def main():
from scripts.utils import create_project_tarball, get_stencil_num
from scripts.conf.conf import machine_conf, machine_info
import os, sys
from csv import DictReader
import time,datetime
dry_run = 1 if len(sys.argv)<2 else int(sys.argv[1])
time_stamp = datetime.datetime.fromtimestamp(time.time()).strftime('%Y%m%d_%H_%M')
exp_name = "pluto_thread_scaling_at_%s_%s" % (machine_info['hostname'], time_stamp)
tarball_dir='results/'+exp_name
if(dry_run==0): create_project_tarball(tarball_dir, "test_"+exp_name)
target_dir='results/' + exp_name
# parse the results to find out which of the already exist
data = []
data_file = os.path.join('results', 'summary.csv')
try:
with open(data_file, 'rb') as output_file:
raw_data = DictReader(output_file)
for k in raw_data:
kernel = get_stencil_num(k)
if(kernel==0):
k['stencil'] ='3d25pt'
elif(kernel==1):
k['stencil'] ='3d7pt'
elif(kernel==4):
k['stencil'] ='3d25pt_var'
elif(kernel==5):
k['stencil'] ='3d7pt_var'
else:
raise
data.append(k)
except:
pass
param_l = dict()
for k in data:
try:
param_l[(k['stencil'], int(k['Global NX']), k['LIKWID performance counter'] ) ] = ([int(k['PLUTO tile size of loop 1']), int(k['PLUTO tile size of loop 3']), int(k['PLUTO tile size of loop 4'])], int(k['Number of time steps']) )
except:
print k
raise
count = 0
for group in ['MEM', 'L2']:
# for group in ['MEM', 'L2', 'L3', 'DATA', 'TLB_DATA', 'ENERGY']:
if(machine_info['hostname']=='Haswell_18core'):
machine_conf['pinning_args'] = " -m -g " + group + " -C S1:0-"
elif(machine_info['hostname']=='IVB_10core'):
if group=='TLB_DATA': group='TLB'
machine_conf['pinning_args'] = " -g " + group + " -C S0:0-"
# for k,v in param_l.iteritems(): print k,v
count = count + thread_scaling_test(dry_run, target_dir, exp_name, param_l=param_l, group=group)
print "experiments count =" + str(count)
def main():
from scripts.utils import create_project_tarball, get_stencil_num
from scripts.conf.conf import machine_conf, machine_info
import os, sys
from csv import DictReader
import time, datetime
dry_run = 1 if len(sys.argv) < 2 else int(sys.argv[1])
time_stamp = datetime.datetime.fromtimestamp(
time.time()).strftime('%Y%m%d_%H_%M')
exp_name = "pochoir_increasing_grid_size_at_%s_%s" % (
machine_info['hostname'], time_stamp)
tarball_dir = 'results/' + exp_name
if (dry_run == 0):
create_project_tarball(tarball_dir, "project_" + exp_name)
target_dir = 'results/' + exp_name
# parse the results to find out which of the already exist
data = []
data_file = os.path.join('results', 'summary.csv')
try:
with open(data_file, 'rb') as output_file:
raw_data = DictReader(output_file)
for k in raw_data:
k['stencil'] = get_stencil_num(k)
data.append(k)
except:
pass
params = set()
for k in data:
try:
params.add((k['stencil'], int(k['Global NX'])))
except:
print k
raise
#update the pinning information to use all cores
th = machine_info['n_cores']
count = 0
for group in ['MEM', 'TLB_DATA', 'L2', 'L3', 'DATA']: #, 'ENERGY']:
if (machine_info['hostname'] == 'IVB_10core'):
if group == 'TLB_DATA': group = 'TLB'
machine_conf[
'pinning_args'] = " -m -g " + group + " -c " + "%d-%d " % (
0, th - 1) + '-- numactl --physcpubind=%d-%d' % (0, th - 1)
# for k in params: print k
count = count + igs_test(
dry_run, target_dir, exp_name, th=th, params=params, group=group)
print "experiments count =" + str(count)
def main():
from scripts.utils import create_project_tarball, get_stencil_num
from scripts.conf.conf import machine_conf, machine_info
import os, sys
from csv import DictReader
import time,datetime
dry_run = 1 if len(sys.argv)<2 else int(sys.argv[1])
time_stamp = datetime.datetime.fromtimestamp(time.time()).strftime('%Y%m%d_%H_%M')
exp_name = "pochoir_increasing_grid_size_at_%s_%s" % (machine_info['hostname'], time_stamp)
tarball_dir='results/'+exp_name
if(dry_run==0): create_project_tarball(tarball_dir, "project_"+exp_name)
target_dir='results/' + exp_name
# parse the results to find out which of the already exist
data = []
data_file = os.path.join('results', 'summary.csv')
try:
with open(data_file, 'rb') as output_file:
raw_data = DictReader(output_file)
for k in raw_data:
k['stencil'] = get_stencil_num(k)
data.append(k)
except:
pass
params = set()
for k in data:
try:
params.add( (k['stencil'], int(k['Global NX'])) )
except:
print k
raise
#update the pinning information to use all cores
th = machine_info['n_cores']
count = 0
for group in ['MEM', 'TLB_DATA', 'L2', 'L3', 'DATA']:#, 'ENERGY']:
if(machine_info['hostname']=='IVB_10core'):
if group=='TLB_DATA': group='TLB'
machine_conf['pinning_args'] = " -m -g " + group + " -c " + "%d-%d "%(0, th-1) + '-- numactl --physcpubind=%d-%d'%(0,th-1)
# for k in params: print k
count = count + igs_test(dry_run, target_dir, exp_name, th=th, params=params, group=group)
print "experiments count =" + str(count)
def main():
from scripts.utils import create_project_tarball, get_stencil_num
from scripts.conf.conf import machine_conf, machine_info
import os, sys
from csv import DictReader
import time, datetime
dry_run = 1 if len(sys.argv) < 2 else int(sys.argv[1])
time_stamp = datetime.datetime.fromtimestamp(time.time()).strftime("%Y%m%d_%H_%M")
exp_name = "pochoir_thread_scaling_at_%s_%s" % (machine_info["hostname"], time_stamp)
tarball_dir = "results/" + exp_name
if dry_run == 0:
create_project_tarball(tarball_dir, "project_" + exp_name)
target_dir = "results/" + exp_name
# parse the results to find out which of the already exist
data = []
data_file = os.path.join("results", "summary.csv")
try:
with open(data_file, "rb") as output_file:
raw_data = DictReader(output_file)
for k in raw_data:
k["stencil"] = get_stencil_num(k)
data.append(k)
except:
pass
params = set()
for k in data:
try:
params.add((k["stencil"], int(k["Global NX"])))
except:
print k
raise
count = 0
for group in ["MEM", "L2"]: # 'TLB_DATA', 'L3', 'DATA', 'ENERGY']:
# for k in params: print k
count = count + thread_scaling_test(dry_run, target_dir, exp_name, params=params, group=group)
print "experiments count =" + str(count)
def main():
from scripts.utils import create_project_tarball, get_stencil_num
from scripts.conf.conf import machine_conf, machine_info
import os, sys
from csv import DictReader
import time,datetime
dry_run = 1 if len(sys.argv)<2 else int(sys.argv[1])
time_stamp = datetime.datetime.fromtimestamp(time.time()).strftime('%Y%m%d_%H_%M')
exp_name = "pochoir_thread_scaling_at_%s_%s" % (machine_info['hostname'], time_stamp)
tarball_dir='results/'+exp_name
if(dry_run==0): create_project_tarball(tarball_dir, "project_"+exp_name)
target_dir='results/' + exp_name
# parse the results to find out which of the already exist
data = []
data_file = os.path.join('results', 'summary.csv')
try:
with open(data_file, 'rb') as output_file:
raw_data = DictReader(output_file)
for k in raw_data:
k['stencil'] = get_stencil_num(k)
data.append(k)
except:
pass
params = set()
for k in data:
try:
params.add( (k['stencil'], int(k['Global NX'])) )
except:
print k
raise
count = 0
for group in ['MEM', 'L2']:# 'TLB_DATA', 'L3', 'DATA', 'ENERGY']:
# for k in params: print k
count = count + thread_scaling_test(dry_run, target_dir, exp_name, params=params, group=group)
print "experiments count =" + str(count)
def parse_entry_info(k, is_tgs_only):
from scripts.utils import get_stencil_num, load_csv
# # get processor name from the file names
# if(k['OpenMP Threads']!=''):
# if(int(k['OpenMP Threads']) == 10):
# machine_name = 'ivb10'
# elif(int(k['OpenMP Threads']) == 18):
# machine_name = 'hw18'
# Use single field to represent the performance
if 'Total RANK0 MStencil/s MAX' in k.keys():
if (k['Total RANK0 MStencil/s MAX'] != ''):
k['MStencil/s MAX'] = k['MWD main-loop RANK0 MStencil/s MAX']
# temporary for deprecated format
if 'RANK0 MStencil/s MAX' in k.keys():
if k['RANK0 MStencil/s MAX'] != '':
k['MStencil/s MAX'] = k['RANK0 MStencil/s MAX']
# add stencil operator
k['stencil'] = get_stencil_num(k)
if k['stencil'] == 0:
k['stencil_name'] = '25_pt_const'
elif k['stencil'] == 1:
k['stencil_name'] = '7_pt_const'
elif k['stencil'] == 4:
k['stencil_name'] = '25_pt_var'
elif k['stencil'] == 5:
k['stencil_name'] = '7_pt_var'
elif k['stencil'] == 6:
k['stencil_name'] = 'solar'
# add the approach
if (k['Time stepper orig name'] == 'Spatial Blocking'):
k['method'] = 'Spt.blk.'
elif (k['Time stepper orig name'] in ['PLUTO', 'Pochoir']):
k['method'] = k['Time stepper orig name']
elif (k['Time stepper orig name'] == 'Diamond'):
if ('_tgs1_' in k['file_name']):
k['method'] = '1WD'
else:
k['method'] = 'MWD'
else:
print("ERROR: Unknow time stepper")
raise
# TEMPORARY: Older versions of Pochoir did not have OpenMP threads count
if (k['method'] == 'Pochoir' and k['OpenMP Threads'] == ''):
k['OpenMP Threads'] = 0
# add mwd type
k['mwdt'] = 'none'
if (k['method'] == 'MWD'):
mwd = k['Wavefront parallel strategy'].lower()
if ('fixed' in mwd) and ('relaxed' in mwd):
k['mwdt'] = 'fers'
elif ('fixed' in mwd):
k['mwdt'] = 'fe'
elif ('relaxed' in mwd):
k['mwdt'] = 'rs'
elif ('wavefront' in mwd):
k['mwdt'] = 'block'
# add precision information
p = 1 if k['Precision'] in 'DP' else 0
k['Precision'] = p
# TLB measurement for LIKWID 4
if 'L1 DTLB load miss rate avg' in k.keys():
if k['L1 DTLB load miss rate avg'] != '':
hw_ctr_fields['TLB'] = [('L1 DTLB load miss rate avg', float)]
hw_ctr_labels['TLB'] = [('L1 DTLB load miss rate avg', 'tlb_',
'tlb')]
# parse the general fileds' format
for f in req_fields + hw_ctr_fields[k['LIKWID performance counter']]:
try:
k[f[0]] = map(f[1], [k[f[0]]])[0]
except:
print("ERROR: results entry missing essential data at file:%s" %
(k['file_name']))
print f[0]
print k
return
# Parse the L2 data per thread
if (k['LIKWID performance counter'] == 'L2'):
k['L2 vol list'] = []
for i in range(k['OpenMP Threads']):
field = 'L2 data volume core %d' % i
if field in k.keys():
k['L2 vol list'].append(float(k['L2 data volume core %d' % i]))
k['tgsl'] = k['Thread group size']
if (is_tgs_only == 0): # regular mode for all MWD
if (k['method'] == 'MWD'):
k['tgsl'] = 100
def main():
import sys
from scripts.utils import get_stencil_num, load_csv
from collections import OrderedDict
raw_data = load_csv(sys.argv[1])
req_fields = [('MStencil/s MAX', float), ('Precision', int),
('Global NX', int), ('Number of time steps', int),
('Number of tests', int)]
hw_ctr_fields = {
'': [],
'TLB': [('L1 DTLB miss rate sum', float)],
'DATA': [('Load to Store ratio avg', float)],
'L2': [('L2 data volume sum', float)],
'L3': [('L3 data volume sum', float)],
'MEM': [('Total Memory Transfer', float),
('Sustained Memory BW', float)],
'ENERGY': [('Energy', float), ('Energy DRAM', float), ('Power', float),
('Power DRAM', float)]
}
duplicates = set()
meas_figs = dict()
perf_fig = dict()
for k in raw_data:
# get processor name from the file names
if (k['OpenMP Threads'] != ''):
if (int(k['OpenMP Threads']) == 10):
machine_name = 'ivb10'
elif (int(k['OpenMP Threads']) == 18):
machine_name = 'hw18'
# Use single field to represent the performance
if 'Total RANK0 MStencil/s MAX' in k.keys():
if (k['Total RANK0 MStencil/s MAX'] != ''):
k['MStencil/s MAX'] = k['MWD main-loop RANK0 MStencil/s MAX']
# temporary for deprecated format
if 'RANK0 MStencil/s MAX' in k.keys():
if k['RANK0 MStencil/s MAX'] != '':
k['MStencil/s MAX'] = k['RANK0 MStencil/s MAX']
# add stencil operator
k['stencil'] = get_stencil_num(k)
if k['stencil'] == 0:
k['stencil_name'] = '25_pt_const'
elif k['stencil'] == 1:
k['stencil_name'] = '7_pt_const'
elif k['stencil'] == 4:
k['stencil_name'] = '25_pt_var'
elif k['stencil'] == 5:
k['stencil_name'] = '7_pt_var'
elif k['stencil'] == 6:
k['stencil_name'] = 'solar'
# add the approach
if (k['Time stepper orig name'] == 'Spatial Blocking'):
k['method'] = 'Spt.blk.'
elif (k['Time stepper orig name'] in ['PLUTO', 'Pochoir']):
k['method'] = k['Time stepper orig name']
elif (k['Time stepper orig name'] == 'Diamond'):
if ('_tgs1_' in k['file_name']):
k['method'] = '1WD'
else:
k['method'] = 'MWD'
else:
print("ERROR: Unknow time stepper")
raise
# add mwd type
k['mwdt'] = 'none'
if (k['method'] == 'MWD'):
mwd = k['Wavefront parallel strategy'].lower()
if ('fixed' in mwd) and ('relaxed' in mwd):
k['mwdt'] = 'fers'
elif ('fixed' in mwd):
k['mwdt'] = 'fe'
elif ('relaxed' in mwd):
k['mwdt'] = 'rs'
elif ('wavefront' in mwd):
k['mwdt'] = 'block'
# add precision information
p = 1 if k['Precision'] in 'DP' else 0
k['Precision'] = p
# TLB measurement for LIKWID 4
if 'L1 DTLB load miss rate avg' in k.keys():
if k['L1 DTLB load miss rate avg'] != '':
hw_ctr_fields['TLB'] = [('L1 DTLB load miss rate avg', float)]
hw_ctr_labels['TLB'] = [('L1 DTLB load miss rate avg', 'tlb_',
'tlb')]
entry = {}
# parse the general fileds' format
for f in req_fields + hw_ctr_fields[k['LIKWID performance counter']]:
try:
entry[f[0]] = map(f[1], [k[f[0]]])[0]
except:
print(
"ERROR: results entry missing essential data at file:%s" %
(k['file_name']))
print f[0]
print k
return
#find repeated data
key = (entry['Precision'], k['stencil_name'],
k['LIKWID performance counter'], k['mwdt'], k['method'],
entry['Global NX'])
if key not in duplicates:
duplicates.add(key)
else:
print("Repeated result at: %s" % (k['file_name']))
continue
# Initialize plot entry if does not exist for current data entry
# for m,n in entry.iteritems(): print m,n
measure_list = [
'n', 'perf', 'total energy', 'tlb', 'mem bw', 'l2 bw', 'l3 bw',
'mem vol', 'l2 vol', 'l3 vol', 'data', 'tgs', 'thx', 'thy', 'thz',
'blk size', 'diam width', 'bs_z'
]
plot_key = (entry['Precision'], k['stencil_name'],
k['LIKWID performance counter'])
line_key = (k['mwdt'], k['method'])
if plot_key not in meas_figs.keys():
meas_figs[plot_key] = {}
if line_key not in meas_figs[plot_key].keys():
meas_figs[plot_key][line_key] = {meas: [] for meas in measure_list}
# append the measurement data
meas_figs[plot_key][line_key]['n'].append(entry['Global NX'])
# meas_figs[plot_key][line_key]['perf'].append(entry['MStencil/s MAX']/1e3)
N = entry['Global NX']**3 * entry['Number of time steps'] * entry[
'Number of tests'] / 1e9
# Memory
if k['LIKWID performance counter'] == 'MEM':
meas_figs[plot_key][line_key]['mem bw'].append(
entry['Sustained Memory BW'] / 1e3)
meas_figs[plot_key][line_key]['mem vol'].append(
entry['Total Memory Transfer'] / N)
# Energy
elif k['LIKWID performance counter'] == 'ENERGY':
entry['cpu energy pj/lup'] = entry['Energy'] / N
entry['dram energy pj/lup'] = entry['Energy DRAM'] / N
entry['total energy pj/lup'] = entry['cpu energy pj/lup'] + entry[
'dram energy pj/lup']
if (entry['total energy pj/lup'] < 3e3):
# entry['total energy pj/lup'] = 0
meas_figs[plot_key][line_key]['total energy'].append(
entry['total energy pj/lup'])
else:
del meas_figs[plot_key][line_key]['n'][-1]
# TLB
elif k['LIKWID performance counter'] == 'TLB':
meas_figs[plot_key][line_key]['tlb'].append(
entry[hw_ctr_fields['TLB'][0][0]])
# L2
elif k['LIKWID performance counter'] == 'L2':
meas_figs[plot_key][line_key]['l2 vol'].append(
entry['L2 data volume sum'] / N)
#L3
elif k['LIKWID performance counter'] == 'L3':
meas_figs[plot_key][line_key]['l3 vol'].append(
entry['L3 data volume sum'] / N)
#CPU
elif k['LIKWID performance counter'] == 'DATA':
meas_figs[plot_key][line_key]['data'].append(
entry['Load to Store ratio avg'])
#Diamond tiling data
if (k['method'] == '1WD' or k['method'] == 'MWD'):
meas_figs[plot_key][line_key]['diam width'].append(
int(k['Intra-diamond width']))
meas_figs[plot_key][line_key]['tgs'].append(
int(k['Thread group size']))
meas_figs[plot_key][line_key]['thx'].append(
int(k['Threads along x-axis']))
meas_figs[plot_key][line_key]['thy'].append(
int(k['Threads along y-axis']))
meas_figs[plot_key][line_key]['thz'].append(
int(k['Threads along z-axis']))
meas_figs[plot_key][line_key]['blk size'].append(
int(k['Total cache block size (kiB)']) / 1024.0)
meas_figs[plot_key][line_key]['bs_z'].append(
int(k['Multi-wavefront updates']))
# append the performance data
plot_key = (entry['Precision'], k['stencil_name'])
line_key = (k['mwdt'], k['method'])
if plot_key not in perf_fig.keys(): # figure
perf_fig[plot_key] = dict()
perf_line = perf_fig[plot_key]
if line_key not in perf_line.keys(): # line
perf_line[line_key] = dict()
perf_point = perf_line[line_key]
nx = entry['Global NX']
if nx not in perf_point.keys(): # points
perf_point[nx] = [entry['MStencil/s MAX'] / 1e3]
else:
perf_point[nx].append(entry['MStencil/s MAX'] / 1e3)
del raw_data
#sort performance results
for k, v in perf_fig.iteritems():
for k2, v2 in perf_fig[k].iteritems():
perf_line = perf_fig[k][k2]
perf_fig[k][k2] = OrderedDict(
sorted(perf_fig[k][k2].iteritems(), key=lambda x: x[0]))
# for k,v in perf_fig.iteritems():
# print(k, "##########")
# for k2,v2 in perf_fig[k].iteritems():
# print(k2,v2)
#sort the plot lines
for p in meas_figs:
for l in meas_figs[p]:
pl = meas_figs[p][l]
#remove unused fields
empty = []
for key, val in pl.iteritems():
if (val == []):
empty.append(key)
for key in empty:
del pl[key]
lines = []
[lines.append(pl[val]) for val in measure_list if val in pl.keys()]
lines = sorted(zip(*lines))
idx = 0
for val in measure_list:
if (val in pl.keys()):
if (pl[val]):
pl[val] = [x[idx] for x in lines]
idx = idx + 1
# for m,n in meas_figs.iteritems():
# print "##############",m
# for i,j in n.iteritems():
# print i,j
plot_all(perf_fig, meas_figs, machine_name)
def main():
import sys
from scripts.utils import get_stencil_num, load_csv
from collections import OrderedDict
raw_data = load_csv(sys.argv[1])
req_fields = [('MStencil/s MAX', float), ('Precision', int), ('Global NX', int), ('Number of time steps', int), ('Number of tests', int)]
hw_ctr_fields = {
'':[],
'TLB':[('L1 DTLB miss rate sum', float)],
'DATA':[('Load to Store ratio avg', float)],
'L2':[('L2 data volume sum', float)],
'L3':[('L3 data volume sum', float)],
'MEM':[('Total Memory Transfer', float),('Sustained Memory BW', float)],
'ENERGY':[('Energy', float), ('Energy DRAM', float), ('Power',float), ('Power DRAM', float)]}
duplicates = set()
meas_figs = dict()
perf_fig = dict()
for k in raw_data:
# get processor name from the file names
if(k['OpenMP Threads']!=''):
if(int(k['OpenMP Threads']) == 10):
machine_name = 'ivb10'
elif(int(k['OpenMP Threads']) == 18):
machine_name = 'hw18'
# Use single field to represent the performance
if 'Total RANK0 MStencil/s MAX' in k.keys():
if(k['Total RANK0 MStencil/s MAX']!=''):
k['MStencil/s MAX'] = k['MWD main-loop RANK0 MStencil/s MAX']
# temporary for deprecated format
if 'RANK0 MStencil/s MAX' in k.keys():
if k['RANK0 MStencil/s MAX']!='':
k['MStencil/s MAX'] = k['RANK0 MStencil/s MAX']
# add stencil operator
k['stencil'] = get_stencil_num(k)
if k['stencil'] == 0:
k['stencil_name'] = '25_pt_const'
elif k['stencil'] == 1:
k['stencil_name'] = '7_pt_const'
elif k['stencil'] == 4:
k['stencil_name'] = '25_pt_var'
elif k['stencil'] == 5:
k['stencil_name'] = '7_pt_var'
elif k['stencil'] == 6:
k['stencil_name'] = 'solar'
# add the approach
if(k['Time stepper orig name'] == 'Spatial Blocking'):
k['method'] = 'Spt.blk.'
elif(k['Time stepper orig name'] in ['PLUTO', 'Pochoir']):
k['method'] = k['Time stepper orig name']
elif(k['Time stepper orig name'] == 'Diamond'):
if('_tgs1_' in k['file_name']):
k['method'] = '1WD'
else:
k['method'] = 'MWD'
else:
print("ERROR: Unknow time stepper")
raise
# add mwd type
k['mwdt']='none'
if(k['method'] == 'MWD'):
mwd = k['Wavefront parallel strategy'].lower()
if('fixed' in mwd) and ('relaxed' in mwd):
k['mwdt'] = 'fers'
elif('fixed' in mwd):
k['mwdt'] = 'fe'
elif('relaxed' in mwd):
k['mwdt'] = 'rs'
elif('wavefront' in mwd):
k['mwdt'] = 'block'
# add precision information
p = 1 if k['Precision'] in 'DP' else 0
k['Precision'] = p
# TLB measurement for LIKWID 4
if 'L1 DTLB load miss rate avg' in k.keys():
if k['L1 DTLB load miss rate avg']!='':
hw_ctr_fields['TLB'] = [('L1 DTLB load miss rate avg', float)]
hw_ctr_labels['TLB'] = [('L1 DTLB load miss rate avg', 'tlb_', 'tlb')]
entry = {}
# parse the general fileds' format
for f in req_fields + hw_ctr_fields[k['LIKWID performance counter']]:
try:
entry[f[0]] = map(f[1], [k[f[0]]] )[0]
except:
print("ERROR: results entry missing essential data at file:%s"%(k['file_name']))
print f[0]
print k
return
#find repeated data
key = (entry['Precision'], k['stencil_name'], k['LIKWID performance counter'], k['mwdt'], k['method'], entry['Global NX'])
if key not in duplicates:
duplicates.add(key)
else:
print("Repeated result at: %s"%(k['file_name']))
continue
# Initialize plot entry if does not exist for current data entry
# for m,n in entry.iteritems(): print m,n
measure_list = ['n', 'perf', 'total energy', 'tlb', 'mem bw', 'l2 bw', 'l3 bw', 'mem vol', 'l2 vol', 'l3 vol', 'data', 'tgs', 'thx', 'thy', 'thz', 'blk size', 'diam width', 'bs_z']
plot_key = (entry['Precision'], k['stencil_name'], k['LIKWID performance counter'])
line_key = (k['mwdt'], k['method'])
if plot_key not in meas_figs.keys():
meas_figs[plot_key] = {}
if line_key not in meas_figs[plot_key].keys():
meas_figs[plot_key][line_key] = {meas:[] for meas in measure_list}
# append the measurement data
meas_figs[plot_key][line_key]['n'].append(entry['Global NX'])
# meas_figs[plot_key][line_key]['perf'].append(entry['MStencil/s MAX']/1e3)
N = entry['Global NX']**3 * entry['Number of time steps'] * entry['Number of tests']/1e9
# Memory
if k['LIKWID performance counter'] == 'MEM':
meas_figs[plot_key][line_key]['mem bw'].append(entry['Sustained Memory BW']/1e3)
meas_figs[plot_key][line_key]['mem vol'].append(entry['Total Memory Transfer']/N)
# Energy
elif k['LIKWID performance counter'] == 'ENERGY':
entry['cpu energy pj/lup'] = entry['Energy']/N
entry['dram energy pj/lup'] = entry['Energy DRAM']/N
entry['total energy pj/lup'] = entry['cpu energy pj/lup'] + entry['dram energy pj/lup']
if (entry['total energy pj/lup'] < 3e3):
# entry['total energy pj/lup'] = 0
meas_figs[plot_key][line_key]['total energy'].append(entry['total energy pj/lup'])
else:
del meas_figs[plot_key][line_key]['n'][-1]
# TLB
elif k['LIKWID performance counter'] == 'TLB':
meas_figs[plot_key][line_key]['tlb'].append(entry[ hw_ctr_fields['TLB'][0][0] ])
# L2
elif k['LIKWID performance counter'] == 'L2':
meas_figs[plot_key][line_key]['l2 vol'].append(entry['L2 data volume sum']/N)
#L3
elif k['LIKWID performance counter'] == 'L3':
meas_figs[plot_key][line_key]['l3 vol'].append(entry['L3 data volume sum']/N)
#CPU
elif k['LIKWID performance counter'] == 'DATA':
meas_figs[plot_key][line_key]['data'].append(entry['Load to Store ratio avg'])
#Diamond tiling data
if(k['method'] == '1WD' or k['method'] == 'MWD'):
meas_figs[plot_key][line_key]['diam width'].append(int(k['Intra-diamond width']))
meas_figs[plot_key][line_key]['tgs'].append(int(k['Thread group size']))
meas_figs[plot_key][line_key]['thx'].append(int(k['Threads along x-axis']))
meas_figs[plot_key][line_key]['thy'].append(int(k['Threads along y-axis']))
meas_figs[plot_key][line_key]['thz'].append(int(k['Threads along z-axis']))
meas_figs[plot_key][line_key]['blk size'].append(int(k['Total cache block size (kiB)'])/1024.0)
meas_figs[plot_key][line_key]['bs_z'].append(int(k['Multi-wavefront updates']))
# append the performance data
plot_key = (entry['Precision'], k['stencil_name'])
line_key = (k['mwdt'], k['method'])
if plot_key not in perf_fig.keys(): # figure
perf_fig[plot_key] = dict()
perf_line = perf_fig[plot_key]
if line_key not in perf_line.keys(): # line
perf_line[line_key] = dict()
perf_point = perf_line[line_key]
nx = entry['Global NX']
if nx not in perf_point.keys(): # points
perf_point[nx] = [entry['MStencil/s MAX']/1e3]
else:
perf_point[nx].append(entry['MStencil/s MAX']/1e3)
del raw_data
#sort performance results
for k,v in perf_fig.iteritems():
for k2,v2 in perf_fig[k].iteritems():
perf_line = perf_fig[k][k2]
perf_fig[k][k2] = OrderedDict(sorted(perf_fig[k][k2].iteritems(), key=lambda x:x[0]))
# for k,v in perf_fig.iteritems():
# print(k, "##########")
# for k2,v2 in perf_fig[k].iteritems():
# print(k2,v2)
#sort the plot lines
for p in meas_figs:
for l in meas_figs[p]:
pl = meas_figs[p][l]
#remove unused fields
empty = []
for key, val in pl.iteritems():
if(val==[]):
empty.append(key)
for key in empty:
del pl[key]
lines = []
[lines.append(pl[val]) for val in measure_list if val in pl.keys()]
lines = sorted(zip(*lines))
idx=0
for val in measure_list:
if(val in pl.keys()):
if(pl[val]):
pl[val] = [x[idx] for x in lines]
idx = idx+1
# for m,n in meas_figs.iteritems():
# print "##############",m
# for i,j in n.iteritems():
# print i,j
plot_all(perf_fig, meas_figs, machine_name)
def main():
from scripts.utils import create_project_tarball, get_stencil_num
from scripts.conf.conf import machine_conf, machine_info
import os, sys
from csv import DictReader
import time, datetime
dry_run = 1 if len(sys.argv) < 2 else int(sys.argv[1])
time_stamp = datetime.datetime.fromtimestamp(
time.time()).strftime('%Y%m%d_%H_%M')
exp_name = "pluto_thread_scaling_at_%s_%s" % (machine_info['hostname'],
time_stamp)
tarball_dir = 'results/' + exp_name
if (dry_run == 0): create_project_tarball(tarball_dir, "test_" + exp_name)
target_dir = 'results/' + exp_name
# parse the results to find out which of the already exist
data = []
data_file = os.path.join('results', 'summary.csv')
try:
with open(data_file, 'rb') as output_file:
raw_data = DictReader(output_file)
for k in raw_data:
kernel = get_stencil_num(k)
if (kernel == 0):
k['stencil'] = '3d25pt'
elif (kernel == 1):
k['stencil'] = '3d7pt'
elif (kernel == 4):
k['stencil'] = '3d25pt_var'
elif (kernel == 5):
k['stencil'] = '3d7pt_var'
else:
raise
data.append(k)
except:
pass
param_l = dict()
for k in data:
try:
param_l[(k['stencil'], int(k['Global NX']),
k['LIKWID performance counter'])] = ([
int(k['PLUTO tile size of loop 1']),
int(k['PLUTO tile size of loop 3']),
int(k['PLUTO tile size of loop 4'])
], int(k['Number of time steps']))
except:
print k
raise
count = 0
for group in ['MEM', 'L2']:
# for group in ['MEM', 'L2', 'L3', 'DATA', 'TLB_DATA', 'ENERGY']:
if (machine_info['hostname'] == 'Haswell_18core'):
machine_conf['pinning_args'] = " -m -g " + group + " -C S1:0-"
elif (machine_info['hostname'] == 'IVB_10core'):
if group == 'TLB_DATA': group = 'TLB'
machine_conf['pinning_args'] = " -g " + group + " -C S0:0-"
# for k,v in param_l.iteritems(): print k,v
count = count + thread_scaling_test(
dry_run, target_dir, exp_name, param_l=param_l, group=group)
print "experiments count =" + str(count)
def parse_entry_info(k, is_tgs_only):
from scripts.utils import get_stencil_num, load_csv
# # get processor name from the file names
# if(k['OpenMP Threads']!=''):
# if(int(k['OpenMP Threads']) == 10):
# machine_name = 'ivb10'
# elif(int(k['OpenMP Threads']) == 18):
# machine_name = 'hw18'
# Use single field to represent the performance
if "Total RANK0 MStencil/s MAX" in k.keys():
if k["Total RANK0 MStencil/s MAX"] != "":
k["MStencil/s MAX"] = k["MWD main-loop RANK0 MStencil/s MAX"]
# temporary for deprecated format
if "RANK0 MStencil/s MAX" in k.keys():
if k["RANK0 MStencil/s MAX"] != "":
k["MStencil/s MAX"] = k["RANK0 MStencil/s MAX"]
# add stencil operator
k["stencil"] = get_stencil_num(k)
if k["stencil"] == 0:
k["stencil_name"] = "25_pt_const"
elif k["stencil"] == 1:
k["stencil_name"] = "7_pt_const"
elif k["stencil"] == 4:
k["stencil_name"] = "25_pt_var"
elif k["stencil"] == 5:
k["stencil_name"] = "7_pt_var"
elif k["stencil"] == 6:
k["stencil_name"] = "solar"
# add the approach
if k["Time stepper orig name"] == "Spatial Blocking":
k["method"] = "Spt.blk."
elif k["Time stepper orig name"] in ["PLUTO", "Pochoir"]:
k["method"] = k["Time stepper orig name"]
elif k["Time stepper orig name"] == "Diamond":
if "_tgs1_" in k["file_name"]:
k["method"] = "1WD"
else:
k["method"] = "MWD"
else:
print ("ERROR: Unknow time stepper")
raise
# TEMPORARY: Older versions of Pochoir did not have OpenMP threads count
if k["method"] == "Pochoir" and k["OpenMP Threads"] == "":
k["OpenMP Threads"] = 0
# add mwd type
k["mwdt"] = "none"
if k["method"] == "MWD":
mwd = k["Wavefront parallel strategy"].lower()
if ("fixed" in mwd) and ("relaxed" in mwd):
k["mwdt"] = "fers"
elif "fixed" in mwd:
k["mwdt"] = "fe"
elif "relaxed" in mwd:
k["mwdt"] = "rs"
elif "wavefront" in mwd:
k["mwdt"] = "block"
# add precision information
p = 1 if k["Precision"] in "DP" else 0
k["Precision"] = p
# TLB measurement for LIKWID 4
if "L1 DTLB load miss rate avg" in k.keys():
if k["L1 DTLB load miss rate avg"] != "":
hw_ctr_fields["TLB"] = [("L1 DTLB load miss rate avg", float)]
hw_ctr_labels["TLB"] = [("L1 DTLB load miss rate avg", "tlb_", "tlb")]
# parse the general fileds' format
for f in req_fields + hw_ctr_fields[k["LIKWID performance counter"]]:
try:
k[f[0]] = map(f[1], [k[f[0]]])[0]
except:
print ("ERROR: results entry missing essential data at file:%s" % (k["file_name"]))
print f[0]
print k
return
# Parse the L2 data per thread
if k["LIKWID performance counter"] == "L2":
k["L2 vol list"] = []
for i in range(k["OpenMP Threads"]):
field = "L2 data volume core %d" % i
if field in k.keys():
k["L2 vol list"].append(float(k["L2 data volume core %d" % i]))
k["tgsl"] = k["Thread group size"]
if is_tgs_only == 0: # regular mode for all MWD
if k["method"] == "MWD":
k["tgsl"] = 100
