Skip to content

Latest commit

 

History

History
133 lines (96 loc) · 8.47 KB

README.md

File metadata and controls

133 lines (96 loc) · 8.47 KB

GIL Knocker

pip install gilknocker

Code Style CI PyPI PyPI - Wheel Downloads

When you thought the GIL was available, and you find yourself suspecting it might be spending time with another.

You probably want py-spy, however if you're looking for a quick-and-dirty way to slip in a GIL contention metric within a specific chunk of code, this might help you.

How?

Unfortunately, there doesn't appear to be any explicit C-API for checking how busy the GIL is. PyGILState_Check won't really work, that's limited to the current thread. PyInterpreterState is an opaque struct, and the PyRuntimeState and other goodies are private in CPython.

So, in ~200 lines of Rusty code, I've conjured up a basic metric that seems to align with what is reported by py-spy when running the same test case. This works by spawning a thread which, at regular intervals, re-acquires the GIL and checks how long it took for the GIL to answer.

Note, the polling_interval_micros, sampling_interval_micros, and sleeping_interval_micros are configurable.

  • polling_interval_micros

    • How frequently to re-acquire the GIL and measure how long it took to acquire. The more frequent, the more likely the contention metric will represent accurate GIL contention. A good value for this is 1-1000.
  • sampling_interval_micros

    • How long to run the polling routine. If this is 1ms, then for 1ms it will try to re-acquire the GIL at polling_interval_micros frequency. Defaults to 10x polling_interval_micros
  • sleeping_interval_micros

    • How long to sleep between sampling routines. Defaults to 100x polling_interval_micros

Use

Look at the tests

from gilknocker import KnockKnock

# These two are equivalent. 
knocker = KnockKnock(1_000) 
knocker = KnockKnock(
  polling_interval_micros=1_000, 
  sampling_interval_micros=10_000, 
  sleeping_interval_micros=100_000
)
knocker.start()

... smart code here ...

knocker.contention_metric  # float between 0-1 indicating roughly how busy the GIL was.
knocker.reset_contention_metric()  # reset timers and meteric calculation

... some more smart code ...

knocker.stop()
knocker.stop()  # Idempodent stopping behavior

knocker.contention_metric  # will stay the same after `stop()` is called.

knocker.is_running  # If you're ever in doubt

How will this impact my program?

Short answer, it depends, but probably not much. As stated above, the more frequent the polling and sampling interval, the more likely non-GIL bound programs will be affected, since there is more room for contention. In GIL heavy programs, the monitoring thread will spend most of its time simply waiting for a lock. This is demonstrated in the benchmarks testing.

In general, it appears that polling_interval_micros=1_000 is a good tradeoff in terms of accurate GIL contention metric and the resulting sampling_interval_micros=10_000 (defaults to 10x polling interval) is high enough to relax performance impact a bit when combined with sleeping_interval_micros=100_000 (defaults to 100x polling interval); but feel free to play with these to conform to your needs.

Below is a summary of benchmarking two different functions, one which uses the GIL, and one which releases it. For interval=None this means no polling was used, effectively just running the function without gilknocker. Otherwise, the interval represents the value passed to KnockKnock(polling_interval_micros=interval)

python -m pytest -v --benchmark-only benchmarks/ --benchmark-histogram

------------------------------------------------------------------------------------ benchmark: 18 tests -------------------------------------------------------------------------------------
Name (time in s)                       Min               Max              Mean            StdDev            Median               IQR            Outliers     OPS            Rounds  Iterations
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
test_bench[a_little_gil-100000]     1.5368 (2.07)     1.6596 (2.23)     1.5968 (2.14)     0.0476 (130.12)   1.5943 (2.14)     0.0719 (140.14)        2;0  0.6262 (0.47)          5           1
test_bench[a_little_gil-10000]      1.5321 (2.06)     1.5989 (2.14)     1.5618 (2.09)     0.0289 (78.95)    1.5610 (2.09)     0.0510 (99.52)         2;0  0.6403 (0.48)          5           1
test_bench[a_little_gil-1000]       1.5246 (2.05)     1.5298 (2.05)     1.5271 (2.05)     0.0019 (5.12)     1.5269 (2.05)     0.0021 (4.00)          2;0  0.6549 (0.49)          5           1
test_bench[a_little_gil-100]        1.5505 (2.09)     1.5543 (2.08)     1.5528 (2.08)     0.0014 (3.96)     1.5533 (2.08)     0.0018 (3.60)          2;0  0.6440 (0.48)          5           1
test_bench[a_little_gil-10]         1.5863 (2.13)     1.6074 (2.16)     1.5928 (2.14)     0.0088 (23.94)    1.5896 (2.13)     0.0111 (21.60)         1;0  0.6278 (0.47)          5           1
test_bench[a_little_gil-None]       1.5043 (2.02)     1.5067 (2.02)     1.5051 (2.02)     0.0011 (2.95)     1.5044 (2.02)     0.0016 (3.17)          1;0  0.6644 (0.50)          5           1
test_bench[a_lotta_gil-100000]      0.7450 (1.00)     0.7458 (1.0)      0.7455 (1.0)      0.0004 (1.0)      0.7457 (1.0)      0.0005 (1.0)           1;0  1.3413 (1.0)           5           1
test_bench[a_lotta_gil-10000]       0.7471 (1.00)     0.8104 (1.09)     0.7601 (1.02)     0.0281 (76.94)    0.7472 (1.00)     0.0168 (32.82)         1;1  1.3156 (0.98)          5           1
test_bench[a_lotta_gil-1000]        0.7436 (1.0)      0.7472 (1.00)     0.7463 (1.00)     0.0015 (4.11)     0.7470 (1.00)     0.0013 (2.54)          1;1  1.3400 (1.00)          5           1
test_bench[a_lotta_gil-100]         0.7558 (1.02)     0.7680 (1.03)     0.7640 (1.02)     0.0050 (13.56)    0.7644 (1.03)     0.0061 (11.97)         1;0  1.3089 (0.98)          5           1
test_bench[a_lotta_gil-10]          0.7542 (1.01)     0.7734 (1.04)     0.7649 (1.03)     0.0084 (23.05)    0.7669 (1.03)     0.0151 (29.45)         2;0  1.3074 (0.97)          5           1
test_bench[a_lotta_gil-None]        0.7437 (1.00)     0.8490 (1.14)     0.8006 (1.07)     0.0501 (137.15)   0.8074 (1.08)     0.0969 (189.03)        1;0  1.2490 (0.93)          5           1
test_bench[some_gil-100000]         1.4114 (1.90)     1.4131 (1.89)     1.4122 (1.89)     0.0007 (1.81)     1.4121 (1.89)     0.0010 (2.00)          2;0  0.7081 (0.53)          5           1
test_bench[some_gil-10000]          1.4115 (1.90)     1.4258 (1.91)     1.4167 (1.90)     0.0059 (16.03)    1.4141 (1.90)     0.0083 (16.19)         1;0  0.7058 (0.53)          5           1
test_bench[some_gil-1000]           1.4169 (1.91)     1.5793 (2.12)     1.4618 (1.96)     0.0690 (188.82)   1.4232 (1.91)     0.0769 (150.04)        1;0  0.6841 (0.51)          5           1
test_bench[some_gil-100]            1.4468 (1.95)     1.6261 (2.18)     1.5701 (2.11)     0.0752 (205.83)   1.5998 (2.15)     0.1004 (195.70)        1;0  0.6369 (0.47)          5           1
test_bench[some_gil-10]             1.5269 (2.05)     1.9894 (2.67)     1.7037 (2.29)     0.1895 (518.49)   1.7301 (2.32)     0.2692 (524.96)        1;0  0.5870 (0.44)          5           1
test_bench[some_gil-None]           1.4115 (1.90)     1.4267 (1.91)     1.4155 (1.90)     0.0063 (17.33)    1.4136 (1.90)     0.0053 (10.24)         1;1  0.7065 (0.53)          5           1
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------


License

Unlicense or MIT, at your discretion.