DLB is a dynamic library designed to speed up HPC hybrid applications (i.e., two levels of parallelism) by improving the load balance of the outer level of parallelism (e.g., MPI) by dynamically redistributing the computational resources at the inner level of parallelism (e.g., OpenMP). at run time.
This dynamism allows DLB to react to different sources of imbalance: Algorithm, data, hardware architecture and resource availability among others.
LeWI (Lend When Idle) is the algorithm used to redistribute the computational resources that are not being used from one process to another process inside the same shared memory node in order to speed up its execution.
DROM (Dynamic Resource Ownership Manager) is the algorithm used to manage the CPU affinity of a process running a shared memory programming model (e.g., OpenMP).
TALP (Tracking Application Live Performance) is the module used to gather performance data from the application. The data can be obtained during the execution or as a report at the end.
-
Build requirements
- A supported platform running GNU/Linux (i386, x86-64, ARM, PowerPC or IA64)
- C compiler
- Python 2.4 or higher
- GNU Autotools, only needed if you want to build from the repository.
-
Download the DLB source code:
- Either from our website: DLB Downloads.
- Or from a git repository
-
Clone DLB repository
-
From GitHub:
git clone https://github.com/bsc-pm/dlb.git
-
From our internal GitLab repository (BSC users only):
git clone https://pm.bsc.es/gitlab/dlb/dlb.git
-
-
Or download from GitHub releases
-
Bootstrap autotools:
cd dlb ./bootstrap
-
-
Run
configure
. Optionally, check the configure flags by running./configure -h
to see detailed information about some features. MPI support must be enabled with--with-mpi
and, optionally, an argument telling where MPI can be located../configure --prefix=<DLB_PREFIX> [<configure-flags>]
-
Build and install
make make install
-
Optionally, add the installed bin directory to your
PATH
export PATH=<DLB_PREFIX>/bin:$PATH
For more information about the autotools installation process, please refer to INSTALL
Choose between linking or preloading the binary with the DLB shared library
libdlb.so
and configure DLB using the environment variable DLB_ARGS
.
-
Example 1: Share CPUs between MPI processes
# Link application with DLB mpicc -o myapp myapp.c -L<DLB_PREFIX>/lib -ldlb -Wl,-rpath,<DLB_PREFIX>/lib # Launch MPI as usual, each process will dynamically adjust the number of threads export DLB_ARGS="--lewi" mpirun -n <np> ./myapp
-
Example 2: Share CPUs between MPI processes with advanced affinity control through OMPT.
# Link application with an OMPT capable OpenMP runtime OMPI_CC=clang mpicc -o myapp myapp.c -fopenmp # Launch application: # * Set environment variables # * DLB library is preloaded # * Run application with binary dlb_run export DLB_ARGS="--lewi --ompt" export OMP_WAIT_POLICY="passive" preload="<DLB_PREFIX>/lib/libdlb.so" mpirun -n <np> <DLB_PREFIX>/bin/dlb_run env LD_PRELOAD="$preload" ./myapp
-
Example 3: Manually reduce assigned CPUs to an OpenMP process.
# Launch an application preloading DLB export OMP_NUM_THREADS=4 export DLB_ARGS="--drom" export LD_PRELOAD=<DLB_PREFIX>/lib/libdlb.so taskset -c 0-3 ./myapp & # Reduce CPU binding to [1,3] and threads to 2 myapp_pid=$! dlb_taskset -p $myapp_pid -c 1,3
-
Example 4: Get a TALP summary report at the end of an execution
export DLB_ARGS="--talp --talp-summary=pop-metrics" PRELOAD=<DLB_PREFIX>/lib/libdlb_mpi.so mpirun <opts> env LD_PRELOAD="$PRELOAD" ./app
Please refer to our DLB User Guide for a more complete documentation.
If you want to cite DLB, you can use the following publications:
-
Hints to improve automatic load balancing with LeWI for hybrid applications at Journal of Parallel and Distributed Computing 2014. (ScienceDirect) (bibtex) (pdf)
-
LeWI: A Runtime Balancing Algorithm for Nested Parallelism at International Conference in Parallel Processing 2009, ICPP09. (IEEE Xplore) (bibtex) (pdf)
For questions, suggestions and bug reports, you can contact us via e-mail at dlb@bsc.es.