Byte Python Concurrent and Parallel Programming V2

Python - Concurrent and
Parallel Programming
threads, locks, processes and events
Introduction
“The free lunch is over” - Herb Sutter, 2005 http://www.gotw.ca/publications/concurrency-ddj.htm
● Moore’s law no longer applies

● Single Core performance stagnates due to physical transistor limits
● Multi Core CPUs in commodity hardware, even Glasses
● It isn’t easy
● But it isn’t going anywhere
Disclaimer
● Python2.7
● Single-machine only
● Not a seasoned python programmer, feel free to correct
Concurrency
A chimpanzee writing Shakespeare and drinking coffee
Parallelism
N chimpanzees writing Shakespeare and drinking coffee, where N > 1.
Processes
● Kernel processes
● Have at least one running thread
● Are scheduled by the OS, not by Python
● multiprocessing allows for IPC
● more on that later
Processes
import multiprocessing as mp
class myProcess(mp.Process):
…
def my_process(foo,bar):
…
p = mp.Process(target=my_process, args=(someFoo, someBar))
p.start() # start the process
p.join() # block the current process until p completes
Threads
● Live within a process
● Share some resources within process
● Are scheduled by the OS
Threads
import threading
class myThread(threading.Thread):
…
def my_thread(foo,bar):
…
t = threading.Thread(target=my_thread, args=(someFoo,
someBar))
t.start() # start the thread

t.join() # block the current thread until t completes
Green threads (gevent, greenlet lib)
● User space threads
● Run within process thread
● Scheduled by the python interpreter
● Live on the memory heap, in process space
Green threads (gevent, greenlet lib)
import gevent
class myGreenlet(Greenlet):
…
def my_greenlet(foo,bar):
…
g = Greenlet(my_greenlet, someFoo, someBar)
g.start()
g.join()
Other cool stuff
● asyncore library
○ socket wrapper for asynchronous eventloop
● eventlet library
○ epoll based (think NodeJS eventloop, GUI)
● asyncio (Py3k, "tulip")
● Twisted (like asyncore, but more of a
framework)
Difficulties with multithreading
"threads it tricky Why use is to"
stdout_lock = threading.Lock()
[thread 1]
with stdout_lock:
print "Why it is tricky"
[thread 2]
with stdout_lock:
print "to use threads"
● Threads are nondeterministic
● Scheduling is done by the OS, not by the Python interpreter
● It is unpredictable when a thread runs, hence code needs to be thread safe
● Threads that use I/O block (wait) for (example filesystem) resources to
become available
● We must use locks to synchronise multithreaded access to shared state
● Synchronising state across threads using locks is difficult and error-prone
Thread locks & synchronisation
● threading.Lock : default lock
○ acquire() closes
○ release() opens
○ subsequent acquire calls on a closed lock block the acquiring thread - even the owning thread
● threading.Rlock : reëntrant lock with ownership counter
○ subsequent acquire calls from the owning thread do not block
■ useful when invoking thread-safe code on the same thread
● threading.Event : simple flag
○ set / clear toggles event to be flagged
○ wait() puts thread in wait state until the event is set
● threading.Condition : conditional lock with notification
○ acquire() blocks
○ notify() signals waiting thread that condition has changed state
○ wait() blocks thread until notify() is called on the lock by owning thread
● threading.Semaphore : lock with a counter
○ acquire() blocks when the amount of acquisitions == semaphore.max
○ release() decreases this
Locking pitfalls
● deadlock
○ thread A → waits on thread B → waits on thread A
○ A waits on C, C waits on B, B waits on A
■ usually circular dependencies or not releasing lock (thread might
be waiting on C for a long time, never releasing resource A)
a.resource_a_lock.acquire() #locks
b.resource_b_lock.acquire() #locks
b.resource_a_lock.acquire() #blocks
b.resource_c_lock.acquire() #blocks
a.resource_b_lock.acquire() #blocks
Locking pitfalls
● race condition
○ thread A retrieves value of bank account f00b4r (EUR 100)
○ thread B gets scheduled and does the same (EUR 100)
○ thread B increases the value by 10% (EUR 110)
○ thread A gets scheduled and increases by 15% (EUR 126,5)
Within shared state multithreading, use a form of lock or event to keep these
Get-Update-Store operations (transactions) atomic (thread-safe)
Use queues to share state between threads
Queue - synchronised task container (it uses simple locks internally)
import Queue
q = Queue.Queue()
q.put() # append a task
q.get() # retrieve a task, wait() if no task available
q.task_done() # decrease task counter
q.join() # block until all tasks are complete
5
B - Buy Beer Brought By Bicycle
How it works
I have:
● a warehouse filled with tasty beer
● a cargo bike
● a cell phone where I can get orders
Pseudocode
while (1)
continue unless order = pop_order_off_stack()
collect_order(order)
ready_bicycle()
route = determine_route(order.address)
bike_to_address(route)
fullfill_transaction(order.price)
return_home(route)
So it got a bit popular
● Hire three employees
○ three messengers (threads)
○ a dispatcher (scheduler)
○ Two order pickers (threads)
● Buy another cargo bike (pool)
1. The dispatcher places the orders in the order queue

2. The order pickers process the order queue and readily available cargo bikes with orders
3. The messengers wait for available ready orders and deliver them. The dispatcher guides them to
their destination.
4. Upon delivery they notify the dispatcher that the order has been delivered. They park their bike
back in the empty bike queue.
def simulate_business():
n_messengers = 10, n_bikes = 15, n_order_pickers = 3, n_orders = 30
q_orders = Queue.Queue()
q_loaded_bikes = Queue.Queue(maxsize=n_bikes)
q_empty_bikes = Queue.Queue(maxsize=n_bikes)
[q_orders.put("order %d" % o) for o in xrange(n_orders)]

[q_empty_bikes.put("bike %d" % b) for b in xrange(n_bikes)]
for m in xrange(n_messengers):
thread = threading.Thread(target=messenger_worker, args=(q_loaded_bikes, q_empty_bikes, q_orders, m))
thread.daemon = True
thread.start()
for o in xrange(n_order_pickers):
thread = threading.Thread(target=order_picker_worker, args=(q_orders, q_empty_bikes, q_loaded_bikes, o))
thread.daemon = True
thread.start()
q_orders.join()
def order_picker_worker(order_queue, empty_queue, ready_queue, n):
def process_order(order):
while True:
try:
empty_bike = empty_queue.get(True, 1)
except Queue.Empty:
continue
sleep(5)
ready_queue.put(empty_bike, True, 1)
while True: ←
try:
beer_order = order_queue.get()
except Queue.Empty:
sleep(5)
continue
process_order(beer_order)
def messenger_worker(full_queue, empty_queue, order_queue, n):
while True:
try:
loaded_bike = full_queue.get(True, 1)
except Queue.Empty:
continue
sleep(10)
empty_queue.put(loaded_bike, True, 1)
order_queue.task_done()
Conditional wait() over while True:
#order picker
ready_cv.acquire()
orders.put(ready_order)
ready_cv.notify()
ready_cv.release()
#messenger
ready_cv.acquire()
while True:
order = orders.get()
if order:
break
ready_cv.wait()
ready_cv.release()
do_stuff_with_order()
Greenlets instead of Threads
import gevent
from gevent.queue import Queue #similar to Queue, tuned for greenlets
gevent.spawn(messenger) #start as greenlet thread

…
gevent.spawn(order_picker)
…
gevent.sleep(0)
…
The GIL holds us back
● The more resources, the slower we get (the dispatcher spends most of her
time switching between messengers)
● Messengers are mostly CPU bound, not I/O. The GIL does not release
often
● We're only ever utilising one core (dispatcher); how can we have multiple
messengers delivering crates in parallel?
Enter multiprocessing!
Difficulties within parallelism
● Amdahl's law
Parallelising any problem by adding more

processes to it is subject to the law of
diminishing returns.
Amdahl's law
S = speedup
N = execution threads
P = parallelisation factor
S(320) = 1 / (1-0.8) + 0.8 / 320 = 5.0025
S(160) = 1 / (1-0.8) + 0.8 / 160 = 5.005
S(80) = 1 / (1-0.8) + 0.8 / 80 = 5.01
S(40) = 1 / (1-0.8) + 0.8 / 40 = 5.02
S(20) = 1 / (1-0.8) + 0.8 / 20 = 5.04
S(15) = 1 / (0.8) + 0.8 / 15 = 5.05
S(10) = 1 / (1-0.8) + 0.8 / 10 = 5.08
S(8) = 1 / (1-0.8) + 0.8 / 8 = 5.1
S(4) = 1 / (1-0.8) + 0.8 / 4 = 5.2
S(2) = 1 / (1-0.8) + 0.8/ 2 = 5.4
Amdahl's law
Multiprocessing
● Spawn/fork processes similarly to thread
● The OS takes care of scheduling the process
● Each has their own Python Interpreter and GIL and can spawn threads
● Process can communicate using Queues, Pipes and Memory maps (shared memory)
○ Or using a different synchronisable medium, Cassandra, SQlite, you name it
● Process isolation means no locking
● Python data can be passed around using pickle, or the faster cpickle
Creating multiple processes
manager = mp.Manager()
queue_1 = manager.Queue()
…
process = mp.Process(target=messenger_worker, args=(queue_1, queue_2…)
process.start()
…
Communication between processes
manager = mp.Manager()
queue_1 = manager.Queue()
…
process = mp.Process(target=messenger_worker, args=(queue_1, queue_2…)
process.start()
…
class Messenger(mp.Process)
…
Communication between processes
current_p, new_p = Pipe() # is actually mp.Connection instances

process = mp.Process(target=messenger_worker, args=(new_p)
process.start()
new_p.send(some_obj)
current_p.rcv() # some_obj
● mp.Connection pickles for you (so does Queue) - this is overhead

Wrapping up
● CPython is multithreading limited b/c GIL
○ GIL can only be lifted (in CPy) using C extensions
● Multithreading with shared state requires synchronisation
● Multiprocessing allows circumvention of GIL
○ multiprocessing library provides many helpful synchronisation
and communication mechanisms
● Consider threading when using lots of I/O
● Consider async (event loop) when using lots of I/O
● Consider MP when using lots of CPU
● Concurrency is fun!
Thanks!
Questions?

Byte Python Concurrent and Parallel Programming V2

Uploaded by

Copyright:

Available Formats

Byte Python Concurrent and Parallel Programming V2

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Byte Python Concurrent and Parallel Programming V2

Uploaded by

Copyright:

Available Formats

Python - Concurrent and

● Moore’s law no longer applies

t.start() # start the thread

1. The dispatcher places the orders in the order queue

[q_orders.put("order %d" % o) for o in xrange(n_orders)]

gevent.spawn(messenger) #start as greenlet thread

Parallelising any problem by adding more

current_p, new_p = Pipe() # is actually mp.Connection instances

● mp.Connection pickles for you (so does Queue) - this is overhead

You might also like