feat: add containr Slurm job and docs

# Conflicts:
#	GPUSimulators/Simulator.py

.gitignore

@@ -186,7 +186,6 @@ cython_debug/
 .pypirc
 
 # CUDA
-cuda_cache/
 
 # Taken from: https://github.com/github/gitignore/blob/main/CUDA.gitignore
 *.i

GPUSimulators/Common.py

@@ -35,6 +35,8 @@ import gc
 import netCDF4
 import json
 
+from tqdm import tqdm
+
 #import pycuda.compiler as cuda_compiler
 #import pycuda.gpuarray
 #import pycuda.driver as cuda
@@ -178,11 +180,11 @@ def runSimulation(simulator, simulator_args, outfile, save_times, save_var_names
         profiling_data_sim_runner["end"]["t_sim_init"] = time.time()
 
         #Start simulation loop
-        progress_printer = ProgressPrinter(save_times[-1], print_every=10)
+        # progress_printer = ProgressPrinter(save_times[-1], print_every=10)
-        for k in range(len(save_times)):
+        for k, t_step in tqdm(enumerate(t_steps), desc="Simulation Loop"):
             #Get target time and step size there
-            t_step = t_steps[k]
+            # t_step = t_steps[k]
-            t_end = save_times[k]
+            # t_end = save_times[k]
             
             #Sanity check simulator
             try:
@@ -194,7 +196,7 @@ def runSimulation(simulator, simulator_args, outfile, save_times, save_var_names
             profiling_data_sim_runner["start"]["t_full_step"] += time.time()
 
             #Simulate
-            if (t_step > 0.0):
+            if t_step > 0.0:
                 sim.simulate(t_step, dt)
 
             profiling_data_sim_runner["end"]["t_full_step"] += time.time()
@@ -211,11 +213,11 @@ def runSimulation(simulator, simulator_args, outfile, save_times, save_var_names
             profiling_data_sim_runner["end"]["t_nc_write"] += time.time()
 
             #Write progress to screen
-            print_string = progress_printer.getPrintString(t_end)
+            # print_string = progress_printer.getPrintString(t_end)
-            if (print_string):
+            # if (print_string):
-                logger.debug(print_string)
+            #     logger.debug(print_string)
                 
-        logger.debug("Simulated to t={:f} in {:d} timesteps (average dt={:f})".format(t_end, sim.simSteps(), sim.simTime() / sim.simSteps()))
+        logger.debug("Simulated to t={:f} in {:d} timesteps (average dt={:f})".format(save_times[-1], sim.simSteps(), sim.simTime() / sim.simSteps()))
         
     return outdata.filename, profiling_data_sim_runner, sim.profiling_data_mpi
     #return outdata.filename
@@ -306,7 +308,7 @@ class IPEngine(object):
         import ipyparallel
         self.cluster = ipyparallel.Client()#profile='mpi')
         time.sleep(3)
-        while(len(self.cluster.ids) != n_engines):
+        while len(self.cluster.ids) != n_engines:
             time.sleep(0.5)
             self.logger.info("Waiting for cluster...")
             self.cluster = ipyparallel.Client()#profile='mpi')
@@ -433,58 +435,58 @@ class DataDumper(object):
         
 
 
-class ProgressPrinter(object):
+# class ProgressPrinter(object):
-    """
+#     """
-    Small helper class for 
+#     Small helper class for 
-    """
+#     """
-    def __init__(self, total_steps, print_every=5):
+#     def __init__(self, total_steps, print_every=5):
-        self.logger = logging.getLogger(__name__)
+#         self.logger = logging.getLogger(__name__)
-        self.start = time.time()
+#         self.start = time.time()
-        self.total_steps = total_steps
+#         self.total_steps = total_steps
-        self.print_every = print_every
+#         self.print_every = print_every
-        self.next_print_time = self.print_every
+#         self.next_print_time = self.print_every
-        self.last_step = 0
+#         self.last_step = 0
-        self.secs_per_iter = None
+#         self.secs_per_iter = None
         
-    def getPrintString(self, step):
+#     def getPrintString(self, step):
-        elapsed =  time.time() - self.start
+#         elapsed =  time.time() - self.start
-        if (elapsed > self.next_print_time):            
+#         if (elapsed > self.next_print_time):            
-            dt = elapsed - (self.next_print_time - self.print_every)
+#             dt = elapsed - (self.next_print_time - self.print_every)
-            dsteps = step - self.last_step
+#             dsteps = step - self.last_step
-            steps_remaining = self.total_steps - step
+#             steps_remaining = self.total_steps - step
                         
-            if (dsteps == 0):
+#             if (dsteps == 0):
-                return
+#                 return
                 
-            self.last_step = step
+#             self.last_step = step
-            self.next_print_time = elapsed + self.print_every
+#             self.next_print_time = elapsed + self.print_every
             
-            if not self.secs_per_iter:
+#             if not self.secs_per_iter:
-                self.secs_per_iter = dt / dsteps
+#                 self.secs_per_iter = dt / dsteps
-            self.secs_per_iter = 0.2*self.secs_per_iter + 0.8*(dt / dsteps)
+#             self.secs_per_iter = 0.2*self.secs_per_iter + 0.8*(dt / dsteps)
             
-            remaining_time = steps_remaining * self.secs_per_iter
+#             remaining_time = steps_remaining * self.secs_per_iter
 
-            return "{:s}. Total: {:s}, elapsed: {:s}, remaining: {:s}".format(
+#             return "{:s}. Total: {:s}, elapsed: {:s}, remaining: {:s}".format(
-                ProgressPrinter.progressBar(step, self.total_steps), 
+#                 ProgressPrinter.progressBar(step, self.total_steps), 
-                ProgressPrinter.timeString(elapsed + remaining_time), 
+#                 ProgressPrinter.timeString(elapsed + remaining_time), 
-                ProgressPrinter.timeString(elapsed), 
+#                 ProgressPrinter.timeString(elapsed), 
-                ProgressPrinter.timeString(remaining_time))
+#                 ProgressPrinter.timeString(remaining_time))
 
-    def timeString(seconds):
+#     def timeString(seconds):
-        seconds = int(max(seconds, 1))
+#         seconds = int(max(seconds, 1))
-        minutes, seconds = divmod(seconds, 60)
+#         minutes, seconds = divmod(seconds, 60)
-        hours, minutes = divmod(minutes, 60)
+#         hours, minutes = divmod(minutes, 60)
-        periods = [('h', hours), ('m', minutes), ('s', seconds)]
+#         periods = [('h', hours), ('m', minutes), ('s', seconds)]
-        time_string = ' '.join('{}{}'.format(value, name)
+#         time_string = ' '.join('{}{}'.format(value, name)
-                                for name, value in periods
+#                                 for name, value in periods
-                                if value)
+#                                 if value)
-        return time_string
+#         return time_string
 
-    def progressBar(step, total_steps, width=30):
+#     def progressBar(step, total_steps, width=30):
-        progress = np.round(width * step / total_steps).astype(np.int32)
+#         progress = np.round(width * step / total_steps).astype(np.int32)
-        progressbar = "0% [" + "#"*(progress) + "="*(width-progress) + "] 100%"
+#         progressbar = "0% [" + "#"*(progress) + "="*(width-progress) + "] 100%"
-        return progressbar
+#         return progressbar
 
 
 """

GPUSimulators/Simulator.py

@@ -25,6 +25,7 @@ import numpy as np
 import math
 import logging
 from enum import IntEnum
+from tqdm import tqdm
 
 #import pycuda.compiler as cuda_compiler
 #import pycuda.gpuarray
@@ -156,7 +157,7 @@ class BaseSimulator(object):
         self.num_substeps = num_substeps
         
         #Handle autotuning block size
-        if (self.context.autotuner):
+        if self.context.autotuner:
             peak_configuration = self.context.autotuner.get_peak_performance(self.__class__)
             block_width = int(peak_configuration["block_width"])
             block_height = int(peak_configuration["block_height"])
@@ -195,42 +196,45 @@ class BaseSimulator(object):
         Requires that the step() function is implemented in the subclasses
         """
 
-        printer = Common.ProgressPrinter(t)
+        # printer = Common.ProgressPrinter(t)
         
         t_start = self.simTime()
         t_end = t_start + t
         
         update_dt = True
-        if (dt is not None):
+        if dt is not None:
             update_dt = False
             self.dt = dt
         
-        while(self.simTime() < t_end):
+        for _ in tqdm(range(math.ceil((t_end - t_start) / self.dt)), desc="Simulation"):
             # Update dt every 100 timesteps and cross your fingers it works
             # for the next 100
-            if (update_dt and (self.simSteps() % 100 == 0)):
+            # TODO this is probably broken now after fixing the "infinite" loop
+            if update_dt and (self.simSteps() % 100 == 0):
                 self.dt = self.computeDt()*self.cfl_scale
-        
+
             # Compute timestep for "this" iteration (i.e., shorten last timestep)
             current_dt = np.float32(min(self.dt, t_end-self.simTime()))
 
             # Stop if end reached (should not happen)
-            if (current_dt <= 0.0):
+            if current_dt <= 0.0:
                 self.logger.warning("Timestep size {:d} is less than or equal to zero!".format(self.simSteps()))
                 break
-        
+
             # Step forward in time
             self.step(current_dt)
 
             #Print info
-            print_string = printer.getPrintString(self.simTime() - t_start)
+            # print_string = printer.getPrintString(self.simTime() - t_start)
-            if (print_string):
+            # if (print_string):
-                self.logger.info("%s: %s", self, print_string)
+            #     self.logger.info("%s: %s", self, print_string)
-                try:
+            #     try:
-                    self.check()
+            #         self.check()
-                except AssertionError as e:
+            #     except AssertionError as e:
-                    e.args += ("Step={:d}, time={:f}".format(self.simSteps(), self.simTime()),)
+            #         e.args += ("Step={:d}, time={:f}".format(self.simSteps(), self.simTime()),)
-                    raise
+            #         raise
+
+        print("Done")
 
 
     def step(self, dt):

Jobs/job_apptainer_lumi.slurm

@@ -0,0 +1,26 @@
+#!/bin/bash -l
+#SBATCH --job-name=lumi
+#SBATCH --account=project_4650000xx
+#SBATCH --time=00:10:00
+#SBATCH --partition=dev-g
+#SBATCH --nodes=1
+#SBATCH --ntasks-per-node=8
+#SBATCH --gpus-per-node=8
+#SBATCH --output=%x-%j.out
+#SBATCH --exclusive
+
+N=$SLURM_JOB_NUM_NODES
+echo "--nbr of nodes:", $N
+echo "--total nbr of gpus:", $SLURM_NTASKS
+
+MyDir=/project/project_4650000xx
+MyApplication=${MyDir}/FiniteVolumeGPU_HIP/mpiTesting.py
+Container=${MyDir}/FiniteVolumeGPU_HIP/my_container.sif
+
+CPU_BIND="map_cpu:49,57,17,25,1,9,33,41"
+
+export MPICH_GPU_SUPPORT_ENABLED=1
+
+srun --cpu-bind=${CPU_BIND} --mpi=pmi2 \
+     apptainer exec "${Container}" \
+     python ${MyApplication} -nx 1024 -ny 1024 --profile

Jobs/job_lumi.slrum

@@ -1,39 +0,0 @@
-#!/bin/bash -e
-#SBATCH --job-name=lumi
-#SBATCH --account=project_4650000xx
-#SBATCH --time=00:10:00
-#SBATCH --partition=dev-g
-#SBATCH --nodes=1
-#SBATCH --ntasks-per-node=8
-#SBATCH --gpus=8
-#SBATCH --gpus-per-node=8
-#SBATCH -o %x-%j.out
-#SBATCH --exclusive
-#
-
-N=$SLURM_JOB_NUM_NODES
-echo "--nbr of nodes:", $N
-echo "--total nbr of gpus:", $SLURM_NTASKS
-
-Mydir=/project/project_4650000xx
-Myapplication=${Mydir}/FiniteVolumeGPU_hip/mpiTesting.py
-
-#modules
-ml LUMI/24.03 partition/G
-ml lumi-container-wrapper
-ml cray-python/3.11.7
-ml rocm/5.4.6
-
-ml craype-accel-amd-gfx90a
-ml cray-mpich/8.1.29
-
-export PATH="/project/project_4650000xx/FiniteVolumeGPU_hip/MyCondaEnv/bin:$PATH"
-
-#missing library
-export LD_LIBRARY_PATH=/opt/cray/pe/mpich/8.1.29/ofi/cray/17.0/lib-abi-mpich:$LD_LIBRARY_PATH
-
-#Binding mask
-bind_mask="0x${fe}000000000000,0x${fe}00000000000000,0x${fe}0000,0x${fe}000000,0x${fe},0x${fe}00,0x${fe}00000000,0x${fe}0000000000"
-
-srun --cpu-bind=mask_cpu:$bind_mask \
-     python ${Myapplication} -nx 1024 -ny 1024 --profile 	

Jobs/job_lumi.slurm

@@ -0,0 +1,27 @@
+#!/bin/bash -l
+#SBATCH --job-name=lumi
+#SBATCH --account=project_4650000xx
+#SBATCH --time=00:10:00
+#SBATCH --partition=dev-g
+#SBATCH --nodes=1
+#SBATCH --ntasks-per-node=8
+#SBATCH --gpus-per-node=8
+#SBATCH --output=%x-%j.out
+#SBATCH --exclusive
+
+N=$SLURM_JOB_NUM_NODES
+echo "--nbr of nodes:", $N
+echo "--total nbr of gpus:", $SLURM_NTASKS
+
+MyDir=/project/project_4650000xx
+MyApplication=${MyDir}/FiniteVolumeGPU_HIP/mpiTesting.py
+CondaEnv=${MyDir}/FiniteVolumeGPU_HIP/MyCondaEnv/bin
+
+export PATH="${CondaEnv}:$PATH"
+
+CPU_BIND="map_cpu:49,57,17,25,1,9,33,41"
+
+export MPICH_GPU_SUPPORT_ENABLED=1
+
+srun --cpu-bind=${CPU_BIND} --mpi=pmi2 \
+     python ${MyApplication} -nx 1024 -ny 1024 --profile

README.md

@@ -5,48 +5,53 @@ This is a HIP version of the [FiniteVolume code](https://github.com/babrodtk/Fin
 ## Setup on LUMI-G
 Here is a step-by-step guide on installing packages on LUMI-G
 
-### Step 1: Install rocm-5.4.6 with Easybuild
+### Step 1: run conda-container
-```
-export EBU_USER_PREFIX=/project/project_xxxxxx/EasyBuild
-ml LUMI/24.03 partition/G
-ml EasyBuild-user
-export PYTHONIOENCODING=utf-8
-eb rocm-5.4.6.eb -r
-```
-
-### Step 2: run conda-container
 Installation via conda can be done as:
-```
+```shell
 ml LUMI/24.03 partition/G
-ml lumi-container-wrapper/0.3.3-cray-python-3.11.7
+ml lumi-container-wrapper
-```
 ```
+```shell
 conda-containerize new --prefix MyCondaEnv conda_environment_lumi.yml
 ```
 where the file `conda_environment_lumi.yml` contains packages to be installed.
 
-### Step 3: Set the env. variable to search for binaries
+### Step 1 alternative: Convert to a singularity container with cotainr
-```
+Load the required modules first
-export the bin path: export PATH="$PWD/MyCondaEnv/bin:$PATH"
+```shell
-```
+ml CrayEnv
-### An alternative: Convert to a singularity container with cotainr
+ml cotainr
 ```
+
+Then build the Singularity/Apptainer container 
+```shell
 cotainr build my_container.sif --system=lumi-g --conda-env=conda_environment_lumi.yml
 ```
 
-### Error when running MPI.
+### Step 2: Modify Slurm Job file
+Depending on your build method, update [`Jobs/job_lumi.slurm`](Jobs/job_lumi.slurm) if `conda-containerize` was used, or [`Jobs/job_apptainer_lumi.slurm`](Jobs/job_apptainer_lumi.slurm) if `containr` was used.
+
+In the job file, the required changes is to match your project allocation,
+and the directories of where the simulator and container is stored.
+
+### Step 3: Run the Slurm Job
+If `conda-containerize` was used for building:
+```shell
+sbatch Jobs/job_lumi.slurm
+```
+
+Otherwise, if `containr` was used for building:
+```shell
+sbatch Jobs/job_apptainer_lumi.slurm
+```
+
+### Troubleshooting
+
+#### Error when running MPI.
 ```
 `MPI startup(): PMI server not found. Please set I_MPI_PMI_LIBRARY variable if it is not a singleton case.
 ```
 This can be resolved by exporting this:
 ```
-export I_MPI_PMI_LIBRARY=/opt/cray/pe/mpich/8.1.27/ofi/cray/14.0/lib/libmpi.so
+export I_MPI_PMI_LIBRARY=/opt/cray/pe/mpich/8.1.29/ofi/cray/17.0/lib/libmpi.so
-```
+```
-### Install hip-python
-```
-python -m pip install -i https://test.pypi.org/simple/ hip-python==5.4.3.470.16
-```
-
-The testing was done with this specific version `hip-python==5.4.3.470.16`
-
- 

conda_environment_lumi.yml

@@ -13,8 +13,9 @@ dependencies:
 - pytools
 - netcdf4
 - scipy
+- tqdm
 - pip:
-  - hip-python==5.4.3.470.16
+  - hip-python==6.2.0.499.16
   - -i https://test.pypi.org/simple/ 
 
 

mpiTesting.py

@@ -70,7 +70,7 @@ def hip_check(call_result):
 
 args = parser.parse_args()
 
-if(args.profile):
+if args.profile:
     profiling_data = {}
     # profiling: total run time
     t_total_start = time.time()
@@ -79,6 +79,8 @@ if(args.profile):
 
 # Get MPI COMM to use
 comm = MPI.COMM_WORLD
+size = comm.Get_size()
+rank = comm.Get_rank()
 
 
 ####
@@ -86,7 +88,7 @@ comm = MPI.COMM_WORLD
 ####
 log_level_console = 20
 log_level_file = 10
-log_filename = 'mpi_' + str(comm.rank) + '.log'
+log_filename = 'mpi_' + str(rank) + '.log'
 logger = logging.getLogger('GPUSimulators')
 logger.setLevel(min(log_level_console, log_level_file))
 
@@ -110,7 +112,7 @@ logger.info("File logger using level %s to %s",
 # Initialize MPI grid etc
 ####
 logger.info("Creating MPI grid")
-grid = MPISimulator.MPIGrid(MPI.COMM_WORLD)
+grid = MPISimulator.MPIGrid(comm)
 
 """
 job_id = int(os.environ["SLURM_JOB_ID"])
@@ -152,7 +154,7 @@ gamma = 1.4
 #save_times = np.linspace(0, 0.000099, 11)
 #save_times = np.linspace(0, 0.000099, 2)
 save_times = np.linspace(0, 0.0000999, 2)
-outfile = "mpi_out_" + str(MPI.COMM_WORLD.rank) + ".nc"
+outfile = "mpi_out_" + str(rank) + ".nc"
 save_var_names = ['rho', 'rho_u', 'rho_v', 'E']
 
 arguments = IC.genKelvinHelmholtz(nx, ny, gamma, grid=grid)
@@ -160,7 +162,7 @@ arguments['context'] = cuda_context
 arguments['theta'] = 1.2
 arguments['grid'] = grid
 
-if(args.profile):
+if args.profile:
     t_init_end = time.time()
     t_init = t_init_end - t_init_start
     profiling_data["t_init"] = t_init
@@ -181,14 +183,14 @@ def genSim(grid, **kwargs):
 (outfile, sim_runner_profiling_data, sim_profiling_data) = Common.runSimulation(
     genSim, arguments, outfile, save_times, save_var_names, dt)
 
-if(args.profile):
+if args.profile:
     t_total_end = time.time()
     t_total = t_total_end - t_total_start
     profiling_data["t_total"] = t_total
-    print("Total run time on rank " + str(MPI.COMM_WORLD.rank) + " is " + str(t_total) + " s")
+    print("Total run time on rank " + str(rank) + " is " + str(t_total) + " s")
 
 # write profiling to json file
-if(args.profile and MPI.COMM_WORLD.rank == 0):
+if args.profile and rank == 0:
     job_id = ""
     if "SLURM_JOB_ID" in os.environ:
         job_id = int(os.environ["SLURM_JOB_ID"])
@@ -199,7 +201,7 @@ if(args.profile and MPI.COMM_WORLD.rank == 0):
             str(job_id) + "_" + str(allocated_nodes) + "_nodes_and_" + str(allocated_gpus) + "_GPUs_profiling.json"
         profiling_data["outfile"] = outfile
     else:
-        profiling_file = "MPI_" + str(MPI.COMM_WORLD.size) + "_procs_and_" + str(num_cuda_devices) + "_GPUs_profiling.json"
+        profiling_file = "MPI_" + str(size) + "_procs_and_" + str(num_cuda_devices) + "_GPUs_profiling.json"
 
     for stage in sim_runner_profiling_data["start"].keys():
         profiling_data[stage] = sim_runner_profiling_data["end"][stage] - sim_runner_profiling_data["start"][stage]
@@ -214,7 +216,7 @@ if(args.profile and MPI.COMM_WORLD.rank == 0):
 
     profiling_data["slurm_job_id"] = job_id
     profiling_data["n_cuda_devices"] = str(num_cuda_devices)
-    profiling_data["n_processes"] = str(MPI.COMM_WORLD.size)
+    profiling_data["n_processes"] = str(size)
     profiling_data["git_hash"] = Common.getGitHash()
     profiling_data["git_status"] = Common.getGitStatus()