StarPlat is designed to make parallel computing easier for users. It reduces the need for complex parallelization instructions while still delivering performance close to hand-optimized code. When choosing between simplicity and performance, StarPlat prioritizes simplicity because it is mainly for domain experts.
StarPlat provides key features and data types for static graph algorithms, including Graph, Node, Edge, and PropNode (for node properties). Users write in a procedural style and suggest parallelization opportunities using constructs like forall. The compiler then decides how to apply parallel execution, including nested forall loops. Many graph algorithms involve repeated execution until they reach a stopping condition, which StarPlat manages with the fixedPointconstruct.
For handling updates efficiently, StarPlat includes batch processing constructs such as Batch, OnDelete (for removing edges), and OnAdd (for adding edges). The Batch construct defines preprocessing steps before executing these handlers. In incremental processing, only changes from added edges are handled, while decremental processing deals with edge deletions.
To run a code in StarPlat, you have to follow the following flow,
DSL → Code Generation → Code Compilation → Running code for different backends.
Note: Try to setup ssh in vscode instead of terminal it will save more time.
To begin, you first need to write the DSL code. Keep the following points in mind:
- Do not include an extension in the DSL file name. For example, if you're writing a DSL file for
triangle_count, name the filetriangle_count, nottriangle_count.ext. - Store static graph codes in the static graph directory and dynamic graph codes in the dynamic graph directory for better organization.
Aquais recommended to run StarPlat. This documentation will follow theAquaworkflow.
Note: you can find DSL code walkthrough (by Nibedita) here.
Now the DSL code is ready to us. So to Compile we have to go **StarPlat/src.**
And Inside this directory you have to run the following commands,
make clean
make
Note: you have to run this commands on first time of using StarPlat or if you have done any changes to StarPlat not for everytime of writing DSL.
Now, to generate code to a desired backend you have run the following command inside same directory,
./StarPlat [-s|-d] -f <dsl.sp(DSL code file path)> -b [cuda|omp|mpi|acc|sycl]
Say we want to generate static code of triangle count for OpenMP backend. So we have to write,
./StarPlat -s -f /lfs/usrhome/mtech/cs23m006/new_repo/starplat/graphcode/staticDSLCodes/triangle_counting_dsl -b omp
Note: Always try to use absolute path not relative path.
If you face any segmentation fault or any error what you need to debug you may use gdb (this is while code generation). It is better to run those three commands separately.
gdb ./Starplat
run -s -f <file_path> -b cuda|omp|mpi
btTo compile code you have to write a main function where you have to include the generated code file from the generated_$backend folder here.
For OpenMP the following main.cpp is shown for triangle count. You can just change the header file name everytime when you compile.
#include <iostream>
#include <fstream>
#include <sstream>
#include <vector>
#include <set>
#include <map>
#include <algorithm>
#include <cstring>
#include <climits>
#include "/lfs/usrhome/mtech/cs23m006/new_repo/starplat/graphcode/generated_omp/triangle_counting_dsl.cc"
int main(int argc, char*argv[]) {
char* filePath;
if (argc == 1) {
std::string inputPath;
std::cout << "Enter the path to the graph file: ";
std::getline(std::cin, inputPath);
filePath = new char[inputPath.length() + 1];
std::strcpy(filePath, inputPath.c_str());
} else if (argc == 2) {
filePath = argv[1];
} else {
return 1;
}
graph g(filePath);
g.parseGraph();
std::cout << "Number of nodes: " << g.num_nodes() << std::endl;
std::cout << "Number of edges: " << g.num_edges() << std::endl;
double starttime = omp_get_wtime();
long num_triangles = Compute_TC(g);
double endtime = omp_get_wtime();
std::cout << "Number of triangles in the graph: " << num_triangles << std::endl;
std::cout<<"Time taken : "<<endtime-starttime<<std::endl;
return 0;
}For OpenMP you can use both terminal and qsub to execute the main.cpp file.
This is the command to compile the OpenMP code:
g++ main.cpp -fopenmp -o main
For MPI the following main.cpp is shown for triangle count. You can just change the header file name everytime when you compile.
#include <iostream>
#include "/lfs/usrhome/oth/rnintern/cs23m006/starplat_new/starplat/graphcode/mpi_header/graph_mpi.h"
#include "/lfs/usrhome/oth/rnintern/cs23m006/starplat_new/starplat/graphcode/mpi_header/profileAndDebug/mpi_debug.c"
#include"/lfs/usrhome/oth/rnintern/cs23m006/starplat_new/starplat/graphcode/generated_mpi/triangle_counting_dsl.cc"
long Compute_TC(Graph&, boost::mpi::communicator);
int main(int argc, char *argv[]) {
char* filePath;
boost::mpi::environment env(argc, argv);
boost::mpi::communicator world;
if (argc == 1) {
std::string inputPath;
std::cout << "Enter the path to the graph file: ";
std::getline(std::cin, inputPath);
filePath = new char[inputPath.length() + 1];
std::strcpy(filePath, inputPath.c_str());
} else if (argc >=2 ) {
filePath = argv[1];
} else {
return 1;
}
Graph graph(argv[1],world,1);
double t1 = MPI_Wtime();
auto ans=Compute_TC(graph,graph.num_nodes(),world);
std::cout << "No of nodes " << graph.num_nodes() << std::endl;
double t2 = MPI_Wtime();
print_mpi_statistics();
if (world.rank() == 0)
{
std::cout << "Total Triangle: " << ans << std::endl;
std::cout << "TIME:[" << (t2 - t1) << "]" << std::endl;
}
world.barrier();
return 0;
}For Cuda the following main.cu is shown for triangle count. You can just change the header file name everytime when you compile.
Note: here main file name should be main.cu not main.cpp.
This is for loading right cuda version.
#module unload $current_gcc (if other than gcc920)
module load gcc920
module unload cuda10.1
module load cuda11.4This is main.cu file sample snippet.
#include <cuda_runtime.h>
#include <cuda.h>
#include <iostream>
#include "/lfs/usrhome/mtech/cs23m006/new_repo/starplat/graphcode/generated_cuda/APFB_Matching.cu"
int main(int argc, char* argv[]) {
char* filePath;
char* updateFile;
if (argc == 2) {
filePath = argv[1];
} else if (argc==3){
filePath = argv[1];
updateFile=argv[2];
}else{
return 1;
}
graph g(filePath);
g.parseGraph();
std::cout << "Number of nodes: " << g.num_nodes() << std::endl;
std::cout << "Number of edges: " << g.num_edges() << std::endl;
cudaEvent_t start, stop;
cudaEventCreate(&start);
cudaEventCreate(&stop);
cudaEventRecord(start, 0);
long ans = Compute_TC(g, g.num_nodes());
cudaEventRecord(stop, 0);
cudaEventSynchronize(stop);
float milliseconds = 0;
cudaEventElapsedTime(&milliseconds, start, stop);
std::cout << "Total Trinagles: " << ans << std::endl;
std::cout << "\nGPU Time: " << milliseconds / 1000.0 << " seconds" << std::endl;
cudaEventDestroy(start);
cudaEventDestroy(stop);
return 0;
}All the graphs are present here,
/lfs1/usrscratch/phd/cs16d003/11suiteDSL/
If you are using the terminal use the following command,
export OMP_NUM_THREADS=16
main $input_file $update_path(if any) > $output_file_path
if you are using qsub to submit jobs, then you following script for the OpenMP can be followed,
#!/bin/bash
#PBS -o logfile.log
#PBS -e errorfile_slash.err
#PBS -l walltime=24:00:00
#PBS -l select=1:ncpus=32
tpdir=`echo $PBS_JOBID | cut -f 1 -d .`
tempdir=$HOME/scratch/job$tpdir
mkdir -p $tempdir
cd $tempdir
cp -R $PBS_O_WORKDIR/* .
export OMP_NUM_THREADS=16
/lfs/usrhome/mtech/cs23m006/new_repo/starplat/graphcode/main /lfs/usrhome/mtech/cs23m006/new_repo/starplat/graphcode/sample_graph_loaded/cleancom-orkutud.txt /lfs/usrhome/mtech/cs23m006/new_repo/starplat/graphcode/sample_graph_loaded/update_orkut.txt > apfb_dyn_orkut_16_0.04.txt
mv ../job$tpdir $PBS_O_WORKDIR/.
Note:
- Add this line
#PBS -q rupesh_gpuqif you are have the permission to use the queue from Prof. Rupesh Nasre. - Don’t use
ncpusvalue other than 32 as there is a high chance that the script will not work. Also we don’t need gpus for OpenMP so we are not using it. walltimecan be reduced.
For MPI we are using rnintern account (access should be given by Rupesh sir).
To compile static code we have to use the following command,
mpicxx -I/lfs/usrhome/oth/rnintern/scratch/MPI_Comparison/boost/install_dir/include $(path to main.cpp) /lfs/usrhome/oth/rnintern/cs23m006/starplat_new/starplat/graphcode/mpi_header/graph_mpi.cc $(add all the headers are using) ../boost/install_dir/lib/libboost_mpi.a ../boost/install_dir/lib/libboost_serialization.a -o main
As the example of above code snippet I am giving an example of APFB matching. In APFB , I am using the following header rma_datatype.cc, edge_property.cc, node_property.cc . So the snippet will look like,
mpicxx -I/lfs/usrhome/oth/rnintern/scratch/MPI_Comparison/boost/install_dir/include /lfs/usrhome/oth/rnintern/cs23m006/starplat_new/starplat/graphcode/main.cpp /lfs/usrhome/oth/rnintern/cs23m006/starplat_new/starplat/graphcode/mpi_header/graph_mpi.cc /lfs/usrhome/oth/rnintern/cs23m006/starplat_new/starplat/graphcode/mpi_header/rma_datatype/rma_datatype.cc /lfs/usrhome/oth/rnintern/cs23m006/starplat_new/starplat/graphcode/mpi_header/graph_properties/edge_property/edge_property.cc /lfs/usrhome/oth/rnintern/cs23m006/starplat_new/starplat/graphcode/mpi_header/graph_properties/node_property/node_property.cc ../boost/install_dir/lib/libboost_mpi.a ../boost/install_dir/lib/libboost_serialization.a -o main
Note: You can use the same snippet for dynamic code by changing header only.
Now to submit a job for MPI using rnintern account you can use the following script.
#!/bin/bash
#PBS -o logfile.log
#PBS -e errorfile_slash.err
#PBS -l walltime=01:00:00
#PBS -l select=1:ncpus=1:ngpus=1
#PBS -q rupesh_gpuq
module load openmpi316
module unload gcc920
module load gcc13.1.0
#tpdir=`echo $PBS_JOBID | cut -f 1 -d .`
#tempdir=$HOME/scratch/job$tpdir
#mkdir -p $tempdir
#cd $tempdir
#mpicxx -I/lfs/usrhome/oth/rnintern/scratch/MPI_Comparison/boost/install_dir/include /lfs/usrhome/oth/rnintern/cs23m006/starplat_new/starplat/graphcode/main.cpp ../boost/install_dir/lib/libboost_mpi.a ../boost/install_dir/lib/libboost_serialization.a -o $PBS_O_WORKDIR/main_exe
/lfs/sware/openmpi316/bin/mpirun -np 2 -hostfile $PBS_NODEFILE /lfs/usrhome/oth/rnintern/scratch/MPI_Comparison/boost/main /lfs1/usrscratch/phd/cs16d003/11suiteDSL/cleaned-graphs/cleancom-orkutud.txt /lfs/usrhome/oth/rnintern/cs23m006/update_sample.txt > $PBS_O_WORKDIR/output_.txt
#mv ../job$tpdir $PBS_O_WORKDIR/.
The script for aqua for cuda code is below.
#PBS -e errorfile.err
#PBS -o logfile.log
#PBS -l walltime=01:00:00
#PBS -l select=1:ncpus=1:ngpus=1
#PBS -q rupesh_gpuq
tpdir=`echo $PBS_JOBID | cut -f 1 -d .`
tempdir=$HOME/scratch/job$tpdir
mkdir -p $tempdir
cd $tempdir
cp -R $PBS_O_WORKDIR/* .
module load gcc920
module load cuda11.4
module load python385
./a.out < input.txt > output.txt
mv * $PBS_O_WORKDIR/.
rmdir $tempdir- To save time, you can use an alias and execute an initial command before proceeding with the rest. I have given an example of my my
.bsh.txtfile.
alias output_dir="cd /lfs/usrhome/oth/rnintern/cs23m006"
alias boost_wd="cd /lfs/usrhome/oth/rnintern/scratch/MPI_Comparison/boost"
alias code_wd="cd /lfs/usrhome/oth/rnintern/cs23m006/starplat_new/starplat"
module load openmpi316
alias qs="qsub /lfs/usrhome/oth/rnintern/cs23m006/script.cmd"
alias qr="qstat rupesh_gpuq"
alias clr="clear"
alias mpimain="mpicxx -I/lfs/usrhome/oth/rnintern/scratch/MPI_Comparison/boost/install_dir/include /lfs/usrhome/oth/rnintern/cs23m006/starplat_new/starplat/graphcode/main.cpp /lfs/usrhome/oth/rnintern/cs23m006/starplat_new/starplat/graphcode/mpi_header/graph_mpi.cc /lfs/usrhome/oth/rnintern/cs23m006/starplat_new/starplat/graphcode/mpi_header/rma_datatype/rma_datatype.cc /lfs/usrhome/oth/rnintern/cs23m006/starplat_new/starplat/graphcode/mpi_header/graph_properties/edge_property/edge_property.cc /lfs/usrhome/oth/rnintern/cs23m006/starplat_new/starplat/graphcode/mpi_header/graph_properties/node_property/node_property.cc ../boost/install_dir/lib/libboost_mpi.a ../boost/install_dir/lib/libboost_serialization.a -o main"
alias mpimaindyn="mpicxx -I/lfs/usrhome/oth/rnintern/scratch/MPI_Comparison/boost/install_dir/include /lfs/usrhome/oth/rnintern/cs23m006/starplat_new/starplat/graphcode/mainDyn.cpp /lfs/usrhome/oth/rnintern/cs23m006/starplat_new/starplat/graphcode/mpi_header/graph_mpi.cc /lfs/usrhome/oth/rnintern/cs23m006/starplat_new/starplat/graphcode/mpi_header/updates.cc /lfs/usrhome/oth/rnintern/cs23m006/starplat_new/starplat/graphcode/mpi_header/rma_datatype/rma_datatype.cc /lfs/usrhome/oth/rnintern/cs23m006/starplat_new/starplat/graphcode/mpi_header/graph_properties/edge_property/edge_property.cc /lfs/usrhome/oth/rnintern/cs23m006/starplat_new/starplat/graphcode/mpi_header/graph_properties/node_property/node_property.cc ../boost/install_dir/lib/libboost_mpi.a ../boost/install_dir/lib/libboost_serialization.a -o main"
alias graph_dir="cd /lfs1/usrscratch/phd/cs16d003/11suiteDSL/"
alias code_mpi="cd /lfs/usrhome/oth/rnintern/cs23m006/starplat_new/starplat/graphcode/generated_mpi"
then just go the the directory of bsh.txt , and run . bsh.txt
- To know the available module in
Aquawe can use the following command,module avail. Thenmodule load $pkg_name,if we want to remove a module we can usemodule unload $pkg_name.