environment variable if you want DLB to intercept MPI calls.
libdlb_mpif.so
Same case as above, but to intercept MPI Fortran calls.

Remember that if the programming model already supports DLB (as in Nanos++), you don’t need to link against any library.

Examples

Currently three examples are distributed with the source code and installed in the ${DLB_PREFIX}/share/doc/dlb/examples/ directory. Each example consists of a README file with a brief description and the steps to follow, a C source code file, a Makefile to compile the source code and a script run.sh to run the example.

Some Makefile variables have been filled at configure time. They should be enough to compile and link the examples with MPI support. Some Makefiles assume that Mercurium is configured in the PATH.

Note

In order to enable tracing you need an Extrae installation and to correctly set the EXTRAE_HOME environment variable.

MPI + OpenMP

PILS is a synthetic MPI program with some predefined load balancing issues. Simply check the Makefile if everything is correct and run make. The run.sh script should also contain the MPI detected at configure time. Apart from that, two options can be modified at the top of the file, whether you want to enable DLB or to enable TRACE mode (or both).

A very similar example but just using OpenMP. Notable differences are the -fopenmp flag used in the Makefile that assumes a GNU-like flag. The run.sh script is also configured to allow two options, DLB and TRACE.

MPI + OmpSs

A very similar example but just using OmpSs. Make sure that Mercurium is in your PATH or modify the Makefile accordingly. Then, you can run it in the same way as the previous example.

Statistics

The last example consists of a PILS program designed to run for a long time, without DLB micro-load balancing, but with the Statistics module enabled. Check the run.sh script. The objective is to let the process run in background while you run one of the other two binaries provided. These two binaries get_pid_list and get_cpu_usage perform basic queries to the first PILS program and obtain some statistics about CPU usage.