I want to provide a service on the web that people can test out the performance of an algo, which is written in python and running on the linux machine
basically what I want to do is that, there is a very trivial PHP handler, let's say start_algo.php, which accepts the request coming from browser, and in the php code through system() or popen() (something like exec( "python algo.py" ) ) to issue a new process running the python script, I think it is doable in this part
problem is that since it is a web service, surely it has to serve multiple users at the same time, but I am quite confused by the Global Interpreter Lock GIL http://wiki.python.org/moin/GlobalInterpreterLock
that the 'standard' CPython has implemented,
does it mean, if I have 3 users running the algo now (which means 3 separated processes, correct me if I am wrong plz), at a particular moment, there is only one user is being served by the Python interpreter and the other 2 are waiting for their turns?
Many thanks in advance
Ted
If you are opening each script by invoking a new process; you will not run afoul of the GIL. Each process gets its own interpreter and therefore its own interpreter lock.
The GIL is per-process. If you start multiple python processes, each will have its own GIL that prevents the interpreter(s) in this specific process from running more than one thread at a time. But independent processes can run at the same time.
Also, multiple threads inside one Python process do take turns on running (rather frequently, IIRC once per hundred opcode instructions or a few dozen milliseconds depending on the version), so it's not like the GIL prevents concurrency at all - it just prevents multi-threading.
Related
I've been trying to wrap my head around how threads work in Python, and it's hard to find good information on how they operate. I may just be missing a link or something, but it seems like the official documentation isn't very thorough on the subject, and I haven't been able to find a good write-up.
From what I can tell, only one thread can be running at once, and the active thread switches every 10 instructions or so?
Where is there a good explanation, or can you provide one? It would also be very nice to be aware of common problems that you run into while using threads with Python.
Yes, because of the Global Interpreter Lock (GIL) there can only run one thread at a time. Here are some links with some insights about this:
http://www.artima.com/weblogs/viewpost.jsp?thread=214235
http://smoothspan.wordpress.com/2007/09/14/guido-is-right-to-leave-the-gil-in-python-not-for-multicore-but-for-utility-computing/
From the last link an interesting quote:
Let me explain what all that means.
Threads run inside the same virtual
machine, and hence run on the same
physical machine. Processes can run
on the same physical machine or in
another physical machine. If you
architect your application around
threads, you’ve done nothing to access
multiple machines. So, you can scale
to as many cores are on the single
machine (which will be quite a few
over time), but to really reach web
scales, you’ll need to solve the
multiple machine problem anyway.
If you want to use multi core, pyprocessing defines an process based API to do real parallelization. The PEP also includes some interesting benchmarks.
Python's a fairly easy language to thread in, but there are caveats. The biggest thing you need to know about is the Global Interpreter Lock. This allows only one thread to access the interpreter. This means two things: 1) you rarely ever find yourself using a lock statement in python and 2) if you want to take advantage of multi-processor systems, you have to use separate processes. EDIT: I should also point out that you can put some of the code in C/C++ if you want to get around the GIL as well.
Thus, you need to re-consider why you want to use threads. If you want to parallelize your app to take advantage of dual-core architecture, you need to consider breaking your app up into multiple processes.
If you want to improve responsiveness, you should CONSIDER using threads. There are other alternatives though, namely microthreading. There are also some frameworks that you should look into:
stackless python
greenlets
gevent
monocle
Below is a basic threading sample. It will spawn 20 threads; each thread will output its thread number. Run it and observe the order in which they print.
import threading
class Foo (threading.Thread):
def __init__(self,x):
self.__x = x
threading.Thread.__init__(self)
def run (self):
print str(self.__x)
for x in xrange(20):
Foo(x).start()
As you have hinted at Python threads are implemented through time-slicing. This is how they get the "parallel" effect.
In my example my Foo class extends thread, I then implement the run method, which is where the code that you would like to run in a thread goes. To start the thread you call start() on the thread object, which will automatically invoke the run method...
Of course, this is just the very basics. You will eventually want to learn about semaphores, mutexes, and locks for thread synchronization and message passing.
Note: wherever I mention thread i mean specifically threads in python until explicitly stated.
Threads work a little differently in python if you are coming from C/C++ background. In python, Only one thread can be in running state at a given time.This means Threads in python cannot truly leverage the power of multiple processing cores since by design it's not possible for threads to run parallelly on multiple cores.
As the memory management in python is not thread-safe each thread require an exclusive access to data structures in python interpreter.This exclusive access is acquired by a mechanism called GIL ( global interpretr lock ).
Why does python use GIL?
In order to prevent multiple threads from accessing interpreter state simultaneously and corrupting the interpreter state.
The idea is whenever a thread is being executed (even if it's the main thread), a GIL is acquired and after some predefined interval of time the
GIL is released by the current thread and reacquired by some other thread( if any).
Why not simply remove GIL?
It is not that its impossible to remove GIL, its just that in prcoess of doing so we end up putting mutiple locks inside interpreter in order to serialize access, which makes even a single threaded application less performant.
so the cost of removing GIL is paid off by reduced performance of a single threaded application, which is never desired.
So when does thread switching occurs in python?
Thread switch occurs when GIL is released.So when is GIL Released?
There are two scenarios to take into consideration.
If a Thread is doing CPU Bound operations(Ex image processing).
In Older versions of python , Thread switching used to occur after a fixed no of python instructions.It was by default set to 100.It turned out that its not a very good policy to decide when switching should occur since the time spent executing a single instruction can
very wildly from millisecond to even a second.Therefore releasing GIL after every 100 instructions regardless of the time they take to execute is a poor policy.
In new versions instead of using instruction count as a metric to switch thread , a configurable time interval is used.
The default switch interval is 5 milliseconds.you can get the current switch interval using sys.getswitchinterval().
This can be altered using sys.setswitchinterval()
If a Thread is doing some IO Bound Operations(Ex filesystem access or
network IO)
GIL is release whenever the thread is waiting for some for IO operation to get completed.
Which thread to switch to next?
The interpreter doesn’t have its own scheduler.which thread becomes scheduled at the end of the interval is the operating system’s decision. .
Use threads in python if the individual workers are doing I/O bound operations. If you are trying to scale across multiple cores on a machine either find a good IPC framework for python or pick a different language.
One easy solution to the GIL is the multiprocessing module. It can be used as a drop in replacement to the threading module but uses multiple Interpreter processes instead of threads. Because of this there is a little more overhead than plain threading for simple things but it gives you the advantage of real parallelization if you need it.
It also easily scales to multiple physical machines.
If you need truly large scale parallelization than I would look further but if you just want to scale to all the cores of one computer or a few different ones without all the work that would go into implementing a more comprehensive framework, than this is for you.
Try to remember that the GIL is set to poll around every so often in order to do show the appearance of multiple tasks. This setting can be fine tuned, but I offer the suggestion that there should be work that the threads are doing or lots of context switches are going to cause problems.
I would go so far as to suggest multiple parents on processors and try to keep like jobs on the same core(s).
I've been trying to create a C++ program that embeds multiple python threads. Due to the nature of the program the advantage of multitasking comes from asynchronous I/O; but due to some variables that need to be altered between context switching I need to control the scheduling. I thought that because of python's GIL lock this would be simple enough, but it's turning out not to be: python wants to use POSIX threads rather than software threads, I can't figure out from the documentation what happens if I store the result of PyEval_SaveThread() and don't call PyEval_RestoreThread() in the same function--so presumably I'm not supposed to be doing that, etc.
Is it possible to create a custom scheduler for embedded python threads, or was python basically designed so that it can't be done?
It turns out that using PyEval_SaveThread() and PyEval_RestoreThread() is unnecessary, basically I used coroutines to run the scripts and control the scheduling. In this case from libPCL. However this isn't really much of a solution because if python encounters a syntax error it will segfault if it is in a coroutine, oddly enough even if there is only one python script running in one coroutine this will still happen. But at the very least they don't seem to conflict with each other.
I've recently started experimenting with using Python for web development. So far I've had some success using Apache with mod_wsgi and the Django web framework for Python 2.7. However I have run into some issues with having processes constantly running, updating information and such.
I have written a script I call "daemonManager.py" that can start and stop all or individual python update loops (Should I call them Daemons?). It does that by forking, then loading the module for the specific functions it should run and starting an infinite loop. It saves a PID file in /var/run to keep track of the process. So far so good. The problems I've encountered are:
Now and then one of the processes will just quit. I check ps in the morning and the process is just gone. No errors were logged (I'm using the logging module), and I'm covering every exception I can think of and logging them. Also I don't think these quitting processes has anything to do with my code, because all my processes run completely different code and exit at pretty similar intervals. I could be wrong of course. Is it normal for Python processes to just die after they've run for days/weeks? How should I tackle this problem? Should I write another daemon that periodically checks if the other daemons are still running? What if that daemon stops? I'm at a loss on how to handle this.
How can I programmatically know if a process is still running or not? I'm saving the PID files in /var/run and checking if the PID file is there to determine whether or not the process is running. But if the process just dies of unexpected causes, the PID file will remain. I therefore have to delete these files every time a process crashes (a couple of times per week), which sort of defeats the purpose. I guess I could check if a process is running at the PID in the file, but what if another process has started and was assigned the PID of the dead process? My daemon would think that the process is running fine even if it's long dead. Again I'm at a loss just how to deal with this.
Any useful answer on how to best run infinite Python processes, hopefully also shedding some light on the above problems, I will accept
I'm using Apache 2.2.14 on an Ubuntu machine.
My Python version is 2.7.2
I'll open by stating that this is one way to manage a long running process (LRP) -- not de facto by any stretch.
In my experience, the best possible product comes from concentrating on the specific problem you're dealing with, while delegating supporting tech to other libraries. In this case, I'm referring to the act of backgrounding processes (the art of the double fork), monitoring, and log redirection.
My favorite solution is http://supervisord.org/
Using a system like supervisord, you basically write a conventional python script that performs a task while stuck in an "infinite" loop.
#!/usr/bin/python
import sys
import time
def main_loop():
while 1:
# do your stuff...
time.sleep(0.1)
if __name__ == '__main__':
try:
main_loop()
except KeyboardInterrupt:
print >> sys.stderr, '\nExiting by user request.\n'
sys.exit(0)
Writing your script this way makes it simple and convenient to develop and debug (you can easily start/stop it in a terminal, watching the log output as events unfold). When it comes time to throw into production, you simply define a supervisor config that calls your script (here's the full example for defining a "program", much of which is optional: http://supervisord.org/configuration.html#program-x-section-example).
Supervisor has a bunch of configuration options so I won't enumerate them, but I will say that it specifically solves the problems you describe:
Backgrounding/Daemonizing
PID tracking (can be configured to restart a process should it terminate unexpectedly)
Log normally in your script (stream handler if using logging module rather than printing) but let supervisor redirect to a file for you.
You should consider Python processes as able to run "forever" assuming you don't have any memory leaks in your program, the Python interpreter, or any of the Python libraries / modules that you are using. (Even in the face of memory leaks, you might be able to run forever if you have sufficient swap space on a 64-bit machine. Decades, if not centuries, should be doable. I've had Python processes survive just fine for nearly two years on limited hardware -- before the hardware needed to be moved.)
Ensuring programs restart when they die used to be very simple back when Linux distributions used SysV-style init -- you just add a new line to the /etc/inittab and init(8) would spawn your program at boot and re-spawn it if it dies. (I know of no mechanism to replicate this functionality with the new upstart init-replacement that many distributions are using these days. I'm not saying it is impossible, I just don't know how to do it.)
But even the init(8) mechanism of years gone by wasn't as flexible as some would have liked. The daemontools package by DJB is one example of process control-and-monitoring tools intended to keep daemons living forever. The Linux-HA suite provides another similar tool, though it might provide too much "extra" functionality to be justified for this task. monit is another option.
I assume you are running Unix/Linux but you don't really say. I have no direct advice on your issue. So I don't expect to be the "right" answer to this question. But there is something to explore here.
First, if your daemons are crashing, you should fix that. Only programs with bugs should crash. Perhaps you should launch them under a debugger and see what happens when they crash (if that's possible). Do you have any trace logging in these processes? If not, add them. That might help diagnose your crash.
Second, are your daemons providing services (opening pipes and waiting for requests) or are they performing periodic cleanup? If they are periodic cleanup processes you should use cron to launch them periodically rather then have them run in an infinite loop. Cron processes should be preferred over daemon processes. Similarly, if they are services that open ports and service requests, have you considered making them work with INETD? Again, a single daemon (inetd) should be preferred to a bunch of daemon processes.
Third, saving a PID in a file is not very effective, as you've discovered. Perhaps a shared IPC, like a semaphore, would work better. I don't have any details here though.
Fourth, sometimes I need stuff to run in the context of the website. I use a cron process that calls wget with a maintenance URL. You set a special cookie and include the cookie info in with wget command line. If the special cookie doesn't exist, return 403 rather than performing the maintenance process. The other benefit here is login to the database and other environmental concerns of avoided since the code that serves normal web pages are serving the maintenance process.
Hope that gives you ideas. I think avoiding daemons if you can is the best place to start. If you can run your python within mod_wsgi that saves you having to support multiple "environments". Debugging a process that fails after running for days at a time is just brutal.
I have a Python program that spawns many threads, runs 4 at a time, and each performs an expensive operation. Pseudocode:
for object in list:
t = Thread(target=process, args=(object))
# if fewer than 4 threads are currently running, t.start(). Otherwise, add t to queue
But when the program is run, Activity Monitor in OS X shows that 1 of the 4 logical cores is at 100% and the others are at nearly 0. Obviously I can't force the OS to do anything but I've never had to pay attention to performance in multi-threaded code like this before so I was wondering if I'm just missing or misunderstanding something.
Thanks.
Note that in many cases (and virtually all cases where your "expensive operation" is a calculation implemented in Python), multiple threads will not actually run concurrently due to Python's Global Interpreter Lock (GIL).
The GIL is an interpreter-level lock.
This lock prevents execution of
multiple threads at once in the Python
interpreter. Each thread that wants to
run must wait for the GIL to be
released by the other thread, which
means your multi-threaded Python
application is essentially single
threaded, right? Yes. Not exactly.
Sort of.
CPython uses what’s called “operating
system” threads under the covers,
which is to say each time a request to
make a new thread is made, the
interpreter actually calls into the
operating system’s libraries and
kernel to generate a new thread. This
is the same as Java, for example. So
in memory you really do have multiple
threads and normally the operating
system controls which thread is
scheduled to run. On a multiple
processor machine, this means you
could have many threads spread across
multiple processors, all happily
chugging away doing work.
However, while CPython does use
operating system threads (in theory
allowing multiple threads to execute
within the interpreter
simultaneously), the interpreter also
forces the GIL to be acquired by a
thread before it can access the
interpreter and stack and can modify
Python objects in memory all
willy-nilly. The latter point is why
the GIL exists: The GIL prevents
simultaneous access to Python objects
by multiple threads. But this does not
save you (as illustrated by the Bank
example) from being a lock-sensitive
creature; you don’t get a free ride.
The GIL is there to protect the
interpreters memory, not your sanity.
See the Global Interpreter Lock section of Jesse Noller's post for more details.
To get around this problem, check out Python's multiprocessing module.
multiple processes (with judicious use
of IPC) are[...] a much better
approach to writing apps for multi-CPU
boxes than threads.
-- Guido van Rossum (creator of Python)
Edit based on a comment from #spinkus:
If Python can't run multiple threads simultaneously, then why have threading at all?
Threads can still be very useful in Python when doing simultaneous operations that do not need to modify the interpreter's state. This includes many (most?) long-running function calls that are not in-Python calculations, such as I/O (file access or network requests)) and [calculations on Numpy arrays][6]. These operations release the GIL while waiting for a result, allowing the program to continue executing. Then, once the result is received, the thread must re-acquire the GIL in order to use that result in "Python-land"
Python has a Global Interpreter Lock, which can prevent threads of interpreted code from being processed concurrently.
http://en.wikipedia.org/wiki/Global_Interpreter_Lock
http://wiki.python.org/moin/GlobalInterpreterLock
For ways to get around this, try the multiprocessing module, as advised here:
Does running separate python processes avoid the GIL?
AFAIK, in CPython the Global Interpreter Lock means that there can't be more than one block of Python code being run at any one time. Although this does not really affect anything in a single processor/single-core machine, on a mulitcore machine it means you have effectively only one thread running at any one time - causing all the other core to be idle.
I've been trying to wrap my head around how threads work in Python, and it's hard to find good information on how they operate. I may just be missing a link or something, but it seems like the official documentation isn't very thorough on the subject, and I haven't been able to find a good write-up.
From what I can tell, only one thread can be running at once, and the active thread switches every 10 instructions or so?
Where is there a good explanation, or can you provide one? It would also be very nice to be aware of common problems that you run into while using threads with Python.
Yes, because of the Global Interpreter Lock (GIL) there can only run one thread at a time. Here are some links with some insights about this:
http://www.artima.com/weblogs/viewpost.jsp?thread=214235
http://smoothspan.wordpress.com/2007/09/14/guido-is-right-to-leave-the-gil-in-python-not-for-multicore-but-for-utility-computing/
From the last link an interesting quote:
Let me explain what all that means.
Threads run inside the same virtual
machine, and hence run on the same
physical machine. Processes can run
on the same physical machine or in
another physical machine. If you
architect your application around
threads, you’ve done nothing to access
multiple machines. So, you can scale
to as many cores are on the single
machine (which will be quite a few
over time), but to really reach web
scales, you’ll need to solve the
multiple machine problem anyway.
If you want to use multi core, pyprocessing defines an process based API to do real parallelization. The PEP also includes some interesting benchmarks.
Python's a fairly easy language to thread in, but there are caveats. The biggest thing you need to know about is the Global Interpreter Lock. This allows only one thread to access the interpreter. This means two things: 1) you rarely ever find yourself using a lock statement in python and 2) if you want to take advantage of multi-processor systems, you have to use separate processes. EDIT: I should also point out that you can put some of the code in C/C++ if you want to get around the GIL as well.
Thus, you need to re-consider why you want to use threads. If you want to parallelize your app to take advantage of dual-core architecture, you need to consider breaking your app up into multiple processes.
If you want to improve responsiveness, you should CONSIDER using threads. There are other alternatives though, namely microthreading. There are also some frameworks that you should look into:
stackless python
greenlets
gevent
monocle
Below is a basic threading sample. It will spawn 20 threads; each thread will output its thread number. Run it and observe the order in which they print.
import threading
class Foo (threading.Thread):
def __init__(self,x):
self.__x = x
threading.Thread.__init__(self)
def run (self):
print str(self.__x)
for x in xrange(20):
Foo(x).start()
As you have hinted at Python threads are implemented through time-slicing. This is how they get the "parallel" effect.
In my example my Foo class extends thread, I then implement the run method, which is where the code that you would like to run in a thread goes. To start the thread you call start() on the thread object, which will automatically invoke the run method...
Of course, this is just the very basics. You will eventually want to learn about semaphores, mutexes, and locks for thread synchronization and message passing.
Note: wherever I mention thread i mean specifically threads in python until explicitly stated.
Threads work a little differently in python if you are coming from C/C++ background. In python, Only one thread can be in running state at a given time.This means Threads in python cannot truly leverage the power of multiple processing cores since by design it's not possible for threads to run parallelly on multiple cores.
As the memory management in python is not thread-safe each thread require an exclusive access to data structures in python interpreter.This exclusive access is acquired by a mechanism called GIL ( global interpretr lock ).
Why does python use GIL?
In order to prevent multiple threads from accessing interpreter state simultaneously and corrupting the interpreter state.
The idea is whenever a thread is being executed (even if it's the main thread), a GIL is acquired and after some predefined interval of time the
GIL is released by the current thread and reacquired by some other thread( if any).
Why not simply remove GIL?
It is not that its impossible to remove GIL, its just that in prcoess of doing so we end up putting mutiple locks inside interpreter in order to serialize access, which makes even a single threaded application less performant.
so the cost of removing GIL is paid off by reduced performance of a single threaded application, which is never desired.
So when does thread switching occurs in python?
Thread switch occurs when GIL is released.So when is GIL Released?
There are two scenarios to take into consideration.
If a Thread is doing CPU Bound operations(Ex image processing).
In Older versions of python , Thread switching used to occur after a fixed no of python instructions.It was by default set to 100.It turned out that its not a very good policy to decide when switching should occur since the time spent executing a single instruction can
very wildly from millisecond to even a second.Therefore releasing GIL after every 100 instructions regardless of the time they take to execute is a poor policy.
In new versions instead of using instruction count as a metric to switch thread , a configurable time interval is used.
The default switch interval is 5 milliseconds.you can get the current switch interval using sys.getswitchinterval().
This can be altered using sys.setswitchinterval()
If a Thread is doing some IO Bound Operations(Ex filesystem access or
network IO)
GIL is release whenever the thread is waiting for some for IO operation to get completed.
Which thread to switch to next?
The interpreter doesn’t have its own scheduler.which thread becomes scheduled at the end of the interval is the operating system’s decision. .
Use threads in python if the individual workers are doing I/O bound operations. If you are trying to scale across multiple cores on a machine either find a good IPC framework for python or pick a different language.
One easy solution to the GIL is the multiprocessing module. It can be used as a drop in replacement to the threading module but uses multiple Interpreter processes instead of threads. Because of this there is a little more overhead than plain threading for simple things but it gives you the advantage of real parallelization if you need it.
It also easily scales to multiple physical machines.
If you need truly large scale parallelization than I would look further but if you just want to scale to all the cores of one computer or a few different ones without all the work that would go into implementing a more comprehensive framework, than this is for you.
Try to remember that the GIL is set to poll around every so often in order to do show the appearance of multiple tasks. This setting can be fine tuned, but I offer the suggestion that there should be work that the threads are doing or lots of context switches are going to cause problems.
I would go so far as to suggest multiple parents on processors and try to keep like jobs on the same core(s).