Performance

[taldcroft]

Description

This is part of a set of performance changes to kadi, chandra_aca, ska_sun, Quaternion, and Chandra.Maneuver to make kadi command states faster. The changes are independent.

This PR includes changes to:

• Use numba - including a numba-ized version of the radec2eci from Ska.quatutil
• Add support for sun_ra, sun_dec kwargs to the pitch and off_nominal_roll functions to allow upstream applications (with values for those) to run faster.
• Use numba caching

Interface impacts

The numba cache uses the NUMBA_CACHE_DIR already configured in ska after sot/ska_helpers#29 .
The pitch and off_nominal_roll functions now support sun_ra and sun_dec kwargs.

Testing

Unit tests

• Mac (as of 6e2d783)

Independent check of unit tests by Jean

• Linux

Functional tests

Also checking numba caching behavior in ska3-flight 2023.3. This demonstrates the creation of numba cache files in the expected user directory.

(ska3) ➜  ska_sun git:(performance) rm -rf ~/.ska3             
(ska3) ➜  ska_sun git:(performance) pytest ska_sun
============================= test session starts ==============================
platform darwin -- Python 3.10.8, pytest-7.2.1, pluggy-1.0.0
rootdir: /Users/aldcroft/git, configfile: pytest.ini
plugins: timeout-2.1.0, anyio-3.6.2
collected 12 items                                                             

ska_sun/tests/test_sun.py ............                                   [100%]

============================== 12 passed in 6.95s ==============================
(ska3) ➜  ska_sun git:(performance) find ~/.ska3     
/Users/aldcroft/.ska3
/Users/aldcroft/.ska3/cache
/Users/aldcroft/.ska3/cache/numba
/Users/aldcroft/.ska3/cache/numba/ska_sun_28343af752654d47d93a0d0bcdf944cc27778384
/Users/aldcroft/.ska3/cache/numba/ska_sun_28343af752654d47d93a0d0bcdf944cc27778384/sun.position_at_jd-135.py310.1.nbc
/Users/aldcroft/.ska3/cache/numba/ska_sun_28343af752654d47d93a0d0bcdf944cc27778384/sun._nominal_roll-307.py310.nbi
/Users/aldcroft/.ska3/cache/numba/ska_sun_28343af752654d47d93a0d0bcdf944cc27778384/sun.sph_dist-213.py310.nbi
/Users/aldcroft/.ska3/cache/numba/ska_sun_28343af752654d47d93a0d0bcdf944cc27778384/sun._radec2eci-266.py310.nbi
/Users/aldcroft/.ska3/cache/numba/ska_sun_28343af752654d47d93a0d0bcdf944cc27778384/sun._nominal_roll-307.py310.1.nbc
/Users/aldcroft/.ska3/cache/numba/ska_sun_28343af752654d47d93a0d0bcdf944cc27778384/sun.position_at_jd-135.py310.nbi
/Users/aldcroft/.ska3/cache/numba/ska_sun_28343af752654d47d93a0d0bcdf944cc27778384/sun._radec2eci-266.py310.1.nbc
/Users/aldcroft/.ska3/cache/numba/ska_sun_28343af752654d47d93a0d0bcdf944cc27778384/sun.sph_dist-213.py310.1.nbc

With regard to performance, on linux (HEAD fido) a comparison with ska_sun 3.10.1 shows the expected result that the numba caching can be slower on the first run of a function (compared to flight / 3.10.1) and then has advantages.

flight:

In [1]: import ska_sun

In [2]: time = '2023:001'

In [3]: %time ska_sun.nominal_roll(10, 20, time=time)
CPU times: user 4.83 ms, sys: 0 ns, total: 4.83 ms
Wall time: 6.41 ms
Out[3]: 292.1368992773556

# This is the version without performance improvements but it still shows some speedup when called again (NFS caching?)
In [4]: %time ska_sun.nominal_roll(10, 20, time=time)
CPU times: user 883 µs, sys: 0 ns, total: 883 µs. 
Wall time: 812 µs                      
Out[4]: 292.1368992773556

In [6]: %timeit ska_sun.nominal_roll(10, 20, time=time)
417 µs ± 118 ns per loop (mean ± std. dev. of 7 runs, 1,000 loops each)

In [7]: ska_sun.__version__
Out[7]: '3.10.1'

This PR

In [1]: import ska_sun

In [2]: time = '2023:001'

In [3]: %time ska_sun.nominal_roll(10, 20, time=time)
CPU times: user 1.92 s, sys: 43.8 ms, total: 1.96 s
Wall time: 2.31 s
Out[3]: -67.86310072264433

In [4]: %time ska_sun.nominal_roll(10, 20, time=time)
CPU times: user 123 µs, sys: 10 µs, total: 133 µs
Wall time: 136 µs
Out[4]: -67.86310072264433

In [5]: %timeit ska_sun.nominal_roll(10, 20, time=time)
35.3 µs ± 1.86 µs per loop (mean ± std. dev. of 7 runs, 10,000 loops each)

And when set with debug printing on the numba caching, the caching saves seem reasonable

ska3-jeanconn-fido> export NUMBA_DEBUG_CACHE=1
ska3-jeanconn-fido> ipython
Python 3.10.8 | packaged by conda-forge | (main, Nov 22 2022, 08:26:04) [GCC 10.4.0]
Type 'copyright', 'credits' or 'license' for more information
IPython 8.8.0 -- An enhanced Interactive Python. Type '?' for help.

In [1]: import ska_sun

In [2]: time = '2023:001'

In [3]: %time ska_sun.nominal_roll(10, 20, time=time)
[cache] index saved to '/home/jeanconn/.ska3/cache/numba/ska_sun_cdeca1941b46e08ad13896c2df292287d50c752d/sun.position_at_jd-135.py310.nbi'
[cache] data saved to '/home/jeanconn/.ska3/cache/numba/ska_sun_cdeca1941b46e08ad13896c2df292287d50c752d/sun.position_at_jd-135.py310.1.nbc'
[cache] index saved to '/home/jeanconn/.ska3/cache/numba/ska_sun_cdeca1941b46e08ad13896c2df292287d50c752d/sun._radec2eci-266.py310.nbi'
[cache] data saved to '/home/jeanconn/.ska3/cache/numba/ska_sun_cdeca1941b46e08ad13896c2df292287d50c752d/sun._radec2eci-266.py310.1.nbc'
[cache] index loaded from '/home/jeanconn/.ska3/cache/numba/ska_sun_cdeca1941b46e08ad13896c2df292287d50c752d/sun._radec2eci-266.py310.nbi'
[cache] index loaded from '/home/jeanconn/.ska3/cache/numba/ska_sun_cdeca1941b46e08ad13896c2df292287d50c752d/sun._radec2eci-266.py310.nbi'
[cache] index saved to '/home/jeanconn/.ska3/cache/numba/ska_sun_cdeca1941b46e08ad13896c2df292287d50c752d/sun._radec2eci-266.py310.nbi'
[cache] data saved to '/home/jeanconn/.ska3/cache/numba/ska_sun_cdeca1941b46e08ad13896c2df292287d50c752d/sun._radec2eci-266.py310.2.nbc'
[cache] index saved to '/home/jeanconn/.ska3/cache/numba/ska_sun_cdeca1941b46e08ad13896c2df292287d50c752d/sun._nominal_roll-307.py310.nbi'
[cache] data saved to '/home/jeanconn/.ska3/cache/numba/ska_sun_cdeca1941b46e08ad13896c2df292287d50c752d/sun._nominal_roll-307.py310.1.nbc'
CPU times: user 1.91 s, sys: 40.1 ms, total: 1.95 s
Wall time: 2.3 s
Out[3]: -67.86310072264433

In [4]: %time ska_sun.nominal_roll(10, 20, time=time)
CPU times: user 169 µs, sys: 16 µs, total: 185 µs
Wall time: 189 µs
Out[4]: -67.86310072264433

In [5]: %time ska_sun.nominal_roll(10, 20, time=time)
CPU times: user 159 µs, sys: 16 µs, total: 175 µs
Wall time: 178 µs
Out[5]: -67.86310072264433

In [6]: %time ska_sun.pitch(10, 20, time=time)
[cache] index saved to '/home/jeanconn/.ska3/cache/numba/ska_sun_cdeca1941b46e08ad13896c2df292287d50c752d/sun.sph_dist-213.py310.nbi'
[cache] data saved to '/home/jeanconn/.ska3/cache/numba/ska_sun_cdeca1941b46e08ad13896c2df292287d50c752d/sun.sph_dist-213.py310.1.nbc'
CPU times: user 84.7 ms, sys: 136 µs, total: 84.8 ms
Wall time: 93.8 ms
Out[6]: 96.6598242368135

In [7]: %time ska_sun.pitch(10, 20, time=time)
CPU times: user 119 µs, sys: 11 µs, total: 130 µs
Wall time: 133 µs
Out[7]: 96.6598242368135

And as mentioned, this PR provides the option to supply sun_ra and sun_dec from the calling code which can provide significant savings.

In [8]: %timeit ska_sun.nominal_roll(10, 20, sun_ra=sun_ra, sun_dec=sun_dec)
1.21 µs ± 0.62 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)

[jeanconn]

I think this would need rebase for tests to pass at this point. @taldcroft Do you want me to go ahead and do that to test this independently from #25?

[jeanconn]

Rebased and force-pushed.

[taldcroft]

@jeanconn - I re-ran unit tests and added some functional test information in the description.