def main():
env = Environment.get()
env.enable_function_timers()
log = Logger.get()
gt = GlobalTimers.get()
gt.start("toast_benchmark (total)")
mpiworld, procs, rank = get_world()
if rank == 0:
log.info("TOAST version = {}".format(env.version()))
log.info("Using a maximum of {} threads per process".format(env.max_threads()))
if mpiworld is None:
log.info("Running serially with one process at {}".format(str(datetime.now())))
else:
if rank == 0:
log.info(
"Running with {} processes at {}".format(procs, str(datetime.now()))
)
cases = {
"tiny": 5000000, # O(1) GB RAM
"xsmall": 50000000, # O(10) GB RAM
"small": 500000000, # O(100) GB RAM
"medium": 5000000000, # O(1) TB RAM
"large": 50000000000, # O(10) TB RAM
"xlarge": 500000000000, # O(100) TB RAM
"heroic": 5000000000000, # O(1000) TB RAM
}
args, comm, n_nodes, n_detector, case, group_seconds, n_group = job_config(
mpiworld, cases
)
# Note: The number of "days" here will just be an approximation of the desired
# data volume since we are doing a realistic schedule for a real observing site.
n_days = int(2.0 * (group_seconds * n_group) / (24 * 3600))
if n_days == 0:
n_days = 1
if rank == 0:
log.info(
"Using {} detectors for approximately {} days".format(n_detector, n_days)
)
# Create the schedule file and input maps on one process
if rank == 0:
create_schedules(args, group_seconds, n_days)
create_input_maps(args)
if mpiworld is not None:
mpiworld.barrier()
if args.dry_run is not None:
if rank == 0:
log.info("Exit from dry run")
# We are done!
sys.exit(0)
gt.start("toast_benchmark (science work)")
# Load and broadcast the schedule file
schedules = pipeline_tools.load_schedule(args, comm)
# Load the weather and append to schedules
pipeline_tools.load_weather(args, comm, schedules)
# Simulate the focalplane
detweights = create_focalplanes(args, comm, schedules, n_detector)
# Create the TOAST data object to match the schedule. This will
# include simulating the boresight pointing.
data, telescope_data, total_samples = create_observations(args, comm, schedules)
# handle = None
# if comm.world_rank == 0:
# handle = open(os.path.join(args.outdir, "distdata.txt"), "w")
# data.info(handle)
# if comm.world_rank == 0:
# handle.close()
# if comm.comm_world is not None:
# comm.comm_world.barrier()
# Split the communicator for day and season mapmaking
time_comms = pipeline_tools.get_time_communicators(args, comm, data)
# Expand boresight quaternions into detector pointing weights and
# pixel numbers
pipeline_tools.expand_pointing(args, comm, data)
# Optionally rewrite the noise PSD:s in each observation to include
# elevation-dependence
pipeline_tools.get_elevation_noise(args, comm, data)
# Purge the pointing if we are NOT going to export the
# data to a TIDAS volume
if (args.tidas is None) and (args.spt3g is None):
for ob in data.obs:
tod = ob["tod"]
tod.free_radec_quats()
# Prepare auxiliary information for distributed map objects
signalname = pipeline_tools.scan_sky_signal(args, comm, data, "signal")
# Set up objects to take copies of the TOD at appropriate times
totalname, totalname_freq = setup_sigcopy(args)
# Loop over Monte Carlos
firstmc = args.MC_start
nsimu = args.MC_count
freqs = [float(freq) for freq in args.freq.split(",")]
nfreq = len(freqs)
for mc in range(firstmc, firstmc + nsimu):
pipeline_tools.simulate_atmosphere(args, comm, data, mc, totalname)
# Loop over frequencies with identical focal planes and identical
# atmospheric noise.
for ifreq, freq in enumerate(freqs):
if comm.world_rank == 0:
log.info(
"Processing frequency {}GHz {} / {}, MC = {}".format(
freq, ifreq + 1, nfreq, mc
)
)
# Make a copy of the atmosphere so we can scramble the gains and apply
# frequency-dependent scaling.
pipeline_tools.copy_signal(args, comm, data, totalname, totalname_freq)
pipeline_tools.scale_atmosphere_by_frequency(
args, comm, data, freq=freq, mc=mc, cache_name=totalname_freq
)
pipeline_tools.update_atmospheric_noise_weights(args, comm, data, freq, mc)
# Add previously simulated sky signal to the atmospheric noise.
pipeline_tools.add_signal(
args, comm, data, totalname_freq, signalname, purge=(nsimu == 1)
)
mcoffset = ifreq * 1000000
pipeline_tools.simulate_noise(
args, comm, data, mc + mcoffset, totalname_freq
)
pipeline_tools.scramble_gains(
args, comm, data, mc + mcoffset, totalname_freq
)
outpath = setup_output(args, comm, mc + mcoffset, freq)
# Bin and destripe maps
pipeline_tools.apply_mapmaker(
args,
comm,
data,
outpath,
totalname_freq,
time_comms=time_comms,
telescope_data=telescope_data,
first_call=(mc == firstmc),
)
if args.apply_polyfilter or args.apply_groundfilter:
# Filter signal
pipeline_tools.apply_polyfilter(args, comm, data, totalname_freq)
pipeline_tools.apply_groundfilter(args, comm, data, totalname_freq)
# Bin filtered maps
pipeline_tools.apply_mapmaker(
args,
comm,
data,
outpath,
totalname_freq,
time_comms=time_comms,
telescope_data=telescope_data,
first_call=False,
extra_prefix="filtered",
bin_only=True,
)
gt.stop_all()
if mpiworld is not None:
mpiworld.barrier()
runtime = gt.seconds("toast_benchmark (science work)")
prefactor = 1.0e-3
kilo_samples = 1.0e-3 * total_samples
sample_factor = 1.2
det_factor = 2.0
metric = (
prefactor
* n_detector ** det_factor
* kilo_samples ** sample_factor
/ (n_nodes * runtime)
)
if rank == 0:
msg = "Science Metric: {:0.1e} * ({:d}**{:0.2f}) * ({:0.3e}**{:0.3f}) / ({:0.1f} * {}) = {:0.2f}".format(
prefactor,
n_detector,
det_factor,
kilo_samples,
sample_factor,
runtime,
n_nodes,
metric,
)
log.info("")
log.info(msg)
log.info("")
with open(os.path.join(args.outdir, "log"), "a") as f:
f.write(msg)
f.write("\n\n")
timer = Timer()
timer.start()
alltimers = gather_timers(comm=mpiworld)
if comm.world_rank == 0:
out = os.path.join(args.outdir, "timing")
dump_timing(alltimers, out)
with open(os.path.join(args.outdir, "log"), "a") as f:
f.write("Copy of Global Timers:\n")
with open("{}.csv".format(out), "r") as t:
f.write(t.read())
timer.stop()
timer.report("Gather and dump timing info")
return
def main():
log = Logger.get()
gt = GlobalTimers.get()
gt.start("toast_ground_sim (total)")
timer0 = Timer()
timer0.start()
mpiworld, procs, rank, comm = get_comm()
args, comm = parse_arguments(comm)
# Initialize madam parameters
madampars = setup_madam(args)
# Load and broadcast the schedule file
schedules = load_schedule(args, comm)
# Load the weather and append to schedules
load_weather(args, comm, schedules)
# load or simulate the focalplane
detweights = load_focalplanes(args, comm, schedules)
# Create the TOAST data object to match the schedule. This will
# include simulating the boresight pointing.
data, telescope_data = create_observations(args, comm, schedules)
# Split the communicator for day and season mapmaking
time_comms = get_time_communicators(args, comm, data)
# Expand boresight quaternions into detector pointing weights and
# pixel numbers
expand_pointing(args, comm, data)
# Purge the pointing if we are NOT going to export the
# data to a TIDAS volume
if (args.tidas is None) and (args.spt3g is None):
for ob in data.obs:
tod = ob["tod"]
tod.free_radec_quats()
# Prepare auxiliary information for distributed map objects
_, localsm, subnpix = get_submaps(args, comm, data)
if args.pysm_model:
focalplanes = [s.telescope.focalplane.detector_data for s in schedules]
signalname = simulate_sky_signal(args, comm, data, focalplanes,
subnpix, localsm, "signal")
else:
signalname = scan_sky_signal(args, comm, data, localsm, subnpix,
"signal")
# Set up objects to take copies of the TOD at appropriate times
totalname, totalname_freq = setup_sigcopy(args)
# Loop over Monte Carlos
firstmc = args.MC_start
nsimu = args.MC_count
freqs = [float(freq) for freq in args.freq.split(",")]
nfreq = len(freqs)
for mc in range(firstmc, firstmc + nsimu):
simulate_atmosphere(args, comm, data, mc, totalname)
# Loop over frequencies with identical focal planes and identical
# atmospheric noise.
for ifreq, freq in enumerate(freqs):
if comm.world_rank == 0:
log.info("Processing frequency {}GHz {} / {}, MC = {}".format(
freq, ifreq + 1, nfreq, mc))
# Make a copy of the atmosphere so we can scramble the gains and apply
# frequency-dependent scaling.
copy_signal(args, comm, data, totalname, totalname_freq)
scale_atmosphere_by_frequency(args,
comm,
data,
freq=freq,
mc=mc,
cache_name=totalname_freq)
update_atmospheric_noise_weights(args, comm, data, freq, mc)
# Add previously simulated sky signal to the atmospheric noise.
add_signal(args,
comm,
data,
totalname_freq,
signalname,
purge=(nsimu == 1))
mcoffset = ifreq * 1000000
simulate_noise(args, comm, data, mc + mcoffset, totalname_freq)
simulate_sss(args, comm, data, mc + mcoffset, totalname_freq)
scramble_gains(args, comm, data, mc + mcoffset, totalname_freq)
if (mc == firstmc) and (ifreq == 0):
# For the first realization and frequency, optionally
# export the timestream data.
output_tidas(args, comm, data, totalname)
output_spt3g(args, comm, data, totalname)
outpath = setup_output(args, comm, mc + mcoffset, freq)
# Bin and destripe maps
apply_madam(
args,
comm,
data,
madampars,
outpath,
detweights,
totalname_freq,
freq=freq,
time_comms=time_comms,
telescope_data=telescope_data,
first_call=(mc == firstmc),
)
if args.apply_polyfilter or args.apply_groundfilter:
# Filter signal
apply_polyfilter(args, comm, data, totalname_freq)
apply_groundfilter(args, comm, data, totalname_freq)
# Bin filtered maps
apply_madam(
args,
comm,
data,
madampars,
outpath,
detweights,
totalname_freq,
freq=freq,
time_comms=time_comms,
telescope_data=telescope_data,
first_call=False,
extra_prefix="filtered",
bin_only=True,
)
gt.stop_all()
if mpiworld is not None:
mpiworld.barrier()
timer = Timer()
timer.start()
alltimers = gather_timers(comm=mpiworld)
if comm.world_rank == 0:
out = os.path.join(args.outdir, "timing")
dump_timing(alltimers, out)
timer.stop()
timer.report("Gather and dump timing info")
timer0.report_clear("toast_ground_sim.py")
return