|
DisCostiC v2.0.0
Distributed Cost in Cluster: a cross-architecture simulator for massively parallel MPI programs
|
First, use the following command to clone the git repository:
Before proceeding, make sure the environment is prepared for the compilation. The installation steps are listed below:
One way to check the installation is to print the version and the help of the DisCostiC using ./discostic --version and ./discostic --help.
β±οΈ Configuration settings π
The test folder holds one header per benchmark and kernel mode. Run_Simulation.sh copies test/<benchmark_kernel>.hpp to test/P2P.hpp, so benchmark_kernel is the header name without .hpp. The shipped config.cfg sets benchmark_kernel = STREAM_COLLECTIVE. Examples of headers:
Header name (benchmark_kernel) | Description |
|---|---|
STREAM_COLLECTIVE | STREAM compute plus an MPI collective |
STREAM_FILE, STREAM_LBL, STREAM_SRC, STREAM_COMP | STREAM triad in one kernel mode |
HEAT_FILE | Two-dimensional five-point Jacobi kernel |
SOR_LBL | Gauss-Seidel successive over-relaxation |
HEATHEAT_FILE, HEATSOR_FILE, HEATDIVIDE_FILE | Back-to-back kernels |
TWOKERNEL, PITRIADFILE, PITRIADLBL, RTOF4 | Multi-kernel workflows |
HPCG | High Performance Conjugate Gradient |
kernel_mode changes how the kernel time is obtained. FILE and SRC run Kerncraft. LBL and COMP use recorded ECM numbers and do not. Further Exec strings are in Essential DisCostiC routines. Modes other than these four are in Modes.
kernel_mode | Integrated tool | Description |
|---|---|---|
COMP | no external tool integrated | This directly embeds the single-core pre-recorded ECM performance model data of the computational kernel into the simulator. |
LBL | no external tool integrated | This reads the pre-recorded ECM performance model data for the computational kernel from an external file located at nodelevel/configs folder. |
SRC | Kerncraft integrated | This directly embeds the source code of the computational kernel into the simulator. |
FILE | Kerncraft integrated | This reads the source code for the computational kernel from an external file located at nodelevel/kernels folder. |
Some *_LBL.hpp headers read message_size from config.cfg and use it as the byte count of the point-to-point exchange. step_size, direction, and periodic are not read by the headers on this tree.
The shipped config.cfg selects the header with benchmark_kernel. FILE and SRC headers run Kerncraft. LBL and COMP headers use recorded ECM numbers. The keys ./discostic --help actually prints are listed in DisCostiC help.
The compilation offers the following choices to enable the output report's data format generation and to enable the tracing of the simulator's own implementation. An executable will be generated after compilation.
| Command | Description |
|---|---|
make | This enables JSON data format without ITAC profiling of simulator implementation. |
make otf2 | This enables both JSON and OTF2 data format without ITAC profiling of simulator implementation. |
make trace_MPI | This enables JSON data format and the standard ITAC tracing mode of simulator implementation with the information about MPI call functions (enabled flag: -trace). |
make trace_all | This enables JSON data format and the verbose ITAC tracing mode of simulator implementation with the information on both user-defined and MPI call functions (enabled flag: -trace -tcollect). |
make stf does not rebuild the executable. It converts the modeled-application ITAC file DisCosTiC.avt to DisCosTiC.single.stf. See ITAC output.
Run_Simulation.sh reads config.cfg in place and launches number_of_processes + 1 ranks on Fritz (singlenode). Submit it from the repository root. The allocation must be at least that many tasks.
| Command | Description |
|---|---|
sbatch Run_Simulation.sh | This runs the simulation described by the current config.cfg. |
| Command | Description |
|---|---|
make clear | This cleans up the working directory by removing all unnecessary DisCostiC files, such as *.dms, *.otf, *.csv files. |
make uninstall | This uninstalls the DisCostiC framework, including installed files and CMake-specific files. |
This work is funded by the DFG MOD4COMP RESEARCH UNIT "Holistic Energy and Performance Modeling for Sustainable Computing".

@AyeshaAfzal91
Ayesha Afzal, Erlangen National High Performance Computing Center (NHR@FAU)
mailto: ayesha.afzal@fau.de