1 David Beazley http://www.dabeaz.com Originally presented at my "Python Concurrency Workshop", May 14-15, 2009 (Chicago) June 11, 2009 @ chipy
can watch the video of this presentation here: https://www.youtube.com/watch?v=ph374fJqFPE It expands upon the slides and is recommended.
programmers know, Python has a Global Interpreter Lock (GIL) • It imposes various restrictions on threads • Namely, you can't utilize multiple CPUs • Thus, it's a (frankly) tired subject for flamewars about how Python "sucks" (along with tail-call optimization, lambda, whitespace, etc.) 3
of a GIL doesn't bother me • I don't have strong feelings about it either way • Bias : For parallel computing involving heavy CPU processing, I much prefer message passing and cooperating processes to thread programming (of course, it depends on the problem) • However, the GIL has some pretty surprising behavior on multicore that interests me 4
Consider this trivial CPU-bound function def count(n): while n > 0: n -= 1 5 • Run it twice in series count(100000000) count(100000000) • Now, run it in parallel in two threads t1 = Thread(target=count,args=(100000000,)) t1.start() t2 = Thread(target=count,args=(100000000,)) t2.start() t1.join(); t2.join()
do I get these performance results on my Dual-Core MacBook? 6 Sequential : 24.6s Threaded : 45.5s (1.8X slower!) • And if I disable one of the CPU cores, why does the threaded performance get better? Threaded : 38.0s • Think about that for a minute... Bloody hell!
like unexplained mysteries or magic • As part of a workshop I ran in May, I went digging into the GIL implementation to see if I could figure out exactly why I was getting those performance results • An exploration that went all the way from Python scripts to the C source code of the pthreads library (yes, I probably need to go outside more often) • So, let's just jump into it... 7
• Python threads are real system threads • POSIX threads (pthreads) • Windows threads • Fully managed by the host operating system • All scheduling/thread switching • Represent threaded execution of the Python interpreter process (written in C) 8
There's not a whole lot going on... • Here's what happens on thread creation • Python creates a small data structure containing some interpreter state • A new thread (pthread) is launched • The thread calls PyEval_CallObject • Last step is just a C function call that runs whatever Python callable was specified 10
thread has its own interpreter specific data structure (PyThreadState) • Current stack frame (for Python code) • Current recursion depth • Thread ID • Some per-thread exception information • Optional tracing/profiling/debugging hooks • It's a small C structure (<100 bytes) 11
interpreter has a global variable that simply points to the ThreadState structure of the currently running thread 13 /* Python/pystate.c */ ... PyThreadState *_PyThreadState_Current = NULL; • Operations in the interpreter implicitly depend this variable to know what thread they're currently working with
Here's the rub... • Only one Python thread can execute in the interpreter at once • There is a "global interpreter lock" that carefully controls thread execution • The GIL ensures that sure each thread gets exclusive access to the interpreter internals when it's running (and that call-outs to C extensions play nice) 14
simple : threads hold the GIL when running • However, they release it when blocking for I/O 15 I/O I/O I/O release acquire release acquire acquire release • So, any time a thread is forced to wait, other "ready" threads get their chance to run • Basically a kind of "cooperative" multitasking run run run run acquire
To deal with CPU-bound threads that never perform any I/O, the interpreter periodically performs a "check" • By default, every 100 interpreter "ticks" 16 CPU Bound Thread Run 100 ticks Run 100 ticks Run 100 ticks check check check • sys.setcheckinterval() changes the setting
The check interval is a global counter that is completely independent of thread scheduling 17 Main Thread 100 ticks check check check 100 ticks 100 ticks Thread 2 Thread 3 Thread 4 100 ticks • A "check" is simply made every 100 "ticks"
What happens during the periodic check? • In the main thread only, signal handlers will execute if there are any pending signals (more shortly) • Release and reacquire the GIL • That last bullet describes how multiple CPU- bound threads get to run (by briefly releasing the GIL, other threads get a chance to run). 18
Python/ceval.c */ ... if (--_Py_Ticker < 0) { ... _Py_Ticker = _Py_CheckInterval; ... if (things_to_do) { if (Py_MakePendingCalls() < 0) { ... } } if (interpreter_lock) { /* Give another thread a chance */ ... PyThread_release_lock(interpreter_lock); /* Other threads may run now */ PyThread_acquire_lock(interpreter_lock, 1); ... }
Let's briefly talk about Ctrl-C • A very common problem encountered with Python thread programming is that threaded programs can no longer be killed with the keyboard interrupt • It is EXTREMELY ANNOYING (you have to use kill -9 in a separate window) • Ever wonder why it doesn't work? 22
a signal arrives, the interpreter runs the "check" after every tick until the main thread runs 23 Main Thread 100 ticks check Thread 2 Thread 3 • Since signal handlers can only run in the main thread, the interpreter quickly acquires/releases the GIL after every tick until it gets scheduled check SIGNAL check check 1 tick 100 ticks check check check signal handler
does not have a thread scheduler • There is no notion of thread priorities, preemption, round-robin scheduling, etc. • All thread scheduling is left to the host operating system (e.g., Linux, Windows, etc.) • This is partly why signals get so weird (the interpreter has no control over scheduling so it just attempts to thread switch as fast as possible with the hope that main will run) 24
reason Ctrl-C doesn't work with threaded programs is that the main thread is often blocked on an uninterruptible thread-join or lock • Since it's blocked, it never gets scheduled to run any kind of signal handler for it • And as an extra little bonus, the interpreter is left in a state where it tries to thread-switch after every tick (so not only can you not interrupt your program, it runs slow as hell!) 25
GIL is not a simple mutex lock • The implementation (Unix) is either... • A POSIX unnamed semaphore • Or a pthreads condition variable • All interpreter locking is based on signaling • To acquire the GIL, check if it's free. If not, go to sleep and wait for a signal • To release the GIL, free it and signal 26
switching is far more subtle than most programmers realize 27 Thread 1 100 ticks check check check 100 ticks Thread 2 ... Operating System signal signal SUSPENDED Thread Context Switch check • The lag between signaling and execution may be significant (depends on the OS) SUSPENDED signal signal check signal
OS is just going to schedule whichever thread has the highest execution "priority" • CPU-bound : low priority • I/O bound : high priority • If a signal is sent to a thread with low priority and the CPUs are busy with higher priority tasks, it won't run until some later point • Read an OS textbook for details 28
we saw earlier, CPU-bound threads have horrible performance properties • Far worse than simple sequential execution • 24.6 seconds (sequential) • 45.5 seconds (2 threads) • A big question : Why? • What is the source of that overhead? 29
thread signaling is the source of that • After every 100 ticks, the interpreter • Locks a mutex • Signals on a condition variable/semaphore where another thread is always waiting • Because another thread is waiting, extra pthreads processing and system calls get triggered to deliver the signal 30
Sequential Execution (OS-X, 1 CPU) • 736 Unix system calls • 117 Mach System Calls • Two CPU-bound threads (OS-X, 1 CPU) • 1149 Unix system calls • ~ 3.3 Million Mach System Calls • Yow! Look at that last figure. 31
The penalty gets far worse on multiple cores • Two CPU-bound threads (OS-X, 1 CPU) • 1149 Unix system calls • ~3.3 Million Mach System Calls • Two CPU-bound threads (OS-X, 2 CPUs) • 1149 Unix system calls • ~9.5 Million Mach System calls 32
did some instrumentation of the Python interpreter to look into this a little deeper • Recorded a real-time trace of all GIL acquisitions, releases, conflicts, retries, etc. • Trying to get an idea of what the interpreter is doing, what different threads are doing, interactions between threads and the GIL, and the overall sequencing of events 33
With multiple cores, CPU-bound threads get scheduled simultaneously (on different cores) and then have a GIL battle 35 Thread 1 (CPU 1) Thread 2 (CPU 2) Release GIL signal Acquire GIL Wake Acquire GIL (fails) Release GIL Acquire GIL signal Wake Acquire GIL (fails) run run run • The waiting thread (T2) may make 100s of failed GIL acquisitions before any success
36 t2 100 5392 ENTRY t2 100 5392 ACQUIRE t2 100 5393 RELEASE t1 100 5393 ACQUIRE t2 100 5393 ENTRY t2 27 5393 BUSY t1 100 5394 RELEASE t1 100 5394 ENTRY t1 100 5394 ACQUIRE t2 74 5394 RETRY t1 100 5395 RELEASE t1 100 5395 ENTRY t1 100 5395 ACQUIRE t2 83 5395 RETRY t1 100 5396 RELEASE t1 100 5396 ENTRY t1 100 5396 ACQUIRE t2 80 5396 RETRY t1 100 5397 RELEASE t1 100 5397 ENTRY t1 100 5397 ACQUIRE t2 79 5397 RETRY ... A thread switch t2 tries to keep running, but immediately has to block because t1 acquired the GIL signal signal signal signal Here, the GIL battle begins. Every RELEASE of the GIL signals t2. Since there are two cores, the OS schedules t2, but leaves t1 running on the other core. Since t1 is left running, it immediately reacquires the GIL before t2 can get to it (so, t2 wakes up, finds the GIL is in use, and blocks again)
What's happening here is that you're seeing a battle between two competing (and ultimately incompatible) goals • Python - only wants to run single- threaded, but doesn't want anything to do with thread scheduling (up to OS) • OS - "Oooh. Multiple cores." Freely schedules processes/threads to take advantage of as many cores as possible 37
Even 1 CPU-bound thread causes problems • It degrades response time of I/O-bound threads 38 Thread 1 (CPU 1) Thread 2 (CPU 2) Network Packet Acquire GIL (fails) run Acquire GIL (fails) Acquire GIL (fails) Acquire GIL (success) signal signal signal signal run sleep
last scenario is a bizarre sort of "priority inversion" problem • A CPU-bound thread (low priority) is blocking the execution of an I/O-bound thread (high priority) • It occurs because the I/O thread can't wake up fast enough to acquire the GIL before the CPU-bound thread reacquires it • And it only happens on multicore... 41
as I can tell, the Python GIL implementation has not changed much (if at all) in the last 10 years • The GIL code in Python 1.5.2 looks almost identical to the code in Python 3.0 • I don't know whether it's even been studied all that much (especially on multicore) • There is more interest in removing the GIL than simply changing the GIL 42
this deserves further study • There is a pretty severe performance penalty for using threads on multicore • The priority inversion for I/O-bound processing is somewhat disturbing • Probably worth fixing--especially if the GIL is going to stick around 43
in the hell would you fix this? • I have some vague ideas, but they're all "hard" • Require Python to do its own form of thread scheduling (or at least cooperate with the OS) • Would involve a non-trivial interaction between the interpreter implementation, the operating system scheduler, the thread library, and C extension modules (egad!) 44
• If you could fix it, it would make thread execution (even with the GIL) more predictable and less resource intensive • Might improve performance/responsiveness of applications that have a mix of CPU and I/O- bound processing • Probably good for libraries that use threads in the background (e.g., multiprocessing) • Might be able to do it without rewriting the whole interpreter. 45
I'm not actively working on any patches or code related to this presentation • However, the problem interests me • If it interests you and you want to hack on any of my code or examples, send me an email ([email protected]) 46
dabeaz
oharu121
utgwkk
naramomi7
naruyoshimi
tosuri13
kazumura
kashewnuts
konokenj
magavel
hiroyukimori
ivargrimstad
gotalab555
techseoconnect
ammeep
ryanjones
akademiya2063
joshbly
orderedlist
elsirapls
badams
inesmontani
hatefulcrawdad
skipperchong
caitiem20
