def ctx(request, dask_executor):
if request.param == 'inline':
yield Context.make_with('inline')
elif request.param == "dask_executor":
yield Context(executor=dask_executor)
elif request.param == "delayed_default":
yield Context(executor=DelayedJobExecutor())
elif request.param == "delayed_dist":
with distributed.Client(n_workers=2,
threads_per_worker=4,
processes=True) as _:
yield Context(executor=DelayedJobExecutor())
elif request.param == "dask_make_default":
try:
ctx = Context.make_with('dask-make-default')
yield ctx
finally:
# cleanup: Close cluster and client
# This is also tested below, here just to make
# sure things behave as expected.
assert isinstance(ctx.executor, DaskJobExecutor)
ctx.executor.is_local = True
ctx.close()
elif request.param == "dask_integration":
with distributed.Client(n_workers=2,
threads_per_worker=4,
processes=False) as _:
yield Context.make_with("dask-integration")
elif request.param == "concurrent":
yield Context.make_with("threads")
elif request.param == "delayed":
yield Context(executor=DelayedJobExecutor())
def setup_executor(address=None, n_workers=None, threads_per_worker=1, **kwds):
""" Setup a Dask distributed cluster scheduler client
Parameters
----------
address : str, optional
This can be the address of a ``Scheduler`` server, like a string
``'127.0.0.1:8786'``. If ``None``, sets up a ``LocalCluster``
n_workers : int, optional
Number of workers. Only used if setting up a ``LocalCluster``
threads_per_worker : int, optional
Number of threads per worker
kwds
Additional options passed to :py:func:`distributed.Client`
Returns
-------
distributed.Client
Distributed compute client
"""
import distributed
try:
client = distributed.Client(address=address,
n_workers=n_workers,
threads_per_worker=threads_per_worker,
**kwds)
except Exception as e:
logger.exception('Could not start `distributed` cluster')
raise
else:
return client
def ee_dask_deploy(config, pb_id, image, n_workers=1, buffers=[], secrets=[]):
"""Deploy Dask execution engine.
:param config: configuration DB handle
:param pb_id: processing block ID
:param image: Docker image to deploy
:param n_workers: number of Dask workers
:param buffers: list of buffers to mount on Dask workers
:param secrets: list of secrets to mount on Dask workers
:return: deployment ID and Dask client handle
"""
# Make deployment
deploy_id = "proc-{}-dask".format(pb_id)
values = {"image": image, "worker.replicas": n_workers}
for i, b in enumerate(buffers):
values["buffers[{}]".format(i)] = b
for i, s in enumerate(secrets):
values["secrets[{}]".format(i)] = s
deploy = ska_sdp_config.Deployment(
deploy_id, "helm", {"chart": "dask", "values": values}
)
for txn in config.txn():
txn.create_deployment(deploy)
# Wait for scheduler to become available
scheduler = deploy_id + "-scheduler." + os.environ["SDP_HELM_NAMESPACE"] + ":8786"
client = None
while client is None:
try:
client = distributed.Client(scheduler, timeout=1)
except:
pass
return deploy_id, client
def test_retrieve() -> None:
cluster = distributed.LocalCluster(
ip='localhost:8786',
# I want a bokeh interface to check progress
dashboard_address='localhost:8787',
# single process, single thread allows ctrl+C backtrace to
# show where the code is getting stuck. Otherwise, it will say,
# "I'm stuck waiting for other processes." It also makes
# time_code more meaningful
processes=False,
threads_per_worker=1,
)
# TODO: put this in a reusable module
with distributed.Client(cluster):
# disable the cache, because I don't want to persist these results
# in the cloud
for cached_func in [
retrieve.get_rfs, retrieve.get_paragraphs,
retrieve.get_raw_forms, retrieve.get_indexes
]:
assert isinstance(cached_func, Cache)
cast(Cache, cached_func).disabled = True
rfs = dask.bag.zip( # pylint: disable=unused-variable
retrieve.get_indexes('10-K', 1995, 1),
retrieve.get_rfs(1995, 1)).take(10, npartitions=1)
def fullReconstruction(self):
from xicam.Tomography.widgets.volumeviewer import VolumeViewer
volumeviewer = VolumeViewer()
self.recontabs.addTab(volumeviewer, '????')
currentitem = self.headermodel.item(self.rawtabview.currentIndex())
if not currentitem: msg.showMessage('Error: You must open files before reconstructing.')
try:
msg.showBusy()
msg.showMessage('Running slice reconstruction...', level=msg.INFO)
currentheader = self.headermodel.item(self.rawtabview.currentIndex()).header
readprocess = self.workflow.processes[0] # hopefully! TODO: require a readprocess first
readprocess.path.value = currentheader.startdoc['path']
numofsinograms = currentheader.meta_array('primary').shape[1]
executor = DaskExecutor()
client = distributed.Client()
def chunkiterator(workflow):
for i in range(0, int(numofsinograms), int(readprocess.chunksize.value)):
readprocess.sinoindex.value = i
yield executor.execute(workflow)
_reconthread = QThreadFutureIterator(chunkiterator, self.workflow,
callback_slot=partial(self.showReconstruction, mode=self.fullrecon),
except_slot=self.exceptionCallback)
_reconthread.start()
except Exception as ex:
msg.logError(ex)
msg.showReady()
msg.clearMessage()
def get_dask(
self,
ca_path="dask_ca.crt",
client_cert_path="dask_client_cert.pem",
hostname=None,
port=8786,
):
with open(ca_path, "w") as fout:
fout.write(self.config["tls_ca"])
userproxy_path = os.environ.get("X509_USER_PROXY",
"/tmp/x509up_u%d" % os.getuid())
with open(userproxy_path, "rb") as fin:
userproxy = fin.read()
result = self.api.post("/clientkey", data={"proxycert": userproxy})
if result.status_code == 401:
raise RuntimeError(
"Authorization denied while retrieving dask certificate")
elif result.status_code != 200:
raise RuntimeError("Error while retrieving dask certificate")
with open(client_cert_path, "w") as fout:
fout.write(result.text)
sec = distributed.security.Security(
tls_ca_file=ca_path,
tls_client_cert=client_cert_path,
require_encryption=True,
)
if hostname is None:
hostname = self.hostname
url = f"tls://{hostname}:{port}"
return distributed.Client(url, security=sec)
def setup(self):
try:
import distributed
except ImportError:
raise NotImplementedError
self.client = distributed.Client()
self.write = create_delayed_write()
def get_client() -> distributed.Client:
address = os.environ.get('SCHEDULER_PORT', None)
for _ in range(10):
try:
client = distributed.Client(address=address)
except OSError:
continue
else:
break
else:
# try > else > break was never hit
raise ValueError('Could not connect to scheduler')
n_workers_ = os.environ.get('N_WORKERS', None)
if n_workers_ is not None:
n_workers = int(n_workers_)
logging.debug('Waiting for: %d workers', n_workers)
client.wait_for_workers(n_workers)
egg_url = os.environ.get('DEPLOY_EGG', None)
if egg_url is not None:
logging.debug('Downloading and uploading egg: %s', egg_url)
with tempfile.TemporaryDirectory() as tempdir:
egg_path = Path(tempdir) / 'code.egg'
src = urllib.request.urlopen(egg_url)
with egg_path.open('wb') as dst:
shutil.copyfileobj(src, dst)
client.upload_file(str(egg_path))
return client
def executor(kind: str,
max_workers: int,
daemon=True) -> typing.Iterator[Executor]:
"""General purpose utility to get an executor with its as_completed handler
This allows us to easily use other executors as needed.
"""
if kind == "thread":
with ThreadPoolExecutor(max_workers=max_workers) as pool_t:
yield pool_t
elif kind == "process":
with ProcessPoolExecutor(max_workers=max_workers) as pool_p:
yield pool_p
elif kind in ["dask", "dask-process", "dask-thread"]:
import dask
import distributed
from distributed.cfexecutor import ClientExecutor
processes = kind == "dask" or kind == "dask-process"
with dask.config.set({"distributed.worker.daemon": daemon}):
with distributed.LocalCluster(
n_workers=max_workers,
processes=processes,
) as cluster:
with distributed.Client(cluster) as client:
yield ClientExecutor(client)
else:
raise NotImplementedError("That kind is not implemented")
def test_prctl_on_docker(module_scoped_container_getter, tmp_path):
network_info = module_scoped_container_getter.get("scheduler").network_info[0]
client = distributed.Client(
f"tcp://{network_info.hostname}:{network_info.host_port}", set_as_default=False
)
core_test(client, tmp_path)
def dask_client():
cluster = distributed.LocalCluster(n_workers=3, threads_per_worker=1)
client = distributed.Client(cluster)
yield client
client.close()
cluster.close()
def test_use_distributed():
# This Client is pretty cheap to start
# since it only uses threads
with distributed.Client(n_workers=1, threads_per_worker=1,
processes=False) as c:
ctx = Context.make_with("dask-integration")
assert isinstance(ctx.executor, DaskJobExecutor)
assert ctx.executor.client is c
def client(self, attempts=10, **kwargs):
'''Wait for scheduler to be initialized and return Client(self)'''
block(self.instances[0][2])
for i in reversed(range(attempts)):
try:
return distributed.Client(self, **kwargs)
except (TimeoutError, ConnectionRefusedError, OSError) as e:
if i == 0: raise e
def external_client():
# Explicitly specify we want only 4 workers so that when running on
# continuous integration we don't request too many.
cluster = distributed.LocalCluster(n_workers=4)
client = distributed.Client(cluster)
yield client
client.close()
cluster.close()
def get_context(self):
"""Returns Dask Client for Scheduler"""
details = self.get_config_data()
if details is not None:
print("Connect to Dask: %s" % details["master_url"])
client = distributed.Client(details["master_url"])
return client
return None
def dispatch_computations(job_args, tmp_dir, timestamp):
client = distributed.Client('localhost:8786')
webbrowser.open('http://localhost:8787')
jobs = []
for job_arg in job_args:
job = client.submit(compute, *job_arg)
jobs.append(job)
check_jobs(jobs, job_args, tmp_dir, timestamp)
def pipeline_scan(st,
segments=None,
cl=None,
host=None,
cfile=None,
vys_timeout=vys_timeout_default,
mem_read=0.,
mem_search=0.,
throttle=False,
mockseg=None):
""" Given rfpipe state and dask distributed client, run search pipline.
"""
if cl is None:
if host is None:
cl = distributed.Client(
n_workers=1,
threads_per_worker=16,
resources={
"READER": 1,
"MEMORY": 16e9
},
local_dir="/lustre/evla/test/realfast/scratch")
else:
cl = distributed.Client('{0}:{1}'.format(host, '8786'))
if not isinstance(segments, list):
segments = list(range(st.nsegment))
futures = []
sleeptime = throttle * 0.8 * st.nints * st.inttime / st.nsegment # bit shorter than scan
for segment in segments:
futures.append(
pipeline_seg(st,
segment,
cl=cl,
cfile=cfile,
vys_timeout=vys_timeout,
mem_read=mem_read,
mem_search=mem_search,
mockseg=mockseg))
if throttle:
sleep(sleeptime)
return futures # list of tuples of futures (seg, data, cc, acc)
def check_dask(self):
try:
import distributed
client = distributed.Client(self.nodes[0].strip() + ":8786")
print "Found %d workers: %s" % (len(brokers.keys()), str(brokers))
return client.scheduler_info()
except:
pass
return None
def test_executor_disables_watch_worker_events_with_false(self):
with distributed.Client(n_workers=1,
processes=False,
set_as_default=False) as client:
executor = DaskExecutor(address=client.scheduler.address,
watch_worker_status=False)
with executor.start():
assert executor.watch_worker_status is False
assert executor._watch_dask_events_task is None
def test_bias_batch_recommend_dask(ml_folds: MLFolds):
algo = Bias(damping=5)
algo = TopN(algo)
with closing(distributed.Client()), joblib.parallel_backend('dask'):
recs = ml_folds.eval_all(algo, dask=True)
assert isinstance(recs, ddf.DataFrame)
ml_folds.check_positive_ndcg(recs)
def dask_client(mock_service_envs: None) -> Iterable[distributed.Client]:
print(pformat(dask.config.get("distributed")))
with distributed.LocalCluster(
worker_class=distributed.Worker,
**{
"resources": {"CPU": 10, "GPU": 10, "MPI": 1},
"preload": "simcore_service_dask_sidecar.tasks",
},
) as cluster:
with distributed.Client(cluster) as client:
yield client
def test_connect_to_running_cluster(self):
with distributed.Client(processes=False, set_as_default=False) as client:
executor = DaskExecutor(address=client.scheduler.address)
assert executor.address == client.scheduler.address
assert executor.cluster_class is None
assert executor.cluster_kwargs is None
assert executor.client_kwargs == {"set_as_default": False}
with executor.start():
res = executor.wait(executor.submit(lambda x: x + 1, 1))
assert res == 2
def setup(self):
try:
import distributed
except ImportError:
raise NotImplementedError()
# TODO: Lazily skipped in CI as it is very demanding and slow.
# Improve times and remove errors.
_skip_slow()
self.client = distributed.Client()
self.write = create_delayed_write()
def test_executor_enables_watch_worker_events_with_true(self):
with distributed.Client(n_workers=1,
processes=False,
set_as_default=False) as client:
executor = DaskExecutor(
address=client.scheduler.address,
watch_worker_status=True,
adapt_kwargs={"maximum": 4},
)
with executor.start():
assert executor.watch_worker_status is True
assert executor._watch_dask_events_task is not None
def init_client(processes, max_memory):
memory_limit = int(max_memory / processes)
memory_limit = '{0:d}MB'.format(memory_limit)
logger.info(
'Initialising client with {0:d} workers and {1:s} per worker'.format(
processes, memory_limit))
cluster = distributed.LocalCluster(n_workers=processes,
threads_per_worker=1,
memory_limit=memory_limit,
local_directory='/scratch/u/u300636')
client = distributed.Client(cluster)
logger.info('Initialised client: {0}'.format(client))
return client
def execute(self, wf, client=None):
# global client, graph
services = {"machines":
[
{
"name": "freyja",
"address": "freyja.nsls2.bnl.gov",
"port": 22,
"username": "******",
"password": "******",
"environment": {"PYTHONUSERBASE": "/tmp"},
"config_dir": "/tmp/camera"
}
],
"graph":
{
"configure":
[
{
"machine": "freyja",
"apps":
[
{"name": "dask/dask-scheduler"}
],
"tasks": ["dask-cluster"]
}
]
}
}
services = json.dumps(services)
print(services)
graph = task_graph.Graph()
graph.parse_stream(services)
graph.start_tasks()
graph.connect()
graph.execute()
meta_data = graph.machines[0].tasks[0].request_meta_data()
local_port = graph.machines[0].node.get_free_local_port()
remote_port = meta_data[0][0]
print(local_port, remote_port)
graph.machines[0].node.forward_tunnel(local_port, "localhost", remote_port)
client = distributed.Client("tcp://localhost:" + str(local_port))
return super(CamLinkExecutor, self).execute(wf, client)[0]
def start(self):
if (self.tls_ca) or (self.tls_key) or (self.tls_cert):
from distributed.security import Security
security = Security(
tls_client_key=self.tls_key,
tls_client_cert=self.tls_cert,
tls_ca_file=self.tls_ca,
)
else:
security = None
self.client = distributed.Client(self.cluster_address,
security=security)
self.futures = {}
def test_setup_executor_distributed(n_workers=1, threads_per_worker=2):
cluster = distributed.LocalCluster(n_workers=n_workers,
threads_per_worker=threads_per_worker)
client = distributed.Client(cluster)
address = cluster.scheduler.address
test = executor.setup_executor(address)
assert test.scheduler.address == cluster.scheduler.address
assert client.scheduler_info() == test.scheduler_info()
test.close()
cluster.close()
client.close()
def start_jupyterlab():
scheduler_address = get_scheduler_address()
client = distributed.Client(f'{scheduler_address}')
host = client.run_on_scheduler(socket.gethostname)
userid = getpass.getuser()
with open(LOG_DIRECTORY + 'jupyterlab.log', 'wb') as fp:
subprocess.run(f'jupyter-lab --port=7787 --no-browser &',
shell=SHELL,
check=True,
stdout=fp,
stderr=subprocess.STDOUT)
ssh_jpylab = f"ssh -t -L 7787:localhost:7787 -L 8787:localhost:8787 {userid}@theta.alcf.anl.gov ssh -t -L 7787:localhost:7787 -L 8787:localhost:8787 {MOMNODE} ssh -t -L 7787:localhost:7787 -L 8787:localhost:8787 {host}"
print(
f"\nTo connect to JupyterLab and Dask dashboard, execute the following command in a shell on your local machine:\n {ssh_jpylab}\n"
)
logging.info(
f"To connect to JupyterLab and Dask dashboard, execute the following command in a shell on your local machine:\n {ssh_jpylab}\n"
)
jpylab_url = ''
counter = 0
while jpylab_url == '':
time.sleep(1)
#f"sed -n '/^[^[].*localhost:7787/p' {LOG_DIRECTORY}jupyterlab.log",
w = subprocess.run(
f"jupyter server list | sed -n '/localhost:7787/ s/ .*//p'",
shell=SHELL,
capture_output=True,
text=True)
jpylab_url = w.stdout.strip()
counter += 1
if counter > 20:
jpylab_url = 'http://localhost:7787/'
break
print(
f"To open JupyterLab, go to (see log file {LOG_DIRECTORY}jupyterlab.log):\n {jpylab_url}\n"
)
logging.info(
f"To open JupyterLab, go to (see log file {LOG_DIRECTORY}jupyterlab.log):\n {jpylab_url}\n"
)
print(
f"To open the Dask dashboard, go to:\n http://localhost:8787/status\n"
)
logging.info(
f"To open the Dask dashboard, go to:\n http://localhost:8787/status\n"
)
def start(self):
if self.tls_ca or self.tls_key or self.tls_cert:
from distributed.security import Security
security = Security(
tls_client_key=self.tls_key,
tls_client_cert=self.tls_cert,
tls_ca_file=self.tls_ca,
require_encryption=True,
)
else:
security = None
self.client = distributed.Client(self.cluster_address, security=security)
self.futures = {}