def main():
stub = sys.argv[1]
cxn = SQLCxn(timeout=None, username='******', db='ubuntu')
if not cxn.table_exists('adclick_clean_vectors_split'):
shape = cxn.get_shape('adclick_clean{}_dense'.format(stub))
stmt = """
CREATE TABLE adclick_clean_vectors_split AS (
SELECT row_num, val[1]::INTEGER y, val[2:{}]::NUMERIC[] indep_vars
FROM adclick_clean{}_dense
) DISTRIBUTED BY (row_num)
""".format(shape[1], stub)
cxn.execute(stmt)
if not cxn.table_exists('adclick_clean_indepvars_long'):
stmt = """
CREATE TABLE adclick_clean_indepvars_long AS (
SELECT row_num, ix AS col_num, indep_vars[ix] AS val
FROM (
SELECT *, GENERATE_SUBSCRIPTS(indep_vars, 1) AS ix
FROM adclick_clean_vectors_split
) tmp
) DISTRIBUTED BY (row_num, col_num)
"""
cxn.execute(stmt)
if not cxn.table_exists('adclick_clean_y'):
stmt = """
CREATE TABLE adclick_clean_y AS (
SELECT row_num, 1 AS col_num, y AS val
FROM adclick_clean_vectors_split
) DISTRIBUTED BY (row_num)
"""
cxn.execute(stmt)
mpath_tall = os.path.abspath(
'../output/M{}{}_sparse_tall.mtx'.format(*fmt))
mpath_wide = os.path.abspath(
'../output/M{}{}_sparse_wide.mtx'.format(*fmt))
data.gen_data_sparse(k, 100, sr, 'M{}{}_sparse_tall'.format(*fmt),
mpath_tall)
data.gen_data_sparse(100, k, sr, 'M{}{}_sparse_wide'.format(*fmt),
mpath_wide)
data.gen_data_disk('../output/y{}_sparse.csv'.format(sparse_gb), k, 1, k,
True)
stmt = """
CREATE VIEW N{0}{1}_sparse_tall AS (
SELECT * FROM M{0}{1}_sparse_tall
)
""".format(*fmt)
if not cxn.table_exists('N{}{}_sparse_tall'.format(*fmt)):
cxn.execute(stmt)
utils.link_if_not('../output/M{}{}_sparse_tall.mtx'.format(*fmt),
'../output/N{}{}_sparse_tall.mtx'.format(*fmt))
cxn.load_dense_matrix('../output/y{}_sparse.csv'.format(sparse_gb),
'y{}_sparse'.format(sparse_gb))
paths = os.listdir('../output')
paths = filter(
lambda x: (x != '.gitignore') and ('.log' not in x) and ('.mtd' not in x),
paths)
paths = map(lambda x: os.path.join('../output', x), paths)
with open('manifest.txt') as fh:
m = k
rows = 2**12
elif mtype == 'tall':
k = int(np.ceil((gb * 1e9) / float(8 * 100)))
m = 100
rows = 2**14
elif mtype == 'wide':
k = 100
m = int(np.ceil((gb * 1e9) / float(8 * 100)))
rows = 1
stub = '_' + mtype
fmt = (gb_stub, stub)
data.gen_data_disk('../output/M{}{}.csv'.format(*fmt), k, m, rows)
if ((mtype == 'wide')
and (not cxn.table_exists('M{}{}'.format(*fmt)))):
print 'CREATING MATRIX: M{}{}'.format(*fmt)
cxn.randomMatrix(k, m, 'M{}{}'.format(*fmt))
if mtype != 'tall':
continue
mpath = os.path.abspath('../output/M{}{}_sparse.mtx'.format(*fmt))
data.gen_data_disk('../output/y{}{}.csv'.format(*fmt), k, 1, rows,
True)
utils.link_if_not('../output/M{}{}.csv'.format(*fmt),
'../output/N{}{}.csv'.format(*fmt))
utils.link_if_not('../output/M{}{}.csv.mtd'.format(*fmt),
'../output/N{}{}.csv.mtd'.format(*fmt))
paths = os.listdir('../output')
paths = filter(
lambda x: (x != '.gitignore') and ('.log' not in x) and ('.mtd' not in x),
def main(kwargs):
op_type = kwargs['opType']
nodes = kwargs['nodes']
stub = kwargs['stub']
colnames = [
'nodes', 'rows', 'cols', 'time1', 'time2', 'time3', 'time4', 'time5'
]
runTimes = pd.DataFrame(np.zeros((1, len(colnames))))
runTimes.columns = colnames
cxn = SQLCxn(username='******', db='ubuntu')
shape = cxn.get_shape_dense('adclick_clean{}_dense'.format(stub))
if not cxn.table_exists('adclick_clean_vectors_split'):
stmt = """
CREATE TABLE adclick_clean_vectors_split AS (
SELECT row_num, val[1]::INTEGER y, val[2:{}]::NUMERIC[] indep_vars
FROM adclick_clean{}_dense
) DISTRIBUTED BY (row_num)
""".format(shape[1], stub)
cxn.execute(stmt)
# need to do a bit of preprocessing
if op_type == 'logit':
cxn.execute('DROP TABLE IF EXISTS adclick_logit_summary')
cxn.execute('DROP TABLE IF EXISTS adclick_logit')
call = """
SELECT madlib.logregr_train('adclick_clean_vectors_split',
'adclick_logit',
'y', 'indep_vars', NULL,
3, 'igd', .000001)
"""
cleanup = ['adclick_logit_summary', 'adclick_logit']
elif op_type == 'reg':
cxn.execute('DROP TABLE IF EXISTS adclick_reg_summary')
cxn.execute('DROP TABLE IF EXISTS adclick_reg')
call = """
SELECT madlib.linregr_train('adclick_clean_vectors_split',
'adclick_reg', 'y', 'indep_vars')
"""
cleanup = ['adclick_reg_summary', 'adclick_reg']
elif op_type == 'pca':
cxn.execute('DROP TABLE IF EXISTS result_table')
cxn.execute('DROP TABLE IF EXISTS result_table_mean')
cxn.execute('DROP TABLE IF EXISTS residual_table')
cxn.execute('DROP TABLE IF EXISTS result_summary_table')
cxn.execute('DROP TABLE IF EXISTS adlick_prj')
stmt = """
CREATE TABLE adclick_clean_depvars AS (
SELECT row_num, val[2:{}]::NUMERIC[] val
FROM adclick_clean{}_dense
) DISTRIBUTED BY (row_num)
""".format(shape[1], stub)
if not cxn.table_exists('adclick_clean_depvars'):
cxn.execute(stmt)
call = """
SELECT madlib.pca_train('adclick_clean_depvars',
'result_table',
'row_num',
5);
SELECT madlib.pca_project('adclick_clean_depvars',
'result_table',
'adclick_prj',
'row_num',
'residual_table',
'result_summary_table')
"""
cleanup = [
'result_table', 'result_table_mean', 'residual_table',
'result_summary_table', 'adclick_prj'
]
#shape = cxn.get_shape_dense('adclick_clean{}_dense'.format(stub))
runTimes.ix[:, ['rows', 'cols']] = shape
path = '../output/madlib_{}{}_dense.txt'.format(op_type, int(nodes))
runTimes.ix[:, 'nodes'] = nodes
res = cxn.time(call, cleanup)
runTimes.ix[:, 3:] = res
runTimes.to_csv(path, index=False)
def doMatrixOp(kwargs):
opType = kwargs.get('opType')
mattype = kwargs.get('mattype')
fixedAxis = int(kwargs.get('fixedAxis'))
nrow_scale = map(lambda x: int(x), kwargs['nrows'].split(' '))
nproc = kwargs.get('nproc')
port = GPDB_PORT_MAP[nproc] if nproc is not None else None
if nproc is not None:
cxn = start_gpdb(port, nproc)
cxn.execute('DROP TABLE IF EXISTS M16_tall')
atexit.register(stop_gpdb, nproc, cxn)
else:
cxn = SQLCxn(username='******', db='ubuntu', timeout=10000)
colnames = ['rows', 'time1', 'time2', 'time3', 'time4', 'time5']
runTimes = pd.DataFrame(np.zeros((1, len(colnames))))
runTimes.columns = colnames
if nproc is None:
path = os.path.join('..', 'output',
'madlib_{}_{}.txt'.format(mattype, opType))
else:
path = os.path.join('..', 'output',
'madlib_cpu_{}_scale.txt'.format(opType))
for nr in nrow_scale:
nrow = fixedAxis if opType == 'GMM' else nr
ncol = nr if opType == 'GMM' else fixedAxis
print nrow
print ncol
Mname = 'M{}{}'.format(nrow, ncol)
if not cxn.table_exists('M{}{}'.format(nrow, ncol)):
cxn.randomMatrix(nrow, ncol, 'M{}{}'.format(nrow, ncol))
if (opType == 'GMM'):
if not cxn.table_exists('N{}{}'.format(ncol, nrow)):
cxn.randomMatrix(ncol, nrow, 'N{}{}'.format(ncol, nrow))
Nname = 'N{}{}'.format(ncol, nrow)
elif (opType == 'ADD'):
if not cxn.table_exists('N{}{}'.format(nrow, ncol)):
cxn.randomMatrix(nrow, ncol, 'N{}{}'.format(nrow, ncol))
Nname = 'N{}{}'.format(nrow, ncol)
cleanup = []
if (opType == 'TRANS'):
call = "matrix_trans('{}',NULL,'Mt',NULL)".format(Mname)
cleanup.append('Mt')
elif (opType == 'NORM'):
call = "matrix_norm('{}',NULL,'fro')".format(Mname)
elif (opType == 'GMM'):
call = "matrix_mult('{}',NULL,'{}',NULL,'MN',NULL)".format(
Mname, Nname)
cleanup.append('MN')
elif (opType == 'MVM'):
array_call = 'SELECT array_agg(random()) FROM generate_series(1,{})'.format(
ncol)
call = "matrix_vec_mult('{}',NULL,({}))".format(Mname, array_call)
elif (opType == 'TSM'):
call = "matrix_mult('{0}','trans=True','{0}',NULL,'MtM',NULL)".format(
Mname)
cleanup.append('MtM')
elif (opType == 'ADD'):
call = "matrix_add('{}',NULL,'{}',NULL,'M_N',NULL)".format(
Mname, Nname)
cleanup.append('M_N')
else:
raise NotImplementedError('Invalid Operation')
sql_call = 'SELECT madlib.{}'.format(call)
runTimes.ix[:, 'rows'] = nr if nproc is None else nproc
runTimes.ix[:, 1:] = cxn.time(sql_call, cleanup)
writeHeader = False if (os.path.exists(path)) else True
runTimes.to_csv(path, index=False, header=writeHeader, mode='a')
def main(kwargs):
op_type = kwargs['opType']
nodes = kwargs['nodes']
stub = kwargs['stub']
colnames = [
'nodes', 'rows', 'cols', 'time1', 'time2', 'time3', 'time4', 'time5'
]
runTimes = pd.DataFrame(np.zeros((1, len(colnames))))
runTimes.columns = colnames
cxn = SQLCxn(username='******', db='ubuntu', timeout=10000)
shape = cxn.get_shape('adclick_clean_1_sparse')
if not cxn.table_exists('adclick_clean_1_vectors_sparse'):
stmt = """
CREATE TABLE adclick_clean_1_vectors_sparse AS (
SELECT x.row_num, madlib.svec_cast_positions_float8arr(
ARRAY_AGG(x.col_num), ARRAY_AGG(x.val), {}, 0.0
) AS indep_vars, y.val AS y
FROM adclick_clean_1_sparse x
INNER JOIN adclick_clean_y y ON x.row_num = y.row_num
GROUP BY x.row_num, y.val
) DISTRIBUTED BY (row_num)
""".format(shape[1])
cxn.execute(stmt)
if op_type == 'logit':
cxn.execute('DROP TABLE IF EXISTS adclick_logit_summary')
cxn.execute('DROP TABLE IF EXISTS adclick_logit')
call = """
SELECT madlib.logregr_train('adclick_clean_1_vectors_sparse',
'adclick_logit',
'y', 'indep_vars', NULL,
3, 'igd', .000001)
"""
cleanup = ['adclick_logit_summary', 'adclick_logit']
elif op_type == 'reg':
cxn.execute('DROP TABLE IF EXISTS adclick_reg_summary')
cxn.execute('DROP TABLE IF EXISTS adclick_reg')
call = """
SELECT madlib.linregr_train('adclick_clean_1_vectors_sparse',
'adclick_reg', 'y', 'indep_vars')
"""
cleanup = ['adclick_reg_summary', 'adclick_reg']
elif op_type == 'pca':
cxn.execute('DROP TABLE IF EXISTS result_table')
cxn.execute('DROP TABLE IF EXISTS result_table_mean')
cxn.execute('DROP TABLE IF EXISTS residual_table')
cxn.execute('DROP TABLE IF EXISTS result_summary_table')
cxn.execute('DROP TABLE IF EXISTS adlick_prj')
call = """
SELECT madlib.pca_sparse_train('adclick_clean_1_sparse',
'result_table',
'row_num',
'col_num',
'val',
'{0}',
'{1}',
5);
SELECT madlib.pca_sparse_project('adclick_clean_1_sparse',
'result_table',
'adclick_prj',
'row_num',
'col_num',
'val',
'{0}',
'{1}',
'residual_table',
'result_summary_table')
""".format(*shape)
cleanup = [
'result_table', 'result_table_mean', 'residual_table',
'result_summary_table', 'adclick_prj'
]
runTimes.ix[:, ['rows', 'cols']] = shape
path = '../output/madlib_{}{}_sparse.txt'.format(op_type, int(nodes))
runTimes.ix[:, 'nodes'] = nodes
res = cxn.time(call, cleanup)
runTimes.ix[:, 3:] = res
runTimes.to_csv(path, index=False)