def main():
from utils import create_project_tarball
import socket
hostname = socket.gethostname()
exp_name = "emmy_omp_barrier"
exe_cmd = "export OMP_NUM_THREADS=10; likwid-perfctr -m -C S0:0-9 -s 0x03 -g L3 "
tarball_dir='results/'+exp_name
create_project_tarball(tarball_dir, "project_"+exp_name)
target_dir='results/' + exp_name
cs = 4096
is_dp = 1
ts = 2
kernel = 1
for tgs in [1, 2, 5, 10]:
for N in [400, 608, 992]:
outfile=(exp_name + '_isDP%d_ts%d_kernel%d_tgs%d_N%d_%s.txt' % (is_dp, ts, kernel,tgs, N, hostname))
nx = N
ny = N
nz = N
nt = max(int(10 * 4e9/(nx*ny*nz*3)), 50)
submit_experiment(kernel, ts=ts, nx=nx, ny=ny, nz=nz, nt=nt, is_dp=is_dp,
outfile=outfile, target_dir=target_dir, tgs=tgs, cs=cs, exe_cmd=exe_cmd)
def main():
from utils import create_project_tarball
import socket
# hostname = socket.gethostname()
exp_name = "hasep1_"
exp_name = exp_name + "threadscaling"#_at_%s" % (hostname)
tarball_dir='results/'+exp_name
create_project_tarball(tarball_dir, "project_"+exp_name)
target_dir='results/' + exp_name
th_l = [1,2,3,4,5,6,7,8,9,10,11,12,13,14]
cs = 4096
for tgs in [1,2,7,14]:
# ts_test(target_dir, exp_name, cs, ts=2, kernel=1, tgs=tgs, N=960, th_l=th_l)
# ts_test(target_dir, exp_name, cs, ts=2, kernel=4, tgs=tgs, N=480, th_l=th_l)
ts_test(target_dir, exp_name, cs, ts=2, kernel=5, tgs=tgs, N=680, th_l=th_l)
cs = 16384
# ts_test(target_dir, exp_name[:-8], cs, ts=0, kernel=1, tgs=0, N=960, th_l=[8])
# ts_test(target_dir, exp_name[:-8], cs, ts=0, kernel=4, tgs=0, N=480, th_l=[3, 7, 10])
ts_test(target_dir, exp_name, cs, ts=0, kernel=5, tgs=1, N=680, th_l=th_l)
def main(): from utils import create_project_tarball import socket hostname = socket.gethostname() exp_name = "hw14c_" exp_name = exp_name + "threadscaling_energy_at_%s" % (hostname) tarball_dir='results/'+exp_name create_project_tarball(tarball_dir, "project_"+exp_name) target_dir='results/' + exp_name th_l = [1, 2, 4, 6, 8, 10, 12, 14] cs = 4096 #ts_test(target_dir, exp_name[:-9], cs, ts=2, kernel=1, tgs=1, N=960, th_l=th_l[1:]) #ts_test(target_dir, exp_name[:-9], cs, ts=2, kernel=1, tgs=2, N=960, th_l=th_l[4:]) #ts_test(target_dir, exp_name[:-9], cs, ts=2, kernel=1, tgs=7, N=960, th_l=[14]) #for i in th_l: # ts_test(target_dir, exp_name[:-9], cs, ts=2, kernel=1, tgs=i, N=960, th_l=[i]) cs = 8192 ts_test(target_dir, exp_name[:-9], cs, ts=0, kernel=1, tgs=1, N=960, th_l=th_l)
def main():
from utils import create_project_tarball
import socket
hostname = socket.gethostname()
exp_name = "emmy_energy_f2.2GHz_at_%s" % hostname
tarball_dir='results/'+exp_name
create_project_tarball(tarball_dir, "project_"+exp_name)
target_dir='results/' + exp_name
wf = 1
is_dp = 1
th = 10
# (stencil, N, nt, experiments)
exp = [(1, 960, 100, [(0,0,0,0)])]; th = 6 # special case for spatial blocking
#exp = [(1, 960, 100, [(2, 2, 5, 12), (2, 5, 11, 20), (2, 10, 11, 50), (0,0,0,0), (2, 1, 5, 1)])]
#exp = exp + [(5, 680, 50, [(2, 2, 3, 2), (2, 5, 3, 15), (2, 10, 9, 10), (0,0,0,0), (2, 1, 3, 1)])]
#exp = exp + [(4, 480, 100, [(2, 2, 1, 2), (2, 5, 1, 5), (2, 10, 3, 10), (0,0,0,0), (2, 1, 1, 9)])]
for kernel, N, nt, exp_l in exp:
for ts, tgs, tb, nwf in exp_l:
outfile=(exp_name + '_ts%d_tgs%d_tb%d_nwf%d_kernel%d_isDP%d_threads%d.txt' % (ts, tgs, tb, nwf, kernel, is_dp, th))
submit_emmy_experiment(N=N, nt=nt, nwf=nwf, tgs=tgs, wf=wf, is_dp=is_dp, th=th, ts=ts, tb=tb, kernel=kernel, outfile=outfile, target_dir=target_dir)
def main():
from utils import create_project_tarball
import socket
# hostname = socket.gethostname()
exp_name = "hasep1_"
exp_name = exp_name + "threadscaling" #_at_%s" % (hostname)
tarball_dir = 'results/' + exp_name
create_project_tarball(tarball_dir, "project_" + exp_name)
target_dir = 'results/' + exp_name
th_l = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14]
cs = 4096
for tgs in [1, 2, 7, 14]:
# ts_test(target_dir, exp_name, cs, ts=2, kernel=1, tgs=tgs, N=960, th_l=th_l)
# ts_test(target_dir, exp_name, cs, ts=2, kernel=4, tgs=tgs, N=480, th_l=th_l)
ts_test(target_dir,
exp_name,
cs,
ts=2,
kernel=5,
tgs=tgs,
N=680,
th_l=th_l)
cs = 16384
# ts_test(target_dir, exp_name[:-8], cs, ts=0, kernel=1, tgs=0, N=960, th_l=[8])
# ts_test(target_dir, exp_name[:-8], cs, ts=0, kernel=4, tgs=0, N=480, th_l=[3, 7, 10])
ts_test(target_dir, exp_name, cs, ts=0, kernel=5, tgs=1, N=680, th_l=th_l)
def main():
from utils import create_project_tarball
import socket
# hostname = socket.gethostname()
exp_name = "emmy_"
exp_name = exp_name + "threadscaling_energy" #_at_%s" % (hostname)
tarball_dir = 'results/' + exp_name
create_project_tarball(tarball_dir, "project_" + exp_name)
target_dir = 'results/' + exp_name
th_l = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
cs = 4096
for tgs in [1, 2, 5, 10]:
# ts_test(target_dir, exp_name, cs, ts=2, kernel=1, tgs=tgs, N=960, th_l=th_l)
# ts_test(target_dir, exp_name, cs, ts=2, kernel=4, tgs=tgs, N=480, th_l=th_l)
ts_test(target_dir,
exp_name,
cs,
ts=2,
kernel=5,
tgs=tgs,
N=680,
th_l=th_l)
def main():
from utils import create_project_tarball
import socket
hostname = socket.gethostname()
exp_name = "emmy_socket_thread_scaling_7pt_const_960cube_f2.2GHz_naive_no_spt_blk_at_%s" % hostname
tarball_dir = 'results/' + exp_name
create_project_tarball(tarball_dir, "project_" + exp_name)
target_dir = 'results/' + exp_name
kl = [1]
wf = 1
for th in list(range(10, 0, -1)):
for is_dp in [1]:
for ts, tgs, tb, nwf in [(0, 0, 0, 0)]:
for kernel in kl:
outfile = (
exp_name +
'_ts%d_tb%d_tgs%d_nwf%d_kernel%d_isDP%d_threads%d.txt'
% (ts, tb, tgs, nwf, kernel, is_dp, th))
submit_emmy_experiment(nwf=nwf,
tgs=tgs,
wf=wf,
is_dp=is_dp,
th=th,
ts=ts,
tb=tb,
kernel=kernel,
outfile=outfile,
target_dir=target_dir)
def main():
from utils import create_project_tarball
import socket
hostname = socket.gethostname()
exp_name = "hw14c_"
exp_name = exp_name + "threadscaling_energy_at_%s" % (hostname)
tarball_dir = 'results/' + exp_name
create_project_tarball(tarball_dir, "project_" + exp_name)
target_dir = 'results/' + exp_name
th_l = [1, 2, 4, 6, 8, 10, 12, 14]
cs = 4096
#ts_test(target_dir, exp_name[:-9], cs, ts=2, kernel=1, tgs=1, N=960, th_l=th_l[1:])
#ts_test(target_dir, exp_name[:-9], cs, ts=2, kernel=1, tgs=2, N=960, th_l=th_l[4:])
#ts_test(target_dir, exp_name[:-9], cs, ts=2, kernel=1, tgs=7, N=960, th_l=[14])
#for i in th_l:
# ts_test(target_dir, exp_name[:-9], cs, ts=2, kernel=1, tgs=i, N=960, th_l=[i])
cs = 8192
ts_test(target_dir,
exp_name[:-9],
cs,
ts=0,
kernel=1,
tgs=1,
N=960,
th_l=th_l)
def main():
from utils import create_project_tarball
import socket
hostname = socket.gethostname()
exp_name = "emmy_socket_thread_scaling_naive_inL3_1to10_threads_f2p2GHz_at_%s" % hostname
tarball_dir = 'results/' + exp_name
create_project_tarball(tarball_dir, "project_" + exp_name)
target_dir = 'results/' + exp_name
size_l = {
0: (128, 64, 64, 1e4),
1: (96, 96, 96, 4e4),
4: (64, 32, 32, 5e4),
5: (64, 48, 48, 1e5)
}
for th in list(range(10, 0, -1)):
for is_dp in [1]:
for kernel in [0, 1, 4, 5]:
nx, ny, nz, nt = size_l[kernel]
outfile = (exp_name + 'naive_kernel%d_isDP%d_threads%d' %
(kernel, is_dp, th))
submit_emmy_experiment(nt=nt,
nx=nx,
ny=ny,
nz=nz,
is_dp=is_dp,
th=th,
kernel=kernel,
outfile=outfile,
target_dir=target_dir)
def main():
from utils import create_project_tarball
import socket
# hostname = socket.gethostname()
exp_name = "emmy_cache_size_vs_code_balace_2p2GHz" # _at_%s" % hostname
tarball_dir = "results/" + exp_name
create_project_tarball(tarball_dir, "project_" + exp_name)
target_dir = "results/" + exp_name
kernel = 1
exp = []
exp = exp + [[0, 0, 960, 30, 0]]
# exp = exp + [[0, 0, 480, 400, 0]]
# exp = exp + [[2, i, 480, 400, 10] for i in range(1, 30, 2)]
exp = exp + [[2, i, 960, 50, 10] for i in range(1, 30, 2)]
for ts, tb, N, nt, tgs in exp:
outfile = "ts%d_kernel%d_tb%d_tgs%d_N%d.txt" % (ts, kernel, tb, tgs, N)
submit_emmy_experiment(
N, N, N, nt, is_dp=1, ts=ts, tb=tb, kernel=kernel, outfile=outfile, target_dir=target_dir, tgs=tgs, nwf=tgs
)
kernel = 5
exp = []
# exp = exp + [[0, 0, 384, 1000, 0]]
exp = exp + [[0, 0, 680, 30, 0]]
# exp = exp + [[2, i, 384, 1000, 10] for i in range(1, 20, 2)]
exp = exp + [[2, i, 680, 50, 10] for i in range(1, 20, 2)]
for ts, tb, N, nt, tgs in exp:
outfile = "ts%d_kernel%d_tb%d_tgs%d_N%d.txt" % (ts, kernel, tb, tgs, N)
submit_emmy_experiment(
N, N, N, nt, is_dp=1, ts=ts, tb=tb, kernel=kernel, outfile=outfile, target_dir=target_dir, tgs=tgs, nwf=tgs
)
kernel = 4
exp = []
# exp = exp + [[0, 0, 240, 500, 0]]
exp = exp + [[0, 0, 480, 30, 0]]
# exp = exp + [[2, i, 240, 500, 10] for i in range(1, 8, 2)]
exp = exp + [[2, i, 480, 50, 10] for i in range(1, 6, 2)]
for ts, tb, N, nt, tgs in exp:
outfile = "ts%d_kernel%d_tb%d_tgs%d_N%d.txt" % (ts, kernel, tb, tgs, N)
submit_emmy_experiment(
N,
N,
N,
nt,
is_dp=1,
ts=ts,
tb=tb,
kernel=kernel,
outfile=outfile,
target_dir=target_dir,
tgs=tgs,
nwf=4 * tgs,
mwdt=0,
)
def main():
from utils import create_project_tarball
import socket
#hostname = socket.gethostname()
exp_name = "emmy_"
exp_name = exp_name + "energy_vs_code_balance" #_at_%s" % (hostname)
tarball_dir = 'results/' + exp_name
create_project_tarball(tarball_dir, "project_" + exp_name)
target_dir = 'results/' + exp_name
cs = 4096
dw_l = [1, 3, 5, 7]
ts_test(target_dir,
exp_name,
cs,
ts=2,
kernel=1,
tgs=5,
N=960,
group='ENERGY',
dw_l=dw_l)
ts_test(target_dir,
exp_name,
cs,
ts=2,
kernel=1,
tgs=5,
N=960,
group='MEM',
dw_l=dw_l)
ts_test(target_dir,
exp_name,
cs,
ts=2,
kernel=5,
tgs=5,
N=480,
group='ENERGY',
dw_l=dw_l)
ts_test(target_dir,
exp_name,
cs,
ts=2,
kernel=5,
tgs=5,
N=480,
group='MEM',
dw_l=dw_l)
def main():
from utils import create_project_tarball
import socket
import time, datetime
time_stamp = datetime.datetime.fromtimestamp(
time.time()).strftime('%Y%m%d_%H_%M')
exp_name = "distributed_strongscaling_at_IVB10_%s" % (time_stamp)
tarball_dir = 'results/' + exp_name
create_project_tarball(tarball_dir, "project_" + exp_name)
target_dir = 'results/' + exp_name
exp = [[0, 5, 768, 0, 8192, proc] for proc in [1, 2, 4, 8, 12, 16, 24, 32]]
exp = exp + [[2, 5, 768, 1, 2048, proc] for proc in [1, 2, 4, 8, 12, 16]]
exp = exp + [[2, 5, 768, -1, 2048, proc]
for proc in [1, 2, 4, 8, 12, 16, 24, 32]]
# exp = exp + [[2, 5, 768, 2, 2048, proc] for proc in [1,2,4,8,12,16,24,32]]
# exp = exp + [[2, 5, 768, 5, 2048, proc] for proc in [1,2,4,8,12,16,24,32]]
# exp = exp + [[2, 5, 768, 10,2048, proc] for proc in [1,2,4,8,12,16,24,32]]
exp = exp + [[0, 4, 512, 0, 8192, proc] for proc in [1, 2, 4, 8, 16]]
exp = exp + [[2, 4, 512, 1, 0, proc] for proc in [1, 2]]
exp = exp + [[2, 4, 512, -1, 0, proc] for proc in [1, 2, 4, 8, 16]]
# exp = exp + [[2, 4, 512, 2, 0, proc] for proc in [1,2,4]]
# exp = exp + [[2, 4, 512, 5, 0, proc] for proc in [1,2,4,8,16]]
# exp = exp + [[2, 4, 512, 10, 0, proc] for proc in [1,2,4,8,16,32]]
exp_f = ['ts', 'kernel', 'N', 'tgs', 'cs', 'procs']
exp = [dict(zip(exp_f, e)) for e in exp]
# nodes_range = [0, 2]
# nodes_range = [2, 8]
nodes_range = [8, 16]
nt = 50
for e in exp:
if (e['procs'] > nodes_range[0] * 2) and (e['procs'] <=
nodes_range[1] * 2):
experiment(target_dir,
exp_name,
ts=e['ts'],
tgs=e['tgs'],
kernel=e['kernel'],
nt=nt,
N=e['N'],
proc_l=[e['procs']],
cs=e['cs'])
def main():
from utils import create_project_tarball
import socket
hostname = socket.gethostname()
exp_name = "emmy_distributed_strongscaling_f2.2GHz_mwd_swd_at_%s" % hostname
tarball_dir = 'results/' + exp_name
create_project_tarball(tarball_dir, "project_" + exp_name)
target_dir = 'results/' + exp_name
exp = [[0, 5, 768, 0, 8192, proc] for proc in [1, 2, 4, 8, 12, 16, 24, 32]]
exp = exp + [[2, 5, 768, 1, 2048, proc] for proc in [1, 2, 4, 8, 12, 16]]
exp = exp + [[2, 5, 768, 2, 2048, proc]
for proc in [1, 2, 4, 8, 12, 16, 24, 32]]
exp = exp + [[2, 5, 768, 5, 2048, proc]
for proc in [1, 2, 4, 8, 12, 16, 24, 32]]
exp = exp + [[2, 5, 768, 10, 2048, proc]
for proc in [1, 2, 4, 8, 12, 16, 24, 32]]
# exp = exp + [[0, 4, 512, 0, 8192, proc] for proc in [1,2,4,8,16,32]]
# exp = exp + [[2, 4, 512, 1, 0, proc] for proc in [1,2]]
# exp = exp + [[2, 4, 512, 2, 0, proc] for proc in [1,2,4]]
# exp = exp + [[2, 4, 512, 5, 0, proc] for proc in [1,2,4,8,16]]
# exp = exp + [[2, 4, 512, 10, 0, proc] for proc in [1,2,4,8,16,32]]
exp_f = ['ts', 'kernel', 'N', 'tgs', 'cs', 'procs']
exp = [dict(zip(exp_f, e)) for e in exp]
# nodes_range = [0, 2]
# nodes_range = [2, 8]
nodes_range = [8, 16]
# nodes_range = [0, 8]
nt = 50
for e in exp:
if (e['procs'] > nodes_range[0] * 2) and (e['procs'] <=
nodes_range[1] * 2):
experiment(target_dir,
exp_name,
ts=e['ts'],
tgs=e['tgs'],
kernel=e['kernel'],
nt=nt,
N=e['N'],
proc_l=[e['procs']],
cs=e['cs'])
def main():
from utils import create_project_tarball
import socket
hostname = socket.gethostname()
exp_name = "emmy_distributed_strongscaling_f2.2GHz_mwd_swd_at_%s" % hostname
tarball_dir = "results/" + exp_name
create_project_tarball(tarball_dir, "project_" + exp_name)
target_dir = "results/" + exp_name
exp = [[0, 5, 768, 0, 8192, proc] for proc in [1, 2, 4, 8, 12, 16, 24, 32]]
exp = exp + [[2, 5, 768, 1, 2048, proc] for proc in [1, 2, 4, 8, 12, 16]]
exp = exp + [[2, 5, 768, 2, 2048, proc] for proc in [1, 2, 4, 8, 12, 16, 24, 32]]
exp = exp + [[2, 5, 768, 5, 2048, proc] for proc in [1, 2, 4, 8, 12, 16, 24, 32]]
exp = exp + [[2, 5, 768, 10, 2048, proc] for proc in [1, 2, 4, 8, 12, 16, 24, 32]]
# exp = exp + [[0, 4, 512, 0, 8192, proc] for proc in [1,2,4,8,16,32]]
# exp = exp + [[2, 4, 512, 1, 0, proc] for proc in [1,2]]
# exp = exp + [[2, 4, 512, 2, 0, proc] for proc in [1,2,4]]
# exp = exp + [[2, 4, 512, 5, 0, proc] for proc in [1,2,4,8,16]]
# exp = exp + [[2, 4, 512, 10, 0, proc] for proc in [1,2,4,8,16,32]]
exp_f = ["ts", "kernel", "N", "tgs", "cs", "procs"]
exp = [dict(zip(exp_f, e)) for e in exp]
# nodes_range = [0, 2]
# nodes_range = [2, 8]
nodes_range = [8, 16]
# nodes_range = [0, 8]
nt = 50
for e in exp:
if (e["procs"] > nodes_range[0] * 2) and (e["procs"] <= nodes_range[1] * 2):
experiment(
target_dir,
exp_name,
ts=e["ts"],
tgs=e["tgs"],
kernel=e["kernel"],
nt=nt,
N=e["N"],
proc_l=[e["procs"]],
cs=e["cs"],
)
def main():
from utils import create_project_tarball
import socket
hostname = socket.gethostname()
exp_name = "emmy_socket_thread_scaling_7pt_var_nosym_680cube_f2.2GHz_mwd_swd_at_%s" % hostname
tarball_dir='results/'+exp_name
create_project_tarball(tarball_dir, "project_"+exp_name)
target_dir='results/' + exp_name
kl = [5]
wf = 1
for th in list(range(10, 0, -2)):
for is_dp in [1]:
for ts, tgs, tb, nwf in [(2, 2, 3, 2)]: # DP
for kernel in kl:
outfile=(exp_name + '_ts%d_tb%d_tgs%d_nwf%d_kernel%d_isDP%d_threads%d.txt' % (ts, tb, tgs, nwf, kernel, is_dp, th))
submit_emmy_experiment(nwf=nwf, tgs=tgs, wf=wf, is_dp=is_dp, th=th, ts=ts, tb=tb, kernel=kernel, outfile=outfile, target_dir=target_dir)
for th in[5, 10]:
for is_dp in [1]:
for ts, tgs, tb, nwf in [(2, 5, 3, 15)]: #DP
for kernel in kl:
outfile=(exp_name + '_ts%d_tb%d_tgs%d_nwf%d_kernel%d_isDP%d_threads%d.txt' % (ts, tb, tgs, nwf, kernel, is_dp, th))
submit_emmy_experiment(nwf=nwf, tgs=tgs, wf=wf, is_dp=is_dp, th=th, ts=ts, tb=tb, kernel=kernel, outfile=outfile, target_dir=target_dir)
for th in[10]:
for is_dp in [1]:
for ts, tgs, tb, nwf in [(2, 10, 9, 10)]: #DP
for kernel in kl:
outfile=(exp_name + '_ts%d_tb%d_tgs%d_nwf%d_kernel%d_isDP%d_threads%d.txt' % (ts, tb, tgs, nwf, kernel, is_dp, th))
submit_emmy_experiment(nwf=nwf, tgs=tgs, wf=wf, is_dp=is_dp, th=th, ts=ts, tb=tb, kernel=kernel, outfile=outfile, target_dir=target_dir)
for th in list(range(10, 0, -1)):
for is_dp in [1]:
for ts, tgs, tb, nwf in [(0,0,0,0), (2, 1, 3, 1)]:
for kernel in kl:
outfile=(exp_name + '_ts%d_tb%d_tgs%d_nwf%d_kernel%d_isDP%d_threads%d.txt' % (ts, tb, tgs, nwf, kernel, is_dp, th))
submit_emmy_experiment(nwf=nwf, tgs=tgs, wf=wf, is_dp=is_dp, th=th, ts=ts, tb=tb, kernel=kernel, outfile=outfile, target_dir=target_dir)
def main():
from utils import create_project_tarball
import socket
import time,datetime
time_stamp = datetime.datetime.fromtimestamp(time.time()).strftime('%Y%m%d_%H_%M')
exp_name = "distributed_strongscaling_at_IVB10_%s" % (time_stamp)
tarball_dir='results/'+exp_name
create_project_tarball(tarball_dir, "project_"+exp_name)
target_dir='results/' + exp_name
exp = [[0, 5, 768, 0, 8192, proc] for proc in [1,2,4,8,12,16,24,32]]
exp = exp + [[2, 5, 768, 1, 2048, proc] for proc in [1,2,4,8,12,16]]
exp = exp + [[2, 5, 768,-1, 2048, proc] for proc in [1,2,4,8,12,16,24,32]]
# exp = exp + [[2, 5, 768, 2, 2048, proc] for proc in [1,2,4,8,12,16,24,32]]
# exp = exp + [[2, 5, 768, 5, 2048, proc] for proc in [1,2,4,8,12,16,24,32]]
# exp = exp + [[2, 5, 768, 10,2048, proc] for proc in [1,2,4,8,12,16,24,32]]
exp = exp + [[0, 4, 512, 0, 8192, proc] for proc in [1,2,4,8,16]]
exp = exp + [[2, 4, 512, 1, 0, proc] for proc in [1,2]]
exp = exp + [[2, 4, 512,-1, 0, proc] for proc in [1,2,4,8,16]]
# exp = exp + [[2, 4, 512, 2, 0, proc] for proc in [1,2,4]]
# exp = exp + [[2, 4, 512, 5, 0, proc] for proc in [1,2,4,8,16]]
# exp = exp + [[2, 4, 512, 10, 0, proc] for proc in [1,2,4,8,16,32]]
exp_f = ['ts', 'kernel', 'N', 'tgs', 'cs', 'procs']
exp = [dict(zip(exp_f,e)) for e in exp]
# nodes_range = [0, 2]
# nodes_range = [2, 8]
nodes_range = [8, 16]
nt = 50
for e in exp:
if (e['procs'] > nodes_range[0]*2) and (e['procs'] <= nodes_range[1]*2):
experiment(target_dir, exp_name, ts=e['ts'], tgs=e['tgs'], kernel=e['kernel'], nt=nt, N=e['N'], proc_l=[e['procs']], cs=e['cs'])
def main():
from utils import create_project_tarball
import socket
hostname = socket.gethostname()
exp_name = "emmy_"
run_cmd = "export OMP_NUM_THREADS=10; likwid-perfctr -m -C S0:0-9 -s 0x03 -g MEM "
cs = 8192
exp_name = exp_name + "test_swd_at_mwd_code_at_%s" % (hostname)
tarball_dir = 'results/' + exp_name
create_project_tarball(tarball_dir, "project_" + exp_name)
target_dir = 'results/' + exp_name
mwd_cs = 4096
mwd(target_dir, exp_name[:-9], mwd_cs, run_cmd, 1)
def main():
from utils import create_project_tarball
import socket
# hostname = socket.gethostname()
exp_name = "emmy_"
exp_name = exp_name + "threadscaling_energy"#_at_%s" % (hostname)
tarball_dir='results/'+exp_name
create_project_tarball(tarball_dir, "project_"+exp_name)
target_dir='results/' + exp_name
th_l = [1,2,3,4,5,6,7,8,9,10]
cs = 4096
for tgs in [1,2,5,10]:
# ts_test(target_dir, exp_name, cs, ts=2, kernel=1, tgs=tgs, N=960, th_l=th_l)
# ts_test(target_dir, exp_name, cs, ts=2, kernel=4, tgs=tgs, N=480, th_l=th_l)
ts_test(target_dir, exp_name, cs, ts=2, kernel=5, tgs=tgs, N=680, th_l=th_l)
def main(): from utils import create_project_tarball import socket hostname = socket.gethostname() exp_name = "emmy_" run_cmd = "export OMP_NUM_THREADS=10; likwid-perfctr -m -C S0:0-9 -s 0x03 -g MEM " cs = 8192 exp_name = exp_name + "test_swd_at_mwd_code_at_%s" % (hostname) tarball_dir='results/'+exp_name create_project_tarball(tarball_dir, "project_"+exp_name) target_dir='results/' + exp_name mwd_cs = 4096 mwd(target_dir, exp_name[:-9], mwd_cs, run_cmd, 1)
def main():
from utils import create_project_tarball
import socket
hostname = socket.gethostname()
exp_name = "emmy_socket_thread_scaling_naive_inL3_1to10_threads_f2p2GHz_at_%s" % hostname
tarball_dir='results/'+exp_name
create_project_tarball(tarball_dir, "project_"+exp_name)
target_dir='results/' + exp_name
size_l = {0:(128, 64, 64, 1e4), 1:(96, 96, 96, 4e4), 4:(64, 32, 32, 5e4), 5:(64, 48,48, 1e5)}
for th in list(range(10, 0, -1)):
for is_dp in [1]:
for kernel in [0, 1, 4, 5]:
nx, ny, nz, nt = size_l[kernel]
outfile=(exp_name + 'naive_kernel%d_isDP%d_threads%d' % (kernel, is_dp, th))
submit_emmy_experiment(nt=nt, nx=nx, ny=ny, nz=nz, is_dp=is_dp, th=th, kernel=kernel, outfile=outfile, target_dir=target_dir)
def main():
from utils import create_project_tarball
import socket
hostname = socket.gethostname()
exp_name = "emmy_energy_f2.2GHz_at_%s" % hostname
tarball_dir = 'results/' + exp_name
create_project_tarball(tarball_dir, "project_" + exp_name)
target_dir = 'results/' + exp_name
wf = 1
is_dp = 1
th = 10
# (stencil, N, nt, experiments)
exp = [(1, 960, 100, [(0, 0, 0, 0)])]
th = 6 # special case for spatial blocking
#exp = [(1, 960, 100, [(2, 2, 5, 12), (2, 5, 11, 20), (2, 10, 11, 50), (0,0,0,0), (2, 1, 5, 1)])]
#exp = exp + [(5, 680, 50, [(2, 2, 3, 2), (2, 5, 3, 15), (2, 10, 9, 10), (0,0,0,0), (2, 1, 3, 1)])]
#exp = exp + [(4, 480, 100, [(2, 2, 1, 2), (2, 5, 1, 5), (2, 10, 3, 10), (0,0,0,0), (2, 1, 1, 9)])]
for kernel, N, nt, exp_l in exp:
for ts, tgs, tb, nwf in exp_l:
outfile = (exp_name +
'_ts%d_tgs%d_tb%d_nwf%d_kernel%d_isDP%d_threads%d.txt' %
(ts, tgs, tb, nwf, kernel, is_dp, th))
submit_emmy_experiment(N=N,
nt=nt,
nwf=nwf,
tgs=tgs,
wf=wf,
is_dp=is_dp,
th=th,
ts=ts,
tb=tb,
kernel=kernel,
outfile=outfile,
target_dir=target_dir)
def main(): from utils import create_project_tarball import socket hostname = socket.gethostname() exp_name = "emmy_" run_cmd = "export OMP_NUM_THREADS=10; likwid-perfctr -m -C S0:0-9 -s 0x03 -g MEM " cs = 8192 exp_name = exp_name + "strongscaling_mwd_naive_at_%s" % (hostname) tarball_dir='results/'+exp_name create_project_tarball(tarball_dir, "project_"+exp_name) target_dir='results/' + exp_name #mwd(target_dir, exp_name, 0, run_cmd, 1) #mwd(target_dir, exp_name, 0, run_cmd, 2) mwd(target_dir, exp_name, 0, run_cmd, 5)
def main():
from utils import create_project_tarball
import socket
hostname = socket.gethostname()
exp_name = "emmy_"
run_cmd = "export OMP_NUM_THREADS=10; likwid-perfctr -m -C S0:0-9 -s 0x03 -g MEM "
cs = 8192
exp_name = exp_name + "test_no_separate_clu_at_%s" % (hostname)
tarball_dir = 'results/' + exp_name
create_project_tarball(tarball_dir, "project_" + exp_name)
target_dir = 'results/' + exp_name
mwd(target_dir, exp_name[:-9], 2048, run_cmd, 1, [1, 4, 5])
mwd(target_dir, exp_name[:-9], 2048, run_cmd, 2, [1, 5])
mwd(target_dir, exp_name[:-9], 2048, run_cmd, 5, [4])
naive(target_dir, exp_name[:-9], cs, run_cmd)
def main():
from utils import create_project_tarball
import socket
hostname = socket.gethostname()
exp_name = "emmy_socket_thread_scaling_25pt_const_960cube_f2.2GHz_naive_at_%s" % hostname
tarball_dir='results/'+exp_name
create_project_tarball(tarball_dir, "project_"+exp_name)
target_dir='results/' + exp_name
kl = [0]
wf = 1
for th in list(range(10, 0, -1)):
for is_dp in [0, 1]:
for ts, tgs, tb, nwf in [(0,0,0,0)]:
for kernel in kl:
outfile=(exp_name + '_ts%d_kernel%d_isDP%d_threads%d.txt' % (ts, kernel, is_dp, th))
submit_emmy_experiment(nwf=nwf, tgs=tgs, wf=wf, is_dp=is_dp, th=th, ts=ts, tb=tb, kernel=kernel, outfile=outfile, target_dir=target_dir)
def main(): from utils import create_project_tarball import socket hostname = socket.gethostname() exp_name = "emmy_" run_cmd = "export OMP_NUM_THREADS=10; likwid-perfctr -m -C S0:0-9 -s 0x03 -g MEM " cs = 8192 exp_name = exp_name + "test_no_separate_clu_at_%s" % (hostname) tarball_dir='results/'+exp_name create_project_tarball(tarball_dir, "project_"+exp_name) target_dir='results/' + exp_name mwd(target_dir, exp_name[:-9], 2048, run_cmd, 1, [1, 4, 5]) mwd(target_dir, exp_name[:-9], 2048, run_cmd, 2, [1, 5]) mwd(target_dir, exp_name[:-9], 2048, run_cmd, 5, [4]) naive(target_dir, exp_name[:-9], cs, run_cmd)
def main():
from utils import create_project_tarball
from csv import DictReader
from utils import ensure_dir
import os
exp_name = "mic_mwd_thread_scaling"
tarball_dir='results/'+exp_name
create_project_tarball(tarball_dir, "project_"+exp_name)
target_dir='results/' + exp_name
target_dir = os.path.join(os.path.abspath("."),target_dir)
ensure_dir(target_dir)
# use auto tuning to find the best params at one group
cs = 64
ts = 2
kernel=1; N=768
# # cache blocks within L2 caches
# cmds = []
# mwdt=2
# for tgs in [1,2,4]:
# for tpc in [1,2,4]:
# if tgs <= tpc:
# th_l = [tpc*t for t in range(6,61,6)]
# for th in th_l:
# aff="E:N:%d:%d:%d" % (th, tpc, 4)
# cmds = cmds + mwd(target_dir, exp_name, cs, N=N, ts=ts, kernel=kernel, tgs=tgs, th=th, mwd_type=mwdt, groups=['CPI'], aff=aff, tpc=tpc)
#
# cmds = [c for sublist in cmds for c in sublist]
# fname = target_dir + "/ts%d_kernel%d_mwdt%d_CPI_autotune.sh"%(ts, kernel,mwdt)
# with open(fname, 'w') as fn:
# fn.write("#!/bin/sh\n\n")
# for c in cmds:
# cp = 'echo ' + str.replace(c,';',"';'")
# cp = str.replace(cp, '|', '#|')
# fn.write('echo "'+c+'"\n')
# fn.write(c+'\n\n')
# parse the results to obtain the selected parameters by the auto tuner
data = []
data_file = os.path.join(target_dir, 'summary.csv')
with open(data_file, 'rb') as output_file:
raw_data = DictReader(output_file)
for k in raw_data:
k['stencil'] = get_knum(k)
k['method'] = 2 if 'Diamond' in k['Time stepper orig name'] else 0
if k['method'] == 2:
if k['Wavefront parallel strategy'] == 'Relaxed synchronization wavefront with fixed execution':
k['mwdt'] = 3
elif k['Wavefront parallel strategy'] == 'Relaxed synchronization wavefront':
k['mwdt'] = 2
if int(k['Thread group size']) == 1:
k['mwdt'] = 0
data.append(k)
params = dict()
for k in data:
try:
if k['method']!=0:
params[(k['mwdt'], k['method'], k['stencil'], int(k['Thread group size']), int(k['Global NX']), int(k['Threads per core']), int(k['OpenMP Threads']) )]= (int(k['Time unroll']), int(k['Multi-wavefront updates']), int(k['Block size in X']))
except:
print k
raise
# for k in params.keys(): print k
# cache blocks within L2 caches
cmds = []
mwdt_r=2
for tgs in [1,2,4]:
for tpc in [1,2,4]:
if tgs <= tpc:
th_l = [tpc*t for t in range(6,61,6)]
for th in th_l:
if ((th/tgs)<=(N/4)):
mwdt=mwdt_r if tgs!=1 else 0
tb, nwf, bsx = params[(mwdt, ts, kernel, tgs, N, tpc, th)]
aff="E:N:%d:%d:%d" % (th, tpc, 4)
cmds = cmds + mwd(target_dir, exp_name, cs, N=N, ts=ts, kernel=kernel, tgs=tgs, th=th, mwd_type=mwdt, groups=['MEM1', 'MEM2', 'MEM3', 'MEM4', 'MEM5'], aff=aff, tpc=tpc, tb=tb, nwf=nwf, bsx=bsx)
cmds = [c for sublist in cmds for c in sublist]
fname = target_dir + "/ts%d_kernel%d_mwdt%d_other_groups.sh"%(ts, kernel,mwdt)
with open(fname, 'w') as fn:
fn.write("#!/bin/sh\n\n")
for c in cmds:
cp = 'echo ' + str.replace(c,';',"';'")
cp = str.replace(cp, '|', '#|')
fn.write('echo "'+c+'"\n')
fn.write(c+'\n\n')
def main():
from utils import create_project_tarball
import socket
from csv import DictReader
import os
hostname = socket.gethostname()
exp_name = "hasep1_"
exp_name = exp_name + "prof_mwd_cache_"
tarball_dir = "results/" + exp_name
create_project_tarball(tarball_dir, "project_" + exp_name)
target_dir = "results/" + exp_name
grid_set = {1: [480, 960], 5: [256, 640]}
# # use auto tuning to find the best params
# th=14
# cs = 6000
# ts = 2
# for kernel in [1, 5]:
# for N in grid_set[kernel]:
# for tgs, mwt_l in [(1,[2]), (2, [2,3]), (7, [2,3]), (14, [2,3])]:
# for mwdt in [2,3]:
# mwd(target_dir, exp_name, cs, N=N, ts=ts, kernel=kernel, tgs=tgs, th=th, mwd_type=mwdt, groups=['DATA'])
#
# parse the results to obtain the selected parameters by the auto tuner
data = []
data_file = os.path.join(target_dir, "summary.csv")
with open(data_file, "rb") as output_file:
raw_data = DictReader(output_file)
for k in raw_data:
k["stencil"] = get_knum(k)
k["method"] = 2 if "Diamond" in k["Time stepper orig name"] else 0
if k["method"] == 2:
if k["Wavefront parallel strategy"] == "Relaxed synchronization wavefront with fixed execution":
k["mwdt"] = 3
elif k["Wavefront parallel strategy"] == "Relaxed synchronization wavefront":
k["mwdt"] = 2
if int(k["Thread group size"]) == 1:
k["mwdt"] = 0
data.append(k)
params = dict()
for k in data:
try:
if k["method"] != 0:
params[(k["mwdt"], k["method"], k["stencil"], int(k["Thread group size"]), int(k["Global NX"]))] = (
int(k["Time unroll"]),
int(k["Multi-wavefront updates"]),
int(k["Block size in X"]),
)
except:
print k
raise
cs = 16000
th_l = [1, 2, 4, 6, 8, 10, 12, 14]
for kernel in [1, 5]:
for N in grid_set[kernel]:
for tgs, ths, mwdt_l in [
(0, th_l, [0]),
(1, th_l, [0]),
(2, th_l[1:], [2, 3]),
(7, [7, 14], [2, 3]),
(14, [14], [2, 3]),
]:
for mwdt in mwdt_l:
for th in ths:
if tgs == 0: # spatial blocking
ts = 0
tb, nwf, bsx = (-1, -1, 100000)
else:
ts = 2
tb, nwf, bsx = params[(mwdt, ts, kernel, tgs, N)]
# mwd(target_dir, exp_name, cs, N=N, ts=ts, kernel=kernel, tgs=tgs, th=th, mwd_type=mwdt, groups=['DATA', 'L2', 'L3', 'MEM'], tb=tb, nwf=nwf, bsx=bsx)
mwd(
target_dir,
exp_name,
cs,
N=N,
ts=ts,
kernel=kernel,
tgs=tgs,
th=th,
mwd_type=mwdt,
groups=["DATA", "L2", "L3"],
tb=tb,
nwf=nwf,
bsx=bsx,
)
def main():
from utils import create_project_tarball
from csv import DictReader
from utils import ensure_dir
import os
exp_name = "mic_mwd_thread_scaling"
tarball_dir = 'results/' + exp_name
create_project_tarball(tarball_dir, "project_" + exp_name)
target_dir = 'results/' + exp_name
target_dir = os.path.join(os.path.abspath("."), target_dir)
ensure_dir(target_dir)
# use auto tuning to find the best params at one group
cs = 64
ts = 2
kernel = 1
N = 768
# # cache blocks within L2 caches
# cmds = []
# mwdt=2
# for tgs in [1,2,4]:
# for tpc in [1,2,4]:
# if tgs <= tpc:
# th_l = [tpc*t for t in range(6,61,6)]
# for th in th_l:
# aff="E:N:%d:%d:%d" % (th, tpc, 4)
# cmds = cmds + mwd(target_dir, exp_name, cs, N=N, ts=ts, kernel=kernel, tgs=tgs, th=th, mwd_type=mwdt, groups=['CPI'], aff=aff, tpc=tpc)
#
# cmds = [c for sublist in cmds for c in sublist]
# fname = target_dir + "/ts%d_kernel%d_mwdt%d_CPI_autotune.sh"%(ts, kernel,mwdt)
# with open(fname, 'w') as fn:
# fn.write("#!/bin/sh\n\n")
# for c in cmds:
# cp = 'echo ' + str.replace(c,';',"';'")
# cp = str.replace(cp, '|', '#|')
# fn.write('echo "'+c+'"\n')
# fn.write(c+'\n\n')
# parse the results to obtain the selected parameters by the auto tuner
data = []
data_file = os.path.join(target_dir, 'summary.csv')
with open(data_file, 'rb') as output_file:
raw_data = DictReader(output_file)
for k in raw_data:
k['stencil'] = get_knum(k)
k['method'] = 2 if 'Diamond' in k['Time stepper orig name'] else 0
if k['method'] == 2:
if k['Wavefront parallel strategy'] == 'Relaxed synchronization wavefront with fixed execution':
k['mwdt'] = 3
elif k['Wavefront parallel strategy'] == 'Relaxed synchronization wavefront':
k['mwdt'] = 2
if int(k['Thread group size']) == 1:
k['mwdt'] = 0
data.append(k)
params = dict()
for k in data:
try:
if k['method'] != 0:
params[(k['mwdt'], k['method'], k['stencil'],
int(k['Thread group size']), int(k['Global NX']),
int(k['Threads per core']),
int(k['OpenMP Threads']))] = (
int(k['Time unroll']),
int(k['Multi-wavefront updates']),
int(k['Block size in X']))
except:
print k
raise
# for k in params.keys(): print k
# cache blocks within L2 caches
cmds = []
mwdt_r = 2
for tgs in [1, 2, 4]:
for tpc in [1, 2, 4]:
if tgs <= tpc:
th_l = [tpc * t for t in range(6, 61, 6)]
for th in th_l:
if ((th / tgs) <= (N / 4)):
mwdt = mwdt_r if tgs != 1 else 0
tb, nwf, bsx = params[(mwdt, ts, kernel, tgs, N, tpc,
th)]
aff = "E:N:%d:%d:%d" % (th, tpc, 4)
cmds = cmds + mwd(
target_dir,
exp_name,
cs,
N=N,
ts=ts,
kernel=kernel,
tgs=tgs,
th=th,
mwd_type=mwdt,
groups=['MEM1', 'MEM2', 'MEM3', 'MEM4', 'MEM5'],
aff=aff,
tpc=tpc,
tb=tb,
nwf=nwf,
bsx=bsx)
cmds = [c for sublist in cmds for c in sublist]
fname = target_dir + "/ts%d_kernel%d_mwdt%d_other_groups.sh" % (ts, kernel,
mwdt)
with open(fname, 'w') as fn:
fn.write("#!/bin/sh\n\n")
for c in cmds:
cp = 'echo ' + str.replace(c, ';', "';'")
cp = str.replace(cp, '|', '#|')
fn.write('echo "' + c + '"\n')
fn.write(c + '\n\n')
def main():
from utils import create_project_tarball
import socket
from csv import DictReader
import os
hostname = socket.gethostname()
exp_name = "emmy_"
exp_name = exp_name + "threadscaling_mem"
tarball_dir='results/'+exp_name
create_project_tarball(tarball_dir, "project_"+exp_name)
target_dir='results/' + exp_name
# parse the results to obtain the selected parameters by the auto tuner
data = []
data_file = os.path.join(target_dir, 'summary.csv')
with open(data_file, 'rb') as output_file:
raw_data = DictReader(output_file)
for k in raw_data:
k['stencil'] = get_knum(k)
k['method'] = 2 if 'Diamond' in k['Time stepper orig name'] else 0
if k['method'] == 2:
if k['Wavefront parallel strategy'] == 'Relaxed synchronization wavefront with fixed execution':
k['mwdt'] = 3
elif k['Wavefront parallel strategy'] == 'Relaxed synchronization wavefront':
k['mwdt'] = 2
elif k['Wavefront parallel strategy'] == 'multi-wavefronts':
k['mwdt'] = 0
if int(k['Thread group size']) == 1:
k['mwdt'] = 0
data.append(k)
params = dict()
for k in data:
try:
if k['method']!=0:
params[(k['mwdt'], k['method'], k['stencil'], int(k['Thread group size']), int(k['Global NX']))]= (int(k['Time unroll']), int(k['Multi-wavefront updates']), int(k['Block size in X']))
except:
print k
raise
cs = 8192
th_l = [1,2,3,4,5,6,7,8,9,10]
exp = [ # ts kernel N mwdt tgs th
# (0, 1, 960, 2, 0, 6 ),
# (2, 1, 960, 0, 1, 10),
# (2, 1, 960, 2, 2, 10),
# (2, 1, 960, 2, 5, 10),
# (2, 1, 960, 2, 10,10),
#
# (0, 5, 680, 2, 0, 6 ),
# (2, 5, 680, 0, 1, 8 ),
# (2, 5, 680, 2, 2, 10),
# (2, 5, 680, 2, 5, 10),
# (2, 5, 680, 2, 10,10),
#
# (0, 4, 480, 0, 0, 8 ),
(2, 4, 480, 0, 1, 7)
# (2, 4, 480, 0, 2, 8 ),
# (2, 4, 480, 0, 5, 10),
# (2, 4, 480, 0, 10,10)
]
for ts, kernel, N, mwdt, tgs, th in exp:
if tgs==0: # spatial blocking
tb, nwf, bsx = (-1,-1,100000)
else:
tb, nwf, bsx = params[(mwdt, ts, kernel, tgs, N)]
mwd(target_dir, exp_name, cs, N=N, ts=ts, kernel=kernel, tgs=tgs, th=th, mwd_type=mwdt, tb=tb, nwf=nwf, bsx=bsx)
def main():
from utils import create_project_tarball
import socket
from csv import DictReader
import os
hostname = socket.gethostname()
exp_name = "emmy_"
exp_name = exp_name + "threadscaling_mem"
tarball_dir = 'results/' + exp_name
create_project_tarball(tarball_dir, "project_" + exp_name)
target_dir = 'results/' + exp_name
# parse the results to obtain the selected parameters by the auto tuner
data = []
data_file = os.path.join(target_dir, 'summary.csv')
with open(data_file, 'rb') as output_file:
raw_data = DictReader(output_file)
for k in raw_data:
k['stencil'] = get_knum(k)
k['method'] = 2 if 'Diamond' in k['Time stepper orig name'] else 0
if k['method'] == 2:
if k['Wavefront parallel strategy'] == 'Relaxed synchronization wavefront with fixed execution':
k['mwdt'] = 3
elif k['Wavefront parallel strategy'] == 'Relaxed synchronization wavefront':
k['mwdt'] = 2
elif k['Wavefront parallel strategy'] == 'multi-wavefronts':
k['mwdt'] = 0
if int(k['Thread group size']) == 1:
k['mwdt'] = 0
data.append(k)
params = dict()
for k in data:
try:
if k['method'] != 0:
params[(k['mwdt'], k['method'], k['stencil'],
int(k['Thread group size']),
int(k['Global NX']))] = (int(
k['Time unroll']), int(
k['Multi-wavefront updates']),
int(k['Block size in X']))
except:
print k
raise
cs = 8192
th_l = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
exp = [ # ts kernel N mwdt tgs th
# (0, 1, 960, 2, 0, 6 ),
# (2, 1, 960, 0, 1, 10),
# (2, 1, 960, 2, 2, 10),
# (2, 1, 960, 2, 5, 10),
# (2, 1, 960, 2, 10,10),
#
# (0, 5, 680, 2, 0, 6 ),
# (2, 5, 680, 0, 1, 8 ),
# (2, 5, 680, 2, 2, 10),
# (2, 5, 680, 2, 5, 10),
# (2, 5, 680, 2, 10,10),
#
# (0, 4, 480, 0, 0, 8 ),
(2, 4, 480, 0, 1, 7)
# (2, 4, 480, 0, 2, 8 ),
# (2, 4, 480, 0, 5, 10),
# (2, 4, 480, 0, 10,10)
]
for ts, kernel, N, mwdt, tgs, th in exp:
if tgs == 0: # spatial blocking
tb, nwf, bsx = (-1, -1, 100000)
else:
tb, nwf, bsx = params[(mwdt, ts, kernel, tgs, N)]
mwd(target_dir,
exp_name,
cs,
N=N,
ts=ts,
kernel=kernel,
tgs=tgs,
th=th,
mwd_type=mwdt,
tb=tb,
nwf=nwf,
bsx=bsx)
def main():
from utils import create_project_tarball
import socket
from csv import DictReader
import os
hostname = socket.gethostname()
exp_name = "emmy_"
exp_name = exp_name + "prof_mwd_cache_"
tarball_dir='results/'+exp_name
create_project_tarball(tarball_dir, "project_"+exp_name)
target_dir='results/' + exp_name
grid_set = {1:[480, 960], 5:[256, 640]}
# use auto tuning to find the best params
th=10
cs = 4096
ts = 2
for kernel in [1, 5]:
for N in grid_set[kernel]:
for tgs, mwt_l in [(1,[2]), (2, [2,3]), (5, [2,3]), (10, [2,3])]:
for mwdt in mwt_l:
mwd(target_dir, exp_name, cs, N=N, ts=ts, kernel=kernel, tgs=tgs, th=th, mwd_type=mwdt, groups=['DATA'])
# parse the results to obtain the selected parameters by the auto tuner
data = []
data_file = os.path.join(target_dir, 'summary.csv')
with open(data_file, 'rb') as output_file:
raw_data = DictReader(output_file)
for k in raw_data:
k['stencil'] = get_knum(k)
k['method'] = 2 if 'Diamond' in k['Time stepper orig name'] else 0
if k['method'] == 2:
if k['Wavefront parallel strategy'] == 'Relaxed synchronization wavefront with fixed execution':
k['mwdt'] = 3
elif k['Wavefront parallel strategy'] == 'Relaxed synchronization wavefront':
k['mwdt'] = 2
if int(k['Thread group size']) == 1:
k['mwdt'] = 0
data.append(k)
params = dict()
for k in data:
try:
if k['method']!=0:
params[(k['mwdt'], k['method'], k['stencil'], int(k['Thread group size']), int(k['Global NX']))]= (int(k['Time unroll']), int(k['Multi-wavefront updates']), int(k['Block size in X']))
except:
print k
raise
cs = 16000
th_l = [1,2,3,4,5,6,7,8,9,10]
for tgs, ths, mwdt_l in [(0,th_l, [0]), (1, th_l, [0]), (2,[2,4,6,8,10], [2,3]), (5, [5,10], [2,3]), (10, [10], [2,3])]:
for mwdt in mwdt_l:
for th in ths:
if tgs==0: # spatial blocking
ts = 0
tb, nwf, bsx = (-1,-1,100000)
else:
ts = 2
tb, nwf, bsx = params[(mwdt, ts, kernel, tgs, N)]
mwd(target_dir, exp_name, cs, N=N, ts=ts, kernel=kernel, tgs=tgs, th=th, mwd_type=mwdt, groups=['DATA', 'L2', 'L3', 'MEM'], tb=tb, nwf=nwf, bsx=bsx)
def main():
from utils import create_project_tarball
import socket
hostname = socket.gethostname()
exp_name = "emmy_socket_thread_scaling_25pt_var_axsym_480cube_f2.2GHz_mwd_swd_at_%s" % hostname
tarball_dir = 'results/' + exp_name
create_project_tarball(tarball_dir, "project_" + exp_name)
target_dir = 'results/' + exp_name
kl = [4]
wf = 1
for th in list(range(10, 0, -2)):
for is_dp in [1]:
for ts, tgs, tb, nwf in [(2, 2, 1, 2)]: # DP
for kernel in kl:
outfile = (
exp_name +
'_ts%d_tb%d_tgs%d_nwf%d_kernel%d_isDP%d_threads%d.txt'
% (ts, tb, tgs, nwf, kernel, is_dp, th))
submit_emmy_experiment(nwf=nwf,
tgs=tgs,
wf=wf,
is_dp=is_dp,
th=th,
ts=ts,
tb=tb,
kernel=kernel,
outfile=outfile,
target_dir=target_dir)
for th in [5, 10]:
for is_dp in [1]:
for ts, tgs, tb, nwf in [(2, 5, 1, 5)]: #DP
for kernel in kl:
outfile = (
exp_name +
'_ts%d_tb%d_tgs%d_nwf%d_kernel%d_isDP%d_threads%d.txt'
% (ts, tb, tgs, nwf, kernel, is_dp, th))
submit_emmy_experiment(nwf=nwf,
tgs=tgs,
wf=wf,
is_dp=is_dp,
th=th,
ts=ts,
tb=tb,
kernel=kernel,
outfile=outfile,
target_dir=target_dir)
for th in [10]:
for is_dp in [1]:
for ts, tgs, tb, nwf in [(2, 10, 3, 10)]: #DP
for kernel in kl:
outfile = (
exp_name +
'_ts%d_tb%d_tgs%d_nwf%d_kernel%d_isDP%d_threads%d.txt'
% (ts, tb, tgs, nwf, kernel, is_dp, th))
submit_emmy_experiment(nwf=nwf,
tgs=tgs,
wf=wf,
is_dp=is_dp,
th=th,
ts=ts,
tb=tb,
kernel=kernel,
outfile=outfile,
target_dir=target_dir)
for th in list(range(10, 0, -1)):
for is_dp in [1]:
for ts, tgs, tb, nwf in [(0, 0, 0, 0), (2, 1, 1, 9)]:
for kernel in kl:
outfile = (
exp_name +
'_ts%d_tb%d_tgs%d_nwf%d_kernel%d_isDP%d_threads%d.txt'
% (ts, tb, tgs, nwf, kernel, is_dp, th))
submit_emmy_experiment(nwf=nwf,
tgs=tgs,
wf=wf,
is_dp=is_dp,
th=th,
ts=ts,
tb=tb,
kernel=kernel,
outfile=outfile,
target_dir=target_dir)
def main():
from utils import create_project_tarball
import socket
# hostname = socket.gethostname()
exp_name = "emmy_cache_size_vs_code_balace_2p2GHz" #_at_%s" % hostname
tarball_dir = 'results/' + exp_name
create_project_tarball(tarball_dir, "project_" + exp_name)
target_dir = 'results/' + exp_name
kernel = 1
exp = []
exp = exp + [[0, 0, 960, 30, 0]]
# exp = exp + [[0, 0, 480, 400, 0]]
# exp = exp + [[2, i, 480, 400, 10] for i in range(1, 30, 2)]
exp = exp + [[2, i, 960, 50, 10] for i in range(1, 30, 2)]
for ts, tb, N, nt, tgs in exp:
outfile = ('ts%d_kernel%d_tb%d_tgs%d_N%d.txt' %
(ts, kernel, tb, tgs, N))
submit_emmy_experiment(N,
N,
N,
nt,
is_dp=1,
ts=ts,
tb=tb,
kernel=kernel,
outfile=outfile,
target_dir=target_dir,
tgs=tgs,
nwf=tgs)
kernel = 5
exp = []
# exp = exp + [[0, 0, 384, 1000, 0]]
exp = exp + [[0, 0, 680, 30, 0]]
# exp = exp + [[2, i, 384, 1000, 10] for i in range(1, 20, 2)]
exp = exp + [[2, i, 680, 50, 10] for i in range(1, 20, 2)]
for ts, tb, N, nt, tgs in exp:
outfile = ('ts%d_kernel%d_tb%d_tgs%d_N%d.txt' %
(ts, kernel, tb, tgs, N))
submit_emmy_experiment(N,
N,
N,
nt,
is_dp=1,
ts=ts,
tb=tb,
kernel=kernel,
outfile=outfile,
target_dir=target_dir,
tgs=tgs,
nwf=tgs)
kernel = 4
exp = []
# exp = exp + [[0, 0, 240, 500, 0]]
exp = exp + [[0, 0, 480, 30, 0]]
# exp = exp + [[2, i, 240, 500, 10] for i in range(1, 8, 2)]
exp = exp + [[2, i, 480, 50, 10] for i in range(1, 6, 2)]
for ts, tb, N, nt, tgs in exp:
outfile = ('ts%d_kernel%d_tb%d_tgs%d_N%d.txt' %
(ts, kernel, tb, tgs, N))
submit_emmy_experiment(N,
N,
N,
nt,
is_dp=1,
ts=ts,
tb=tb,
kernel=kernel,
outfile=outfile,
target_dir=target_dir,
tgs=tgs,
nwf=4 * tgs,
mwdt=0)