Python GIL

Python GIL (Global Interpreter Lock) is a mutex that prevents multiple threads from executing Python bytecode simultaneously in CPython. Only one thread can hold the GIL at a time, making CPU-bound multithreading ineffective. The GIL exists because CPython's memory management isn't thread-safe; removing it hurts single-threaded performance.

Use multiprocessing for CPU-bound parallelism, threading for I/O-bound concurrency, or async for high-concurrency I/O.

This example demonstrates GIL limitations and workarounds.

# run: python3 gil.py
# description: GIL effects on threading and multiprocessing
 
import threading
import time
import multiprocessing
 
def cpu_bound_task(n):
    """Pure CPU work (no I/O)."""
    total = 0
    for i in range(n):
        total += i
    return total
 
# Threading: GIL prevents parallelism
print("Threading (GIL limited):")
start = time.time()
 
def thread_worker():
    cpu_bound_task(100_000_000)
 
threads = []
for _ in range(2):
    t = threading.Thread(target=thread_worker)
    threads.append(t)
    t.start()
 
for t in threads:
    t.join()
 
thread_elapsed = time.time() - start
print(f"  2 threads: {thread_elapsed:.3f}s")
 
# Multiprocessing: separate processes, no GIL
print("\nMultiprocessing (no GIL):")
start = time.time()
 
processes = []
for _ in range(2):
    p = multiprocessing.Process(target=cpu_bound_task, args=(100_000_000,))
    processes.append(p)
    p.start()
 
for p in processes:
    p.join()
 
process_elapsed = time.time() - start
print(f"  2 processes: {process_elapsed:.3f}s")
 
print(f"\nSpeedup: {thread_elapsed / process_elapsed:.1f}x")
 
# I/O-bound: threading is fine
print("\n" + "="*60)
print("I/O-bound threading (GIL released during I/O):")
 
def io_bound_task():
    """I/O work (releases GIL)."""
    time.sleep(1)  # GIL released during sleep
 
start = time.time()
threads = []
for _ in range(3):
    t = threading.Thread(target=io_bound_task)
    threads.append(t)
    t.start()
 
for t in threads:
    t.join()
 
io_elapsed = time.time() - start
print(f"  3 I/O tasks in threads: {io_elapsed:.3f}s")
print(f"  (should be ~1s, not ~3s)")
 
# GIL release points
print("\n" + "="*60)
print("GIL is released during:")
print("  - I/O operations (read, write, network)")
print("  - time.sleep()")
print("  - C extensions (NumPy, etc.) often release GIL")
print("  - threading.Lock/RLock acquisition")
 
# sys.getswitchinterval() and GIL
import sys
print(f"\nGIL switch interval: {sys.getswitchinterval()}s")
print("  (how often to switch between threads)")

GIL behavior:

• CPython has GIL; PyPy, Jython, IronPython don't
• Released during I/O operations
• Acquired before executing Python bytecode
• Prevents true parallelism in CPU-bound code

For CPU-bound work:

• Use multiprocessing: separate process, separate GIL
• Each process has its own interpreter and GIL
• Trade-off: more overhead, but true parallelism

For I/O-bound work:

• Use threading: efficient, GIL released during I/O
• Great for network requests, file I/O, databases
• Each thread can wait independently

For high-concurrency I/O:

• Use asyncio or trio: even lighter than threads
• Thousands of concurrent tasks with minimal overhead
• No thread switching; cooperative multitasking

Alternatives to threading:

• multiprocessing.Pool: pool of worker processes
• concurrent.futures.ThreadPoolExecutor: thread pool for I/O
• concurrent.futures.ProcessPoolExecutor: process pool for CPU
• asyncio: event loop for I/O

Thread safety without GIL:

• GIL doesn't prevent all race conditions
• Still need locks for shared mutable state
• Even atomic operations (e.g., x += 1) aren't thread-safe without locking

NumPy and GIL:

• NumPy releases GIL for large operations
• Can use threading with NumPy for some parallelism
• True NumPy parallelism better with multiprocessing

GIL internals:

• sys.getswitchinterval(): how often to switch threads (default 5ms)
• sys.setswitchinterval(seconds): adjust (rarely needed)
• GIL contention visible with many threads competing

Common misconceptions:

• “Threading is slow in Python”: not for I/O
• “Multiprocessing always better”: overhead not worth it for I/O
• “Locks solve all concurrency”: still need careful design
• “PyPy has no GIL”: true, but PyPy is not CPython

When to use each:

• CPU-bound: multiprocessing
• I/O-bound (<100 concurrent): threading
• I/O-bound (100+): asyncio
• Mixed: multiprocessing + asyncio inside processes