What is the correct method for using multiple CPU cores with jax.pmap?
The following example creates an environment variable for SPMD on CPU core backends, tests that JAX recognises the devices, and attempts a device lock.
import os
os.environ["XLA_FLAGS"] = '--xla_force_host_platform_device_count=2'
import jax as jx
import jax.numpy as jnp
jx.local_device_count()
# WARNING:absl:No GPU/TPU found, falling back to CPU. (Set TF_CPP_MIN_LOG_LEVEL=0 and rerun for more info.)
# 2
jx.devices("cpu")
# [CpuDevice(id=0), CpuDevice(id=1)]
def sfunc(x): while True: pass
jx.pmap(sfunc)(jnp.arange(2))
Executing from a jupyter kernel and observing htop shows that only one core is locked
I receive the same output from htop when omitting the first two lines and running:
$ env XLA_FLAGS=--xla_force_host_platform_device_count=2 python test.py
Replacing sfunc with
def sfunc(x): return 2.0*x
and calling
jx.pmap(sfunc)(jnp.arange(2))
# ShardedDeviceArray([0., 2.], dtype=float32, weak_type=True)
does return a SharedDeviecArray.
Clearly I am not correctly configuring JAX/XLA to use two cores. What am I missing and what can I do to diagnose the problem?
As far as I can tell, you are configuring the cores correctly (see e.g. Issue #2714). The problem lies in your test function:
def sfunc(x): while True: pass
This function gets stuck in an infinite loop at trace-time, not at run-time. Tracing happens in your host Python process on a single CPU (see How to think in JAX for an introduction to the idea of tracing within JAX transformations).
If you want to observe CPU usage at runtime, you'll have to use a function that finishes tracing and begins running. For that you could use any long-running function that actually produces results. Here is a simple example:
def sfunc(x):
for i in range(100):
x = (x # x)
return x
jx.pmap(sfunc)(jnp.zeros((2, 1000, 1000)))
How can I retrieve a list of tasks in a queue that are yet to be processed?
EDIT: See other answers for getting a list of tasks in the queue.
You should look here:
Celery Guide - Inspecting Workers
Basically this:
my_app = Celery(...)
# Inspect all nodes.
i = my_app.control.inspect()
# Show the items that have an ETA or are scheduled for later processing
i.scheduled()
# Show tasks that are currently active.
i.active()
# Show tasks that have been claimed by workers
i.reserved()
Depending on what you want
If you are using Celery+Django simplest way to inspect tasks using commands directly from your terminal in your virtual environment or using a full path to celery:
Doc: http://docs.celeryproject.org/en/latest/userguide/workers.html?highlight=revoke#inspecting-workers
$ celery inspect reserved
$ celery inspect active
$ celery inspect registered
$ celery inspect scheduled
Also if you are using Celery+RabbitMQ you can inspect the list of queues using the following command:
More info: https://linux.die.net/man/1/rabbitmqctl
$ sudo rabbitmqctl list_queues
if you are using rabbitMQ, use this in terminal:
sudo rabbitmqctl list_queues
it will print list of queues with number of pending tasks. for example:
Listing queues ...
0b27d8c59fba4974893ec22d478a7093 0
0e0a2da9828a48bc86fe993b210d984f 0
10#torob2.celery.pidbox 0
11926b79e30a4f0a9d95df61b6f402f7 0
15c036ad25884b82839495fb29bd6395 1
celerey_mail_worker#torob2.celery.pidbox 0
celery 166
celeryev.795ec5bb-a919-46a8-80c6-5d91d2fcf2aa 0
celeryev.faa4da32-a225-4f6c-be3b-d8814856d1b6 0
the number in right column is number of tasks in the queue. in above, celery queue has 166 pending task.
If you don't use prioritized tasks, this is actually pretty simple if you're using Redis. To get the task counts:
redis-cli -h HOST -p PORT -n DATABASE_NUMBER llen QUEUE_NAME
But, prioritized tasks use a different key in redis, so the full picture is slightly more complicated. The full picture is that you need to query redis for every priority of task. In python (and from the Flower project), this looks like:
PRIORITY_SEP = '\x06\x16'
DEFAULT_PRIORITY_STEPS = [0, 3, 6, 9]
def make_queue_name_for_pri(queue, pri):
"""Make a queue name for redis
Celery uses PRIORITY_SEP to separate different priorities of tasks into
different queues in Redis. Each queue-priority combination becomes a key in
redis with names like:
- batch1\x06\x163 <-- P3 queue named batch1
There's more information about this in Github, but it doesn't look like it
will change any time soon:
- https://github.com/celery/kombu/issues/422
In that ticket the code below, from the Flower project, is referenced:
- https://github.com/mher/flower/blob/master/flower/utils/broker.py#L135
:param queue: The name of the queue to make a name for.
:param pri: The priority to make a name with.
:return: A name for the queue-priority pair.
"""
if pri not in DEFAULT_PRIORITY_STEPS:
raise ValueError('Priority not in priority steps')
return '{0}{1}{2}'.format(*((queue, PRIORITY_SEP, pri) if pri else
(queue, '', '')))
def get_queue_length(queue_name='celery'):
"""Get the number of tasks in a celery queue.
:param queue_name: The name of the queue you want to inspect.
:return: the number of items in the queue.
"""
priority_names = [make_queue_name_for_pri(queue_name, pri) for pri in
DEFAULT_PRIORITY_STEPS]
r = redis.StrictRedis(
host=settings.REDIS_HOST,
port=settings.REDIS_PORT,
db=settings.REDIS_DATABASES['CELERY'],
)
return sum([r.llen(x) for x in priority_names])
If you want to get an actual task, you can use something like:
redis-cli -h HOST -p PORT -n DATABASE_NUMBER lrange QUEUE_NAME 0 -1
From there you'll have to deserialize the returned list. In my case I was able to accomplish this with something like:
r = redis.StrictRedis(
host=settings.REDIS_HOST,
port=settings.REDIS_PORT,
db=settings.REDIS_DATABASES['CELERY'],
)
l = r.lrange('celery', 0, -1)
pickle.loads(base64.decodestring(json.loads(l[0])['body']))
Just be warned that deserialization can take a moment, and you'll need to adjust the commands above to work with various priorities.
To retrieve tasks from backend, use this
from amqplib import client_0_8 as amqp
conn = amqp.Connection(host="localhost:5672 ", userid="guest",
password="guest", virtual_host="/", insist=False)
chan = conn.channel()
name, jobs, consumers = chan.queue_declare(queue="queue_name", passive=True)
A copy-paste solution for Redis with json serialization:
def get_celery_queue_items(queue_name):
import base64
import json
# Get a configured instance of a celery app:
from yourproject.celery import app as celery_app
with celery_app.pool.acquire(block=True) as conn:
tasks = conn.default_channel.client.lrange(queue_name, 0, -1)
decoded_tasks = []
for task in tasks:
j = json.loads(task)
body = json.loads(base64.b64decode(j['body']))
decoded_tasks.append(body)
return decoded_tasks
It works with Django. Just don't forget to change yourproject.celery.
This worked for me in my application:
def get_celery_queue_active_jobs(queue_name):
connection = <CELERY_APP_INSTANCE>.connection()
try:
channel = connection.channel()
name, jobs, consumers = channel.queue_declare(queue=queue_name, passive=True)
active_jobs = []
def dump_message(message):
active_jobs.append(message.properties['application_headers']['task'])
channel.basic_consume(queue=queue_name, callback=dump_message)
for job in range(jobs):
connection.drain_events()
return active_jobs
finally:
connection.close()
active_jobs will be a list of strings that correspond to tasks in the queue.
Don't forget to swap out CELERY_APP_INSTANCE with your own.
Thanks to #ashish for pointing me in the right direction with his answer here: https://stackoverflow.com/a/19465670/9843399
The celery inspect module appears to only be aware of the tasks from the workers perspective. If you want to view the messages that are in the queue (yet to be pulled by the workers) I suggest to use pyrabbit, which can interface with the rabbitmq http api to retrieve all kinds of information from the queue.
An example can be found here:
Retrieve queue length with Celery (RabbitMQ, Django)
I think the only way to get the tasks that are waiting is to keep a list of tasks you started and let the task remove itself from the list when it's started.
With rabbitmqctl and list_queues you can get an overview of how many tasks are waiting, but not the tasks itself: http://www.rabbitmq.com/man/rabbitmqctl.1.man.html
If what you want includes the task being processed, but are not finished yet, you can keep a list of you tasks and check their states:
from tasks import add
result = add.delay(4, 4)
result.ready() # True if finished
Or you let Celery store the results with CELERY_RESULT_BACKEND and check which of your tasks are not in there.
As far as I know Celery does not give API for examining tasks that are waiting in the queue. This is broker-specific. If you use Redis as a broker for an example, then examining tasks that are waiting in the celery (default) queue is as simple as:
connect to the broker
list items in the celery list (LRANGE command for an example)
Keep in mind that these are tasks WAITING to be picked by available workers. Your cluster may have some tasks running - those will not be in this list as they have already been picked.
The process of retrieving tasks in particular queue is broker-specific.
I've come to the conclusion the best way to get the number of jobs on a queue is to use rabbitmqctl as has been suggested several times here. To allow any chosen user to run the command with sudo I followed the instructions here (I did skip editing the profile part as I don't mind typing in sudo before the command.)
I also grabbed jamesc's grep and cut snippet and wrapped it up in subprocess calls.
from subprocess import Popen, PIPE
p1 = Popen(["sudo", "rabbitmqctl", "list_queues", "-p", "[name of your virtula host"], stdout=PIPE)
p2 = Popen(["grep", "-e", "^celery\s"], stdin=p1.stdout, stdout=PIPE)
p3 = Popen(["cut", "-f2"], stdin=p2.stdout, stdout=PIPE)
p1.stdout.close()
p2.stdout.close()
print("number of jobs on queue: %i" % int(p3.communicate()[0]))
If you control the code of the tasks then you can work around the problem by letting a task trigger a trivial retry the first time it executes, then checking inspect().reserved(). The retry registers the task with the result backend, and celery can see that. The task must accept self or context as first parameter so we can access the retry count.
#task(bind=True)
def mytask(self):
if self.request.retries == 0:
raise self.retry(exc=MyTrivialError(), countdown=1)
...
This solution is broker agnostic, ie. you don't have to worry about whether you are using RabbitMQ or Redis to store the tasks.
EDIT: after testing I've found this to be only a partial solution. The size of reserved is limited to the prefetch setting for the worker.
from celery.task.control import inspect
def key_in_list(k, l):
return bool([True for i in l if k in i.values()])
def check_task(task_id):
task_value_dict = inspect().active().values()
for task_list in task_value_dict:
if self.key_in_list(task_id, task_list):
return True
return False
With subprocess.run:
import subprocess
import re
active_process_txt = subprocess.run(['celery', '-A', 'my_proj', 'inspect', 'active'],
stdout=subprocess.PIPE).stdout.decode('utf-8')
return len(re.findall(r'worker_pid', active_process_txt))
Be careful to change my_proj with your_proj
To get the number of tasks on a queue you can use the flower library, here is a simplified example:
from flower.utils.broker import Broker
from django.conf import settings
def get_queue_length(queue):
broker = Broker(settings.CELERY_BROKER_URL)
queues_result = broker.queues([queue])
return queues_result.result()[0]['messages']
I am running airlfow local executor with a postgres database and I am getting a: (psycopg2.OperationalError) FATAL: remaining connection slots are reserved
My configs:
sql_alchemy_pool_size = 5
sql_alchemy_pool_recycle = 1800
sql_alchemy_reconnect_timeout = 300
Still I can see that airflow always keeps more connections open than the pool size.
How can we limit Airflow to actually use the pool limit?
airflow_version = 1.10.2; postgres_version = 9.6
You have forks of the main process as workers, each of which manage their own thread pool.
Check the implementation of LocalExecutor; as it is using multiprocessing under the hood. SqlAlchemy will close any open connections upon forking the LocalWorker; but the pool size will be equivalent to the parent, so at max, theoretically you'd have k * (n + 1) connections, where n is your parallelism constant and k is your sql_alchemy_pool_size.
People having the same problem with celery executor can try reducing scheduler.max_threads (AIRFLOW__SCHEDULER__MAX_THREADS) .
Theoretically Max no connections from scheduler = (sql_alchemy_pool_size + sql_alchemy_max_overflow) x max_threads + 1[Dag processor Manager] + 1[Main scheduler process]
Check out https://stackoverflow.com/a/57411598/5860608 for further info.
I am trying to use parfor loop in Matlab 2013a. However, it gives the error cited below when I try to open matlabpool.
matlabpool open
Starting matlabpool using the 'local' profile ... stopped.
Error using matlabpool (line 144)
Failed to open matlabpool. (For information in addition to the causing error,
validate the profile 'local' in the Cluster Profile Manager.)
Caused by:
Error using parallel.internal.pool.InteractiveClient/start (line 281)
Failed to start matlabpool.
Error using parallel.Cluster/createCommunicatingJob (line 82)
The property "NumWorkersRange" cannot be set after submission.
Or when I use:
n = matlabpool('size')
it gives me 0 as answer.
How can I fix this problem?
My system is equipped with two 2.66 GHz 6core Intel Xeon processors.
MathWorks completely changed the command structure for the Parallel Toolbox. The command you want is parpool. I'd start from there. One nice new feature is that if you don't explicitly start a pool and then call a command like parfor, MATLAB will automatically start one for you.
I have an embarrassingly parallel job that requires no communication between the workers. I'm trying to use the dfeval function, but the overhead seems to be enormous. To get started, I'm trying to run the example from the documentation.
>> matlabpool open
Starting matlabpool using the 'local' configuration ... connected to 8 labs.
>> sched = findResource('scheduler','type','local')
sched =
Local Scheduler Information
===========================
Type : local
ClusterOsType : pc
ClusterSize : 8
DataLocation : C:\Users\~\AppData\Roaming\MathWorks\MATLAB\local_scheduler_data\R2010a
HasSharedFilesystem : true
- Assigned Jobs
Number Pending : 0
Number Queued : 0
Number Running : 1
Number Finished : 8
- Local Specific Properties
ClusterMatlabRoot : C:\Program Files\MATLAB\R2010a
>> matlabpool close force local
Sending a stop signal to all the labs ... stopped.
Did not find any pre-existing parallel jobs created by matlabpool.
>> sched = findResource('scheduler','type','local')
sched =
Local Scheduler Information
===========================
Type : local
ClusterOsType : pc
ClusterSize : 8
DataLocation : C:\Users\~\AppData\Roaming\MathWorks\MATLAB\local_scheduler_data\R2010a
HasSharedFilesystem : true
- Assigned Jobs
Number Pending : 0
Number Queued : 0
Number Running : 0
Number Finished : 8
- Local Specific Properties
ClusterMatlabRoot : C:\Program Files\MATLAB\R2010a
>> tic;y = dfeval(#rand,{1 2 3},'Configuration', 'local');toc
Elapsed time is 4.442944 seconds.
Running subsequent times produces similar timings. So my questions are:
Why do I need to run matlabpool close force local to get the Number Running to zero, given that I run matlabpool open in a fresh instance?
Is five seconds of overhead really necessary for such a trivial example? especially given the Matlab workers have already been started up?
The DFEVAL function is a wrapper around submitting a job with one or more tasks to a given scheduler, in your case the 'local' scheduler. With the 'local' scheduler, each new task runs in a fresh MATLAB worker session, which is why you see the 4.5 second overhead - that's the time take to launch the worker, work out what to do, do it, and then quit.
The reason that you need the number of running jobs to be zero is that the local scheduler can only run a restricted number of workers.
In general, PARFOR with MATLABPOOL is an easier combination to use than DFEVAL. Also, when you open a MATLABPOOL, the workers are launched and ready, so the overhead of PARFOR is much less (but still not zero as the body of the loop needs to be sent to the workers).