def run_mpi(self):
while True:
bd_comm.Send(np.array([bd_rank], dtype='i'), dest=0)
info = MPI.Status()
bd_comm.Probe(source=0, tag=MPI.ANY_TAG, status=info)
count = info.Get_elements(MPI.BYTE)
chunk = bytearray(count)
bd_comm.Recv(chunk, source=0)
if count == 0:
break
if chunk == bytearray(b'wait'):
continue
pf = PacketFooter(view=chunk)
smd_chunk, epics_chunk = pf.split_packets()
pfe = PacketFooter(view=epics_chunk)
epics_views = pfe.split_packets()
self.run.epics_store.update(epics_views)
if self.run.scan:
yield Step(self.run, smd_batch=smd_chunk)
else:
for event in self.evt_man.events(smd_chunk):
yield event
def run_mpi(self):
rankreq = np.empty(1, dtype='i')
smd_comm = self.run.comms.smd_comm
n_bd_nodes = self.run.comms.bd_comm.Get_size() - 1
bd_comm = self.run.comms.bd_comm
smd_rank = self.run.comms.smd_rank
step_hist = StepHistory(self.run.comms.bd_size, len(self.run.configs))
while True:
smd_comm.Send(np.array([smd_rank], dtype='i'), dest=0)
info = MPI.Status()
smd_comm.Probe(source=0, status=info)
count = info.Get_elements(MPI.BYTE)
smd_chunk = bytearray(count)
smd_comm.Recv(smd_chunk, source=0)
if not smd_chunk:
break
eb_man = EventBuilderManager(smd_chunk, self.run)
# Build batch of events
for smd_batch_dict in eb_man.batches():
# If single item and dest_rank=0, send to any bigdata nodes.
if 0 in smd_batch_dict.keys():
smd_batch, _ = smd_batch_dict[0]
bd_comm.Recv(rankreq, source=MPI.ANY_SOURCE)
missing_step_views = step_hist.get_buffer(rankreq[0])
step_pf = PacketFooter(view=eb_man.step_chunk())
step_views = step_pf.split_packets()
step_hist.extend_buffers(step_views, rankreq[0])
batch = repack_for_bd(smd_batch, missing_step_views,
self.run.configs)
bd_comm.Send(batch, dest=rankreq[0])
# With > 1 dest_rank, start looping until all dest_rank batches
# have been sent.
else:
while smd_batch_dict:
bd_comm.Recv(rankreq, source=MPI.ANY_SOURCE)
if rankreq[0] in smd_batch_dict:
smd_batch, _ = smd_batch_dict[rankreq[0]]
missing_step_views = step_hist.get_buffer(
rankreq[0])
step_pf = PacketFooter(view=eb_man.step_chunk())
step_hist.extend_buffers(step_pf.split_packets(),
rankreq[0])
batch = repack_for_bd(smd_batch,
missing_step_views,
self.run.configs)
bd_comm.Send(batch, dest=rankreq[0])
del smd_batch_dict[rankreq[0]] # done sending
else:
bd_comm.Send(bytearray(b'wait'), dest=rankreq[0])
# Done - kill all alive bigdata nodes
for i in range(n_bd_nodes):
bd_comm.Recv(rankreq, source=MPI.ANY_SOURCE)
bd_comm.Send(bytearray(), dest=rankreq[0])
def run_mpi(self):
rankreq = np.empty(1, dtype='i')
for (smd_chunk, update_chunk) in self.smdr_man.chunks():
# Creates a chunk from smd and epics data to send to SmdNode
# Anatomy of a chunk (pf=packet_footer):
# [ [smd0][smd1][smd2][pf] ][ [epics0][epics1][epics2][pf] ][ pf ]
# ----- smd_chunk ------ ---------epics_chunk-------
# -------------------------- chunk ------------------------------
# Read new epics data as available in the queue
# then send only unseen portion of data to the evtbuilder rank.
update_pf = PacketFooter(view=update_chunk)
self.epics_man.extend_buffers(update_pf.split_packets())
smd_comm.Recv(rankreq, source=MPI.ANY_SOURCE)
epics_chunk = self.epics_man.get_buffer(rankreq[0])
pf = PacketFooter(2)
pf.set_size(0, memoryview(smd_chunk).shape[0])
pf.set_size(1, memoryview(epics_chunk).shape[0])
chunk = smd_chunk + epics_chunk + pf.footer
smd_comm.Send(chunk, dest=rankreq[0])
for i in range(PS_SMD_NODES):
smd_comm.Recv(rankreq, source=MPI.ANY_SOURCE)
smd_comm.Send(bytearray(), dest=rankreq[0])
def run_mpi(self):
rankreq = np.empty(1, dtype='i')
for (smd_chunk, step_chunk) in self.smdr_man.chunks():
# Creates a chunk from smd and epics data to send to SmdNode
# Anatomy of a chunk (pf=packet_footer):
# [ [smd0][smd1][smd2][pf] ][ [epics0][epics1][epics2][pf] ][ pf ]
# ----- smd_chunk ------ ---------epics_chunk-------
# -------------------------- chunk ------------------------------
# Read new epics data as available in the queue
# then send only unseen portion of data to the evtbuilder rank.
if not smd_chunk: break
self.run.comms.smd_comm.Recv(rankreq, source=MPI.ANY_SOURCE)
# Check missing steps for the current client
missing_step_views = self.step_hist.get_buffer(rankreq[0])
# Update this node's step buffers
step_pf = PacketFooter(view=step_chunk)
step_views = step_pf.split_packets()
self.step_hist.extend_buffers(step_views, rankreq[0])
smd_extended = repack_for_eb(smd_chunk, missing_step_views,
self.run.configs)
self.run.comms.smd_comm.Send(smd_extended, dest=rankreq[0])
for i in range(self.run.comms.n_smd_nodes):
self.run.comms.smd_comm.Recv(rankreq, source=MPI.ANY_SOURCE)
self.run.comms.smd_comm.Send(bytearray(), dest=rankreq[0])
def events(self, view):
pf = PacketFooter(view=view)
views = pf.split_packets()
# Keeps offset, size, & timestamp for all events in the batch
# for batch reading (if filter_fn is not given).
ofsz_batch = np.zeros((pf.n_packets, self.n_smd_files, 2),
dtype=np.intp)
has_offset = True
for i, event_bytes in enumerate(views):
if event_bytes:
evt = Event._from_bytes(self.smd_configs, event_bytes)
if not evt._has_offset:
has_offset = False
yield evt # no offset info in the smalldata event
else:
segment = 0 # "d.info" "detectors" have only one segment
ofsz = np.asarray([[d.info[segment].offsetAlg.intOffset, d.info[segment].offsetAlg.intDgramSize] \
for d in evt])
ofsz_batch[i, :, :] = ofsz
# Only get big data one event at a time when filter is off
if self.filter_fn:
bd_evt = self.dm.jump(ofsz[:, 0], ofsz[:, 1])
yield bd_evt
if self.filter_fn == 0 and has_offset:
# Read chunks of 'size' bytes and store them in views
views = [None] * self.n_smd_files
view_sizes = np.zeros(self.n_smd_files)
for i in range(self.n_smd_files):
# If no data were filtered, we can assume that all bigdata
# dgrams starting from the first offset are stored consecutively
# in the file. We read a chunk of sum(all dgram sizes) and
# store in a view.
offset = ofsz_batch[0, i, 0]
size = np.sum(ofsz_batch[:, i, 1])
view_sizes[i] = size
os.lseek(self.dm.fds[i], offset, 0)
views[i] = os.read(self.dm.fds[i], size)
# Build each event from these views
dgrams = [None] * self.n_smd_files
offsets = [0] * self.n_smd_files
for i in range(pf.n_packets):
for j in range(self.n_smd_files):
if offsets[j] >= view_sizes[j]:
continue
size = ofsz_batch[i, j, 1]
if size:
dgrams[j] = dgram.Dgram(view=views[j],
config=self.dm.configs[j],
offset=offsets[j])
offsets[j] += size
bd_evt = Event(dgrams)
yield bd_evt
def __init__(self, view, run):
self.configs = run.configs
self.batch_size = run.batch_size
self.filter_fn = run.filter_callback
self.destination = run.destination
self.n_files = len(self.configs)
pf = PacketFooter(view=view)
views = pf.split_packets()
self.eb = EventBuilder(views)
def __init__(self, view, run):
self.configs = run.configs
self.batch_size = run.batch_size
self.filter_fn = run.filter_callback
self.destination = run.destination
self.n_files = len(self.configs)
pf = PacketFooter(view=view)
views = pf.split_packets()
self.eb = EventBuilder(views, self.configs)
self.c_filter = PrometheusManager.get_metric('psana_eb_filter')
def __init__(self,
view,
configs,
batch_size=1,
filter_fn=0,
destination=0):
self.configs = configs
self.batch_size = batch_size
self.filter_fn = filter_fn
self.destination = destination
pf = PacketFooter(view=view)
views = pf.split_packets()
self.eb = EventBuilder(views)
def test_contents(self):
view = bytearray()
pf = PacketFooter(2)
for i, msg in enumerate([b'packet0', b'packet1']):
view.extend(msg)
pf.set_size(i, memoryview(msg).shape[0])
view.extend(pf.footer)
pf2 = PacketFooter(view=view)
assert pf2.n_packets == 2
views = pf2.split_packets()
assert memoryview(views[0]).shape[0] == 7
assert memoryview(views[1]).shape[0] == 7
def extend_buffers(self, views, client_id, as_event=False):
idx = client_id - 1 # rank 0 has no send history.
# Views is either list of smdchunks or events
if not as_event:
# For Smd0
for i_smd, view in enumerate(views):
self.bufs[i_smd].extend(view)
self.send_history[idx][i_smd] += view.nbytes
else:
# For EventBuilder
for i_evt, evt_bytes in enumerate(views):
pf = PacketFooter(view=evt_bytes)
assert pf.n_packets == self.n_smds
for i_smd, dg_bytes in enumerate(pf.split_packets()):
self.bufs[i_smd].extend(dg_bytes)
self.send_history[idx][i_smd] += dg_bytes.nbytes
def _from_bytes(cls, configs, event_bytes, run=None):
dgrams = []
if event_bytes:
pf = PacketFooter(view=event_bytes)
views = pf.split_packets()
assert len(configs) == pf.n_packets
dgrams = [None]*pf.n_packets # make sure that dgrams are arranged
# according to the smd files.
for i in range(pf.n_packets):
if views[i].shape[0] > 0: # do not include any missing dgram
dgrams[i] = dgram.Dgram(config=configs[i], view=views[i])
evt = cls(dgrams, run=run)
return evt
def __init__(self, view, smd_configs, dm, filter_fn=0):
if view:
pf = PacketFooter(view=view)
self.smd_events = pf.split_packets()
self.n_events = pf.n_packets
else:
self.smd_events = None
self.n_events = 0
self.smd_configs = smd_configs
self.dm = dm
self.n_smd_files = len(self.smd_configs)
self.filter_fn = filter_fn
self.cn_events = 0
if not self.filter_fn and len(self.dm.xtc_files) > 0:
self._read_bigdata_in_chunk()
def run_mpi(self):
rankreq = np.empty(1, dtype='i')
for (smd_chunk, step_chunk) in self.smdr_man.chunks():
# Creates a chunk from smd and epics data to send to SmdNode
# Anatomy of a chunk (pf=packet_footer):
# [ [smd0][smd1][smd2][pf] ][ [epics0][epics1][epics2][pf] ][ pf ]
# ----- smd_chunk ------ ---------epics_chunk-------
# -------------------------- chunk ------------------------------
# Read new epics data as available in the queue
# then send only unseen portion of data to the evtbuilder rank.
if not smd_chunk: break
st_req = time.time()
self.run.comms.smd_comm.Recv(rankreq, source=MPI.ANY_SOURCE)
en_req = time.time()
# Check missing steps for the current client
missing_step_views = self.step_hist.get_buffer(rankreq[0])
# Update step buffers (after getting the missing steps
step_pf = PacketFooter(view=step_chunk)
step_views = step_pf.split_packets()
self.step_hist.extend_buffers(step_views, rankreq[0])
smd_extended = repack_for_eb(smd_chunk, missing_step_views,
self.run.configs)
self.run.comms.smd_comm.Send(smd_extended, dest=rankreq[0])
# sending data to prometheus
self.c_sent.labels('evts',
rankreq[0]).inc(self.smdr_man.got_events)
self.c_sent.labels('batches', rankreq[0]).inc()
self.c_sent.labels('MB', rankreq[0]).inc(
memoryview(smd_extended).nbytes / 1e6)
self.c_sent.labels('seconds', rankreq[0]).inc(en_req - st_req)
logging.debug(
f'node.py: Smd0 sent {self.smdr_man.got_events} events to {rankreq[0]} (waiting for this rank took {en_req-st_req:.5f} seconds)'
)
for i in range(self.run.comms.n_smd_nodes):
self.run.comms.smd_comm.Recv(rankreq, source=MPI.ANY_SOURCE)
self.run.comms.smd_comm.Send(bytearray(), dest=rankreq[0])
def repack_for_eb(smd_chunk, step_views, configs):
""" Smd0 uses this to prepend missing step views
to the smd_chunk (just data with the same limit timestamp from all
smd files - not event-built yet).
"""
if step_views:
smd_chunk_pf = PacketFooter(view=smd_chunk)
new_chunk_pf = PacketFooter(n_packets=smd_chunk_pf.n_packets)
new_chunk = bytearray()
for i, (smd_view, step_view) in enumerate(
zip(smd_chunk_pf.split_packets(), step_views)):
new_chunk.extend(step_view + bytearray(smd_view))
new_chunk_pf.set_size(
i,
memoryview(step_view).nbytes + smd_view.nbytes)
new_chunk.extend(new_chunk_pf.footer)
return new_chunk
else:
return smd_chunk
def _send_to_dest(self, dest_rank, smd_batch_dict, step_batch_dict,
eb_man):
bd_comm = self.run.comms.bd_comm
smd_batch, _ = smd_batch_dict[dest_rank]
missing_step_views = self.step_hist.get_buffer(dest_rank)
batch = repack_for_bd(smd_batch,
missing_step_views,
self.run.configs,
client=dest_rank)
bd_comm.Send(batch, dest=dest_rank)
del smd_batch_dict[dest_rank] # done sending
step_batch, _ = step_batch_dict[dest_rank]
if eb_man.eb.nsteps > 0 and memoryview(step_batch).nbytes > 0:
step_pf = PacketFooter(view=step_batch)
self.step_hist.extend_buffers(step_pf.split_packets(),
dest_rank,
as_event=True)
del step_batch_dict[dest_rank] # done adding
def steps(self):
current_step_pos = 0
""" Generates events between steps. """
smdr_man = SmdReaderManager(self.smd_dm.fds, self.max_events)
for i, (smd_chunk, step_chunk) in enumerate(smdr_man.chunks()):
# Update step stores
step_pf = PacketFooter(view=step_chunk)
step_views = step_pf.split_packets()
self.epics_store.update(step_views)
self.step_store.update(step_views)
eb_man = EventBuilderManager(smd_chunk, self.configs, \
batch_size=self.batch_size, filter_fn=self.filter_callback)
for i, step_dgram in enumerate(
self.step_store.dgrams(from_pos=current_step_pos + 1)):
if step_dgram:
limit_ts = step_dgram.seq.timestamp()
current_step_pos += 1
else:
limit_ts = -1
yield Step(self, eb_man=eb_man, limit_ts=limit_ts)
def __init__(self, view, smd_configs, dm, filter_fn=0):
if view:
if view == bytearray(
b'wait'
): # RunParallel (unused bigdata nodes get this wait msg)
self.smd_events = None
self.n_events = 0
else:
pf = PacketFooter(view=view)
self.smd_events = pf.split_packets()
self.n_events = pf.n_packets
else:
self.smd_events = None
self.n_events = 0
self.smd_configs = smd_configs
self.dm = dm
self.n_smd_files = len(self.smd_configs)
self.filter_fn = filter_fn
self.cn_events = 0
if not self.filter_fn and len(self.dm.xtc_files) > 0:
self._read_bigdata_in_chunk()
def events(self):
ev_man = EventManager(self.configs, self.dm, \
filter_fn=self.filter_callback)
#get smd chunks
smdr_man = SmdReaderManager(self.smd_dm.fds, self.max_events)
for (smd_chunk, step_chunk) in smdr_man.chunks():
# Update epics_store for each chunk
step_pf = PacketFooter(view=step_chunk)
step_views = step_pf.split_packets()
self.epics_store.update(step_views)
eb_man = EventBuilderManager(smd_chunk,
self.configs,
batch_size=self.batch_size,
filter_fn=self.filter_callback)
for batch_dict in eb_man.batches():
batch, _ = batch_dict[
0] # there's only 1 dest_rank for serial run
for evt in ev_man.events(batch):
if evt._dgrams[0].seq.service() != 12: continue
yield evt
def run_mpi(self):
rankreq = np.empty(1, dtype='i')
current_update_pos = 0
while True:
# handles requests from smd_0
smd_comm.Send(np.array([smd_rank], dtype='i'), dest=0)
info = MPI.Status()
smd_comm.Probe(source=0, status=info)
count = info.Get_elements(MPI.BYTE)
chunk = bytearray(count)
smd_comm.Recv(chunk, source=0)
if count == 0:
break
# Unpack the chunk received from Smd0
pf = PacketFooter(view=chunk)
smd_chunk, update_chunk = pf.split_packets()
eb_man = EventBuilderManager(smd_chunk, self.run.configs, \
batch_size=self.run.batch_size, filter_fn=self.run.filter_callback, \
destination=self.run.destination)
# Unpack epics chunk and updates run's epics_store and epics_manager
pfe = PacketFooter(view=update_chunk)
update_views = pfe.split_packets()
self.run.epics_store.update(update_views)
self.run.step_store.update(update_views)
self.epics_man.extend_buffers(update_views)
# Build batch of events
for update_dgram in self.run.step_store.dgrams(
from_pos=current_update_pos + 1, scan=self.run.scan):
if update_dgram:
limit_ts = update_dgram.seq.timestamp()
current_update_pos += 1
else:
limit_ts = -1
for smd_batch_dict in eb_man.batches(limit_ts=limit_ts):
# If single item and dest_rank=0, send to any bigdata nodes.
if 0 in smd_batch_dict.keys():
smd_batch, _ = smd_batch_dict[0]
bd_comm.Recv(rankreq, source=MPI.ANY_SOURCE)
epics_batch = self.epics_man.get_buffer(rankreq[0])
batch = self.pack(smd_batch, epics_batch)
bd_comm.Send(batch, dest=rankreq[0])
# With > 1 dest_rank, start looping until all dest_rank batches
# have been sent.
else:
while smd_batch_dict:
bd_comm.Recv(rankreq, source=MPI.ANY_SOURCE)
if rankreq[0] in smd_batch_dict:
smd_batch, _ = smd_batch_dict[rankreq[0]]
epics_batch = self.epics_man.get_buffer(
rankreq[0])
batch = self.pack(smd_batch, epics_batch)
bd_comm.Send(batch, dest=rankreq[0])
del smd_batch_dict[rankreq[0]] # done sending
else:
bd_comm.Send(bytearray(b'wait'),
dest=rankreq[0])
# Done - kill all alive bigdata nodes
for i in range(self.n_bd_nodes):
bd_comm.Recv(rankreq, source=MPI.ANY_SOURCE)
bd_comm.Send(bytearray(), dest=rankreq[0])
def run_mpi(self):
rankreq = np.empty(1, dtype='i')
smd_comm = self.run.comms.smd_comm
n_bd_nodes = self.run.comms.bd_comm.Get_size() - 1
bd_comm = self.run.comms.bd_comm
smd_rank = self.run.comms.smd_rank
while True:
smd_comm.Send(np.array([smd_rank], dtype='i'), dest=0)
info = MPI.Status()
smd_comm.Probe(source=0, status=info)
count = info.Get_elements(MPI.BYTE)
smd_chunk = bytearray(count)
smd_comm.Recv(smd_chunk, source=0)
if not smd_chunk:
break
eb_man = EventBuilderManager(smd_chunk, self.run)
# Build batch of events
for smd_batch_dict, step_batch_dict in eb_man.batches():
# If single item and dest_rank=0, send to any bigdata nodes.
if 0 in smd_batch_dict.keys():
smd_batch, _ = smd_batch_dict[0]
step_batch, _ = step_batch_dict[0]
bd_comm.Recv(rankreq, source=MPI.ANY_SOURCE)
missing_step_views = self.step_hist.get_buffer(rankreq[0])
batch = repack_for_bd(smd_batch,
missing_step_views,
self.run.configs,
client=rankreq[0])
bd_comm.Send(batch, dest=rankreq[0])
if eb_man.eb.nsteps > 0 and memoryview(
step_batch).nbytes > 0:
step_pf = PacketFooter(view=step_batch)
self.step_hist.extend_buffers(step_pf.split_packets(),
rankreq[0],
as_event=True)
# With > 1 dest_rank, start looping until all dest_rank batches
# have been sent.
else:
# Check if destinations are valid
destinations = np.asarray(list(smd_batch_dict.keys()))
if any(destinations > n_bd_nodes):
print(
f"Found invalid destination ({destinations}). Must be <= {n_bd_nodes} (#big data nodes)"
)
break
while smd_batch_dict:
sent = False
if self.waiting_bds: # Check first if there are bd nodes waiting
copied_waiting_bds = self.waiting_bds[:]
for dest_rank in copied_waiting_bds:
if dest_rank in smd_batch_dict:
self._send_to_dest(dest_rank,
smd_batch_dict,
step_batch_dict, eb_man)
self.waiting_bds.remove(dest_rank)
sent = True
if not sent:
bd_comm.Recv(rankreq, source=MPI.ANY_SOURCE)
dest_rank = rankreq[0]
if dest_rank in smd_batch_dict:
self._send_to_dest(dest_rank, smd_batch_dict,
step_batch_dict, eb_man)
else:
self.waiting_bds.append(dest_rank)
# Done
# - kill idling nodes
for dest_rank in self.waiting_bds:
bd_comm.Send(bytearray(), dest=dest_rank)
# - kill all other nodes
for i in range(n_bd_nodes - len(self.waiting_bds)):
bd_comm.Recv(rankreq, source=MPI.ANY_SOURCE)
bd_comm.Send(bytearray(), dest=rankreq[0])
from psana.parallelreader import ParallelReader
from psana.psexp.packet_footer import PacketFooter
import os, glob
from psana import dgram
tmp_dir = os.path.join(os.environ.get('TEST_XTC_DIR', os.getcwd()), '.tmp')
xtc_files = [
os.path.join(tmp_dir, 'data-r0001-s00.xtc2'),
os.path.join(tmp_dir, 'data-r0001-s01.xtc2')
]
fds = [os.open(xtc_file, os.O_RDONLY) for xtc_file in xtc_files]
configs = [dgram.Dgram(file_descriptor=fd) for fd in fds]
prl_reader = ParallelReader(fds)
block = prl_reader.get_block()
pf = PacketFooter(view=block)
views = pf.split_packets()
for i in range(len(views)):
config, view = configs[i], views[i]
d = dgram.Dgram(config=config, view=view)
#assert getattr(d.epics[0].fast, 'HX2:DVD:GCC:01:PMON') == 41.0
#assert getattr(d.epics[0].slow, 'XPP:GON:MMS:01:RBV') == 41.0
def run_mpi(self):
rankreq = np.empty(1, dtype='i')
smd_comm = self.run.comms.smd_comm
n_bd_nodes = self.run.comms.bd_comm.Get_size() - 1
bd_comm = self.run.comms.bd_comm
smd_rank = self.run.comms.smd_rank
while True:
smd_chunk = self._request_data(smd_comm)
if not smd_chunk:
break
eb_man = EventBuilderManager(smd_chunk, self.run)
# Build batch of events
for smd_batch_dict, step_batch_dict in eb_man.batches():
# If single item and dest_rank=0, send to any bigdata nodes.
if 0 in smd_batch_dict.keys():
smd_batch, _ = smd_batch_dict[0]
step_batch, _ = step_batch_dict[0]
self._request_rank(rankreq)
missing_step_views = self.step_hist.get_buffer(rankreq[0])
batch = repack_for_bd(smd_batch,
missing_step_views,
self.run.configs,
client=rankreq[0])
bd_comm.Send(batch, dest=rankreq[0])
# sending data to prometheus
logging.debug(
'node.py: EventBuilder sent %d events (%.2f MB) to rank %d'
% (eb_man.eb.nevents, memoryview(batch).nbytes / 1e6,
rankreq[0]))
self.c_sent.labels('evts',
rankreq[0]).inc(eb_man.eb.nevents)
self.c_sent.labels('batches', rankreq[0]).inc()
self.c_sent.labels('MB', rankreq[0]).inc(
memoryview(batch).nbytes / 1e6)
if eb_man.eb.nsteps > 0 and memoryview(
step_batch).nbytes > 0:
step_pf = PacketFooter(view=step_batch)
self.step_hist.extend_buffers(step_pf.split_packets(),
rankreq[0],
as_event=True)
# With > 1 dest_rank, start looping until all dest_rank batches
# have been sent.
else:
# Check if destinations are valid
destinations = np.asarray(list(smd_batch_dict.keys()))
if any(destinations > n_bd_nodes):
print(
f"Found invalid destination ({destinations}). Must be <= {n_bd_nodes} (#big data nodes)"
)
break
while smd_batch_dict:
sent = False
if self.waiting_bds: # Check first if there are bd nodes waiting
copied_waiting_bds = self.waiting_bds[:]
for dest_rank in copied_waiting_bds:
if dest_rank in smd_batch_dict:
self._send_to_dest(dest_rank,
smd_batch_dict,
step_batch_dict, eb_man)
self.waiting_bds.remove(dest_rank)
sent = True
if not sent:
self._request_rank(rankreq)
dest_rank = rankreq[0]
if dest_rank in smd_batch_dict:
self._send_to_dest(dest_rank, smd_batch_dict,
step_batch_dict, eb_man)
else:
self.waiting_bds.append(dest_rank)
# Done
# - kill idling nodes
for dest_rank in self.waiting_bds:
bd_comm.Send(bytearray(), dest=dest_rank)
# - kill all other nodes
for i in range(n_bd_nodes - len(self.waiting_bds)):
self._request_rank(rankreq)
bd_comm.Send(bytearray(), dest=rankreq[0])
