alg benchmarking runner and processor

launch_alg_bench.py

@@ -0,0 +1,203 @@
+import os
+import subprocess
+from datetime import datetime
+
+################ HELPER FUNCTIONS ################
+def load_template(template_path: str):
+    output_template = ""
+    with open(template_path, "r") as handle:
+        output_template = handle.read()
+    return output_template
+
+
+def write_batch(batch_fpath: str, batch_content: str):
+    with open(batch_fpath, "w") as handle:
+        _ = handle.write(batch_content)
+
+################### SETUP DIRS ###################
+output_dir = os.getcwd()+"/output/"
+err_dir = os.getcwd()+"/error/"
+batch_files_dir = os.getcwd()+"/batchs/"
+data_dir = os.getcwd()+"/data/"
+
+if os.path.isdir(output_dir) == False:
+    os.mkdir(output_dir)
+if os.path.isdir(err_dir) == False:
+    os.mkdir(err_dir)
+if os.path.isdir(data_dir) == False:
+    os.mkdir(data_dir)
+if os.path.isdir(batch_files_dir) == False:
+    os.mkdir(batch_files_dir)
+
+################ GLOBAL DEFAULTS #################
+mpi1_bin = "/home/hpc/ihpc/ihpc136h/workspace/mpi-benchmark-tool/bin/IMB-MPI1"
+default_parameter = {
+    "time_stamp": datetime.now().strftime("%y_%m_%d_%H-%M-%S"),
+    "job_name": "",
+    "output_dir": os.getcwd()+"/output/",
+    "err_dir": os.getcwd()+"/error/",
+    "data_dir": os.getcwd()+"/data/",
+    "n_procs": 18,
+    "off_cache_flag": "",
+    "bin": mpi1_bin,
+    "n_nodes": 1
+}
+
+algs_dic = [{'name': "Allgather",
+             'flag': "I_MPI_ADJUST_ALLGATHER",
+             'algs': [
+                 "Recursive doubling ",
+                 "Bruck`s ",
+                 "Ring ",
+                 "Topology aware Gatherv + Bcast ",
+                 "Knomial ",
+             ]},
+            {'name': "Allreduce",
+             'flag': "I_MPI_ADJUST_ALLREDUCE",
+             'algs': [
+                 "Recursive doubling ",
+                 "Rabenseifner`s ",
+                 "Reduce + Bcast ",
+                 "Topology aware Reduce + Bcast ",
+                 "Binomial gather + scatter ",
+                 "Topology aware binominal gather + scatter ",
+                 "Shumilin`s ring ",
+                 "Ring ",
+                 "Knomial ",
+                 "Topology aware SHM-based flat ",
+                 "Topology aware SHM-based Knomial ",
+                 "Topology aware SHM-based Knary ",
+             ]},
+
+            {'name': "Alltoall",
+             'flag': "I_MPI_ADJUST_ALLTOALL",
+             'algs': [
+                 "Bruck`s ",
+                 "Isend/Irecv + waitall ",
+                 "Pair wise exchange ",
+                 "Plum`s ",
+             ]},
+            {'name': "Barrier",
+             'flag': "I_MPI_ADJUST_BARRIER",
+             'algs': [
+                 "Dissemination ",
+                 "Recursive doubling ",
+                 "Topology aware dissemination ",
+                 "Topology aware recursive doubling ",
+                 "Binominal gather + scatter ",
+                 "Topology aware binominal gather + scatter ",
+                 "Topology aware SHM-based flat ",
+                 "Topology aware SHM-based Knomial ",
+                 "Topology aware SHM-based Knary ",
+             ]},
+            {'name': "Bcast",
+             'flag': "I_MPI_ADJUST_BCAST",
+             'algs': [
+                 "Binomial ",
+                 "Recursive doubling ",
+                 "Ring ",
+                 "Topology aware binomial ",
+                 "Topology aware recursive doubling ",
+                 "Topology aware ring ",
+                 "Shumilin`s ",
+                 "Knomial ",
+                 "Topology aware SHM-based flat ",
+                 "Topology aware SHM-based Knomial ",
+                 "Topology aware SHM-based Knary ",
+                 "NUMA aware SHM-based (SSE4.2) ",
+                 "NUMA aware SHM-based (AVX2) ",
+                 "NUMA aware SHM-based (AVX512) ",
+             ]},
+
+            {'name': "Gather",
+             'flag': "I_MPI_ADJUST_GATHER",
+             'algs': [
+                 "Binomial ",
+                 "Topology aware binomial ",
+                 "Shumilin`s ",
+                 "Binomial with segmentation ",
+             ]},
+
+            {'name': "Reduce_scatter",
+             'flag': "I_MPI_ADJUST_REDUCE_SCATTER",
+             'algs': [
+                 "Recursive halving ",
+                 "Pair wise exchange ",
+                 "Recursive doubling ",
+                 "Reduce + Scatterv ",
+                 "Topology aware Reduce + Scatterv ",
+             ]},
+
+            {'name': "Reduce",
+             'flag': "I_MPI_ADJUST_REDUCE",
+             'algs': [
+                 "Shumilin`s ",
+                 "Binomial ",
+                 "Topology aware Shumilin`s ",
+                 "Topology aware binomial ",
+                 "Rabenseifner`s ",
+                 "Topology aware Rabenseifner`s ",
+                 "Knomial ",
+                 "Topology aware SHM-based flat ",
+                 "Topology aware SHM-based Knomial ",
+                 "Topology aware SHM-based Knary ",
+                 "Topology aware SHM-based binomial ",
+             ]},
+
+            {'name': "Scatter",
+             'flag': "I_MPI_ADJUST_SCATTER",
+             'algs': [
+                 "Binomial ",
+                 "Topology aware binomial ",
+                 "Shumilin`s ",
+             ]},
+            ]
+log = ""
+
+############## MULTIPLE-NODE LAUNCH ##############
+off_cache_flags = [
+    "-off_cache -1",
+    "-off_cache 50",
+    ""
+]
+
+ndcnt = [
+    2,
+    3,
+    4,
+    5,
+    6,
+    7,
+    8,
+    9,
+    10
+]
+
+proc_per_node = 72
+multiple_node_parameter = dict(default_parameter)
+multiple_node_template = load_template("./templates/multinode_algs.template")
+
+for flag in off_cache_flags:
+    multiple_node_parameter["off_cache_flag"] = flag
+    for n_nodes in ndcnt:
+        n_procs = n_nodes*proc_per_node
+        multiple_node_parameter["n_procs"] = int(n_procs)
+        multiple_node_parameter["n_nodes"] = n_nodes
+        for alg_conf in algs_dic:
+            collective = alg_conf['name']
+            multiple_node_parameter["job_name"] = collective
+            multiple_node_parameter["alg_flag"] = alg_conf['flag']
+            algs = alg_conf["algs"]
+            for idx, alg in enumerate(algs):
+                multiple_node_parameter["alg_name"] = alg
+                multiple_node_parameter["alg_idx"] = idx
+                batch_file = os.path.join(batch_files_dir,
+                                          f"{collective}_{alg.strip().replace('`','').replace(' ','_').replace('/','_')}.sh")
+                write_batch(batch_file,
+                            multiple_node_template.format(**multiple_node_parameter))
+                result = subprocess.run(["sbatch", batch_file],
+                                        capture_output=True, text=True)
+                log += f"#{collective} {n_procs}" + "\n"
+                log += "\tSTDOUT:" + result.stdout + "\n"
+                log += "\tSTDERR:" + result.stderr + "\n"
+print(log)

postprocess_data_algs.py

@@ -0,0 +1,112 @@
+from venv import create
+import pandas as pd
+import os
+
+data_markers = {
+    "block_separator": "#----------------------------------------------------------------",
+    "benchmark_type": "# Benchmarking",
+    "processes_num": "# #processes = ",
+    "min_bytelen": "# Minimum message length in bytes",
+    "max_bytelen": "# Maximum message length in bytes",
+    "mpi_datatype": "# MPI_Datatype                   :",
+    "mpi_red_datatype": "# MPI_Datatype for reductions    :",
+    "mpi_red_op": "# MPI_Op",
+    "end_of_table": "# All processes entering MPI_Finalize",
+    "creation_time": "# CREATION_TIME :",
+    "n_nodes": "# N_NODES       :",
+    "off_cache_flag": "# OFF_CACHE_FLAG  :",
+    "algorithm":"# ALGORITHM  :"
+}
+
+column_names = [
+    "benchmark_type",
+    "proc_num",
+    "msg_size_bytes",
+    "repetitions",
+    "t_min_usec",
+    "t_max_usec",
+    "t_avg_usec",
+    "mpi_datatype",
+    "mpi_red_datatype",
+    "mpi_red_op",
+    "creation_time",
+    "n_nodes",
+    "off_cache_flag",
+    "algorithm"
+]
+
+data = list()
+
+for file in os.listdir("data/"):
+    with open("data/"+file, 'r') as f:
+        lines = f.readlines()
+
+        past_preheader = False
+        in_header = False
+        in_body = False
+
+        btype = "NA"
+        proc_num = "NA"
+        mpi_datatype = "NA"
+        mpi_red_datatype = "NA"
+        mpi_red_op = "NA"
+        creation_time = "NA"
+        n_nodes = "NA"
+        off_cache_flag = "NA"
+        algorithm = "NA"
+
+        for line in lines:
+            if data_markers["block_separator"] in line:
+                if in_header and not past_preheader:
+                    past_preheader = True
+                elif in_header and past_preheader:
+                    in_body = True
+                in_header = not in_header
+                continue
+            if not in_header and not in_body and past_preheader:
+                if data_markers["mpi_datatype"] in line:
+                    mpi_datatype = line.split()[-1]
+                elif data_markers["mpi_red_datatype"] in line:
+                    mpi_red_datatype = line.split()[-1]
+                elif data_markers["mpi_red_op"] in line:
+                    mpi_red_op = line.split()[-1]
+
+            if not in_header and not in_body and not past_preheader:
+                if data_markers["n_nodes"] in line:
+                    n_nodes = line.split()[-1]
+                if data_markers["creation_time"] in line:
+                    creation_time = line.split()[-1]
+                if data_markers["off_cache_flag"] in line:
+                    off_cache_flag = line.split(":")[-1].strip()
+                    if off_cache_flag == "": off_cache_flag = "NA"
+                    else: off_cache_flag = off_cache_flag.replace("-off_cache","")
+                if data_markers["algorithm"] in line:
+                    algorithm = line.split(":")[-1].strip()
+
+            if past_preheader and in_header:
+                if data_markers["benchmark_type"] in line:
+                    btype = line.split()[2]
+                if data_markers["processes_num"] in line:
+                    proc_num = int(line.split()[3])
+
+            if in_body:
+                if "#" in line or "".join(line.split()) == "":
+                    continue
+                if data_markers["end_of_table"] in line:
+                    break
+                if("int-overflow" in line) : continue 
+                if("out-of-mem" in line) : continue 
+                data.append([btype, proc_num]+[int(s) if s.isdigit()
+                            else float(s) for s in line.split()] +
+                            [
+                                mpi_datatype,
+                                mpi_red_datatype,
+                                mpi_red_op,
+                                creation_time,
+                                n_nodes,
+                                off_cache_flag,
+                                algorithm
+                ])
+
+df = pd.DataFrame(data, columns=column_names)
+df.to_csv("data.csv", index=False)

templates/multinode.template

@@ -3,13 +3,36 @@
 #SBATCH --output={output_dir}{job_name}_{n_procs}.out
 #SBATCH --error={err_dir}{job_name}_{n_procs}.err
 #SBATCH --nodes={n_nodes}
-#SBATCH --time=00:10:00
+#SBATCH --nodelist=f01[01-64]
+#SBATCH --time=00:30:00
 #SBATCH --export=NONE
 
+# Switch Help Table
+# SwitchName=fswibl01 Level=0 LinkSpeed=1 Nodes=f01[01-64]
+# SwitchName=fswibl02 Level=0 LinkSpeed=1 Nodes=f02[01-64]
+# SwitchName=fswibl03 Level=0 LinkSpeed=1 Nodes=f03[01-64]
+# SwitchName=fswibl04 Level=0 LinkSpeed=1 Nodes=f04[01-64]
+# SwitchName=fswibl05 Level=0 LinkSpeed=1 Nodes=f05[01-64]
+# SwitchName=fswibl06 Level=0 LinkSpeed=1 Nodes=f06[01-64]
+# SwitchName=fswibl07 Level=0 LinkSpeed=1 Nodes=f01[65-88],f02[65-88]
+# SwitchName=fswibl08 Level=0 LinkSpeed=1 Nodes=f03[65-88],f04[65-88],fritz[1-2]
+# SwitchName=fswibl09 Level=0 LinkSpeed=1 Nodes=f05[65-88],f06[65-88],fritz[3-4],fviz1
+# SwitchName=fswibl10 Level=0 LinkSpeed=1 Nodes=f07[01-64]
+# SwitchName=fswibl11 Level=0 LinkSpeed=1 Nodes=f08[01-64]
+# SwitchName=fswibl12 Level=0 LinkSpeed=1 Nodes=f09[01-64]
+# SwitchName=fswibl13 Level=0 LinkSpeed=1 Nodes=f10[01-64]
+
 unset SLURM_EXPORT_ENV
 
 module load intel intelmpi 
 
+# Enable tuned collectives
+export I_MPI_TUNING=on
+export I_MPI_TUNING_MODE=auto        # or 'collectives'
+
+# Options: 0=auto, 1=recursive doubling, 2=ring, 3=binomial tree, 4=scatter-allgather
+export I_MPI_COLL_ALLGATHER=2 
+export I_MPI_COLL_GATHER=2     
 
 OUTPUT_FILENAME="{data_dir}/{job_name}_$SLURM_JOB_ID.dat"
 
@@ -17,6 +40,5 @@ echo  "# CREATION_TIME : {time_stamp}" > $OUTPUT_FILENAME
 echo  "# N_NODES       : {n_nodes}" >> $OUTPUT_FILENAME
 echo  "# OFF_CACHE_FLAG  : {off_cache_flag}">> $OUTPUT_FILENAME
 
-srun --cpu-freq=2000000-2000000:performance -N {n_nodes} -n{n_procs} {bin} {job_name} -npmin {n_procs} {off_cache_flag}  >> $OUTPUT_FILENAME 
-
+srun --cpu-freq=2000000-2000000:performance -N {n_nodes} -n{n_procs} {bin} {job_name} -npmin {n_procs} {off_cache_flag} -mem 2 -time 60 >> $OUTPUT_FILENAME 
 

templates/multinode_algs.template

@@ -0,0 +1,38 @@
+#!/bin/bash -l
+#SBATCH --job-name={job_name}_{n_procs}_{alg_idx}
+#SBATCH --output={output_dir}{job_name}_{n_procs}.out
+#SBATCH --error={err_dir}{job_name}_{n_procs}.err
+#SBATCH --nodes={n_nodes}
+#SBATCH --nodelist=f01[01-64]
+#SBATCH --time=00:30:00
+#SBATCH --export=NONE
+
+# SwitchName=fswibl01 Level=0 LinkSpeed=1 Nodes=f01[01-64]
+# SwitchName=fswibl02 Level=0 LinkSpeed=1 Nodes=f02[01-64]
+# SwitchName=fswibl03 Level=0 LinkSpeed=1 Nodes=f03[01-64]
+# SwitchName=fswibl04 Level=0 LinkSpeed=1 Nodes=f04[01-64]
+# SwitchName=fswibl05 Level=0 LinkSpeed=1 Nodes=f05[01-64]
+# SwitchName=fswibl06 Level=0 LinkSpeed=1 Nodes=f06[01-64]
+# SwitchName=fswibl07 Level=0 LinkSpeed=1 Nodes=f01[65-88],f02[65-88]
+# SwitchName=fswibl08 Level=0 LinkSpeed=1 Nodes=f03[65-88],f04[65-88],fritz[1-2]
+# SwitchName=fswibl09 Level=0 LinkSpeed=1 Nodes=f05[65-88],f06[65-88],fritz[3-4],fviz1
+# SwitchName=fswibl10 Level=0 LinkSpeed=1 Nodes=f07[01-64]
+# SwitchName=fswibl11 Level=0 LinkSpeed=1 Nodes=f08[01-64]
+# SwitchName=fswibl12 Level=0 LinkSpeed=1 Nodes=f09[01-64]
+# SwitchName=fswibl13 Level=0 LinkSpeed=1 Nodes=f10[01-64]
+
+unset SLURM_EXPORT_ENV
+
+module load intel intelmpi 
+
+export {alg_flag}={alg_idx}
+
+OUTPUT_FILENAME="{data_dir}/{job_name}_$SLURM_JOB_ID.dat"
+
+echo  "# CREATION_TIME : {time_stamp}" > $OUTPUT_FILENAME
+echo  "# N_NODES       : {n_nodes}" >> $OUTPUT_FILENAME
+echo  "# OFF_CACHE_FLAG  : {off_cache_flag}">> $OUTPUT_FILENAME
+echo  "# ALGORITHM  : {alg_name}">> $OUTPUT_FILENAME
+
+srun --cpu-freq=2000000-2000000:performance -N {n_nodes} -n{n_procs} {bin} {job_name} -npmin {n_procs} {off_cache_flag} -mem 2 -time 60 >> $OUTPUT_FILENAME 
+