def getMetaData(explist):
# Get DataFrame of All Experiments
df_All = analysis.getAllExperiments()
# Get Cori Runs
df_Cori = df_All.loc[df_All['Experiment'].isin(explist)]
# Create Meta DataFrame
df_Cori_Meta = pd.DataFrame(
columns=['Workload', 'Ranks', 'Stencil', 'Runtime'])
# Loops Over All Cori Runs
for df in [df_Cori]:
for run in range(0, len(df)):
currentRun = df.iloc[run]
eid = currentRun['Experiment']
workload = currentRun['workload']
ranks = currentRun['processors']
stencil = currentRun['stencil_size']
rab_work = currentRun['rabbit_workload']
rid = currentRun['expid']
currentPath = './mlruns/' + str(eid) + '/' + str(
rid) + '/artifacts/bsp-trace.json'
# Print Progress
print(eid, currentPath)
# Check that Rabbit is Off
if (rab_work == 0):
# Get Data
currentData = analysis.getData(currentPath)
currentData = currentData[currentData['rank'] == 0]
# Calculate Total Runtime
currentRunTime = currentData['interval_max_usec'].sum(
) / 1000000
# Add to Metadata
df_Cori_Meta.loc[len(df_Cori_Meta)] = [
workload,
int(ranks),
int(stencil), currentRunTime
]
return df_Cori_Meta
def get_formated_data(df_platform_NoStencil, workload, processors, fname):
if path.isfile(fname):
data = np.load(fname)
else:
ppn = 32
data_platform = analysis.getData(df_platform_NoStencil)
data_platform = data_platform[[
'workload', 'node', 'rank', 'iteration', 'iterations',
'workload_usec'
]]
data_platform_Iteration0 = data_platform[(
data_platform['iteration'] == 0)]
print(data_platform.head())
nodes = data_platform['node'].unique()
node_counter = np.zeros((len(nodes), 1))
iterations = data_platform['iterations'].iloc[0]
data = np.zeros((int(processors / ppn), ppn, iterations))
for item in range(0, len(data_platform_Iteration0)):
cur = data_platform_Iteration0.iloc[item]
for node in nodes:
if node == cur['node']:
i = int(np.where(nodes == node)[0][0])
j = int(node_counter[i])
node_counter[i] = node_counter[i] + 1
curData = data_platform[data_platform['rank'] == cur['rank']]
for k in range(0, iterations):
temp = curData.iloc[k]
data[i, j, k] = temp['workload_usec']
# save data into a npy
np.save(fname, data)
return data
def generatePredictionFigure( df_All, workload, baseRanks, CI, fname ):
print( 'Generating Prediction for: ' + workload )
outfile1 = 'Figures/Attaway_' + workload + '_prediction'
outfile2 = 'Figures/Attaway_' + workload + '_prediction_stencil'
coriData = pd.read_csv( fname )
workloadData = coriData[ ( coriData['Workload'] == workload ) & ( coriData['Ranks'] == baseRanks )]
kvals = np.array( [1, 2, 4, 8, 16] )
iterations = 50
MILLION = 1000000
eps = 10 ** -5
yMin = 0
yMax = 0
count = 0
fig1 = plt.figure( 1, figsize=( 7, 3 ) )
fig2 = plt.figure( 2, figsize=( 7, 3 ) )
count = 0
labels = ['No Stencil', 'Stencil']
if workload == 'hpcg' or workload == 'lammps':
df_Attaway_NoStencil = df_All[ ( df_All['machine'] == 'Attaway' ) & ( df_All['processors'] == baseRanks ) & ( df_All['workload'] == workload ) & ( df_All['stencil_size'] == 0 ) ]
df_List = [df_Attaway_NoStencil]
else:
df_Attaway_NoStencil = df_All[ ( df_All['machine'] == 'Attaway' ) & ( df_All['processors'] == baseRanks ) & ( df_All['workload'] == workload ) & ( df_All['stencil_size'] == 0 ) ]
df_Attaway_Stencil = df_All [ ( df_All['machine'] == 'Attaway' ) & ( df_All['processors'] == baseRanks ) & ( df_All['workload'] == workload ) & ( df_All['stencil_size'] != 0 ) ]
df_List = [df_Attaway_NoStencil, df_Attaway_Stencil]
for df in df_List:
print( labels[count] )
lowerBound = []
middle = []
upperBound = []
if ( count == 0 ):
runtimeData = workloadData[workloadData['Stencil'] == 0]['Runtime']
else:
runtimeData = workloadData[workloadData['Stencil'] != 0]['Runtime']
runtimeData = runtimeData.sort_values()
minRuntime = min( runtimeData )
medianRuntime = runtimeData.iloc[ int( len( runtimeData ) / 2 ) ]
maxRuntime = max( runtimeData )
print( workload, minRuntime, medianRuntime, maxRuntime )
for k in kvals:
expListMin = []
expListMed = []
expListMax = []
for run in range( 0, len( df ) ):
currentRun = df.iloc[run]
eid = currentRun['Experiment']
rid = currentRun['expid']
currentPath = './mlruns/' + str( eid ) + '/' + str( rid ) + '/artifacts/bsp-trace.json'
currentData = analysis.getData( currentPath )
currentData = currentData[currentData['rank'] == 0]
currentRuntime = sum( currentData['interval_max_usec'] ) / MILLION
if ( abs( currentRuntime - maxRuntime ) < eps ) or ( abs( currentRuntime - medianRuntime ) < eps ) or ( abs( currentRuntime - minRuntime ) < eps ):
for i in range( iterations ):
data = analysis.resample_project( currentData, len( currentData ), k, col='interval_max_usec' )
projectedRunTime = sum( data ) / MILLION
print( 'Run:', run + 1, '/', len( df ), '\tIteration:', i + 1, '/', iterations, '\tWorkload:', workload, '\tk:', k, '\tProjected Runtime:', projectedRunTime )
if ( abs( currentRuntime - minRuntime ) < eps ):
expListMin.append( projectedRunTime )
if ( abs( currentRuntime - medianRuntime ) < eps ):
expListMed.append( projectedRunTime )
if ( abs( currentRuntime - maxRuntime ) < eps ):
expListMax.append( projectedRunTime )
expListMin.sort()
expListMed.sort()
expListMax.sort()
meanMin = np.mean( np.array( expListMin ) )
stvMin = stats.sem( np.array( expListMin ) )
meanMed = np.mean( np.array( expListMed ) )
stvMed = stats.sem( np.array( expListMed ) )
meanMax = np.mean( np.array( expListMax ) )
stvMax = stats.sem( np.array( expListMax ) )
intervalMin = stats.norm.interval( CI, loc=meanMin, scale=stvMin )
intervalMed = stats.norm.interval( CI, loc=meanMed, scale=stvMed )
intervalMax = stats.norm.interval( CI, loc=meanMax, scale=stvMax )
print( CI, '$ Confidence Interval:', intervalMin[0], intervalMax[1], '\tMedian:', ( intervalMed[0] + intervalMed[1] ) / 2 )
lowerBound.append( intervalMin[0] )
middle.append( ( intervalMed[0] + intervalMed[1] ) / 2 )
upperBound.append( intervalMax[1] )
if count == 0:
plt.figure( 1 )
_ = plt.plot( baseRanks * kvals, middle, color='black' )
_ = plt.fill_between( baseRanks * kvals, lowerBound, upperBound, color='blue', alpha=0.9 )
if count == 1:
plt.figure( 2 )
_ = plt.plot( baseRanks * kvals, middle, color='black' )
_ = plt.fill_between( baseRanks * kvals, lowerBound, upperBound, color='blue', alpha=0.9 )
if count == 0:
yMin = min( lowerBound)
yMax = max( upperBound)
else:
if yMin > min( lowerBound ):
yMin = min( lowerBound )
if yMax < max( upperBound ):
yMax = max( upperBound )
count = count + 1
attawayData = pd.read_csv( fname )
baseranks0 = list()
baseRuntime0 = list()
ranks0 = list()
currentRuntime0 = list()
baseranks1 = list()
baseRuntime1 = list()
ranks1 = list()
currentRuntime1 = list()
for run in range( 0, len( attawayData ) ):
label = ''
myworkload = attawayData['Workload'][run]
ranks = attawayData['Ranks'][run]
stencil = attawayData['Stencil'][run]
currentRunTime = attawayData['Runtime'][run]
if int( stencil ) > 0:
label = label + 'Stencil'
if label == '':
if myworkload == workload:
if ranks == baseRanks:
baseranks0.append( int( ranks ) )
baseRuntime0.append( currentRunTime )
print( 'No Stencil - BaseRanks', workload, ranks, currentRunTime )
else:
ranks0.append( int( ranks ) )
currentRuntime0.append( currentRunTime )
print( 'No Stencil - Other', workload, ranks, currentRunTime )
if not ( workload == 'hpcg' or workload == 'lammps' ):
if label == 'Stencil':
if myworkload == workload:
if ranks == baseRanks:
baseranks1.append( int( ranks ) )
baseRuntime1.append( currentRunTime )
print( 'Stencil - BaseRanks', workload, ranks, currentRunTime )
else:
ranks1.append( int( ranks ) )
currentRuntime1.append( currentRunTime )
print( 'Stencil - Other', workload, ranks, currentRunTime )
print( baseranks0 )
print( baseRuntime0 )
print( ranks0 )
print( currentRuntime0 )
print( baseranks1 )
print( baseRuntime1 )
print( ranks1 )
print( currentRuntime1 )
if ( workload == 'sleep' ):
w = 'ftq'
else:
w = workload
plt.figure( 1 )
_ = plt.plot( baseranks0, baseRuntime0, 'o', color='black', label='Sample workload' )
_ = plt.plot( ranks0, currentRuntime0, 'o', color='red', label='Scaled-up workload' )
_ = plt.ylim( 0.95 * yMin, 1.05 * yMax )
_ = plt.title( 'Attaway ' + w + ' - per run bootstrap, global CIs' )
_ = plt.xlabel( 'Number of Ranks' )
_ = plt.ylabel( 'Runtime (s)' )
_ = plt.legend( loc='lower right', borderaxespad=0.5 )
plt.tight_layout()
fig1.savefig( outfile1 )
plt.close( fig1 )
if ( workload != 'hpcg' and workload != 'lammps' ):
plt.figure( 2 )
_ = plt.plot( baseranks1, baseRuntime1, 'o', color='black', label='Sample workload' )
_ = plt.plot( ranks1, currentRuntime1, 'o', color='red', label='Scaled-up workload' )
_ = plt.ylim( 0.95 * yMin, 1.05 * yMax )
_ = plt.title( 'Attaway ' + w + ' - per run bootstrap, global CIs - Stencil' )
_ = plt.xlabel( 'Number of Ranks' )
_ = plt.ylabel( 'Runtime (s)' )
_ = plt.legend( loc='lower right', borderaxespad=0.5 )
plt.tight_layout()
fig2.savefig( outfile2 )
plt.close( fig2 )