Freeze/link all stdlib modules imported during startup.

[ericsnowcurrently]

During runtime startup we import a bunch of modules, particularly if the site module is used (without -S). Any of these that aren't frozen/linked incur a substantial IO overhead. We already freeze importlib and link a number of builtin modules. The idea here is that we freeze/link the rest.

To do:

• re-run a large number of times times to stabilize the results
• make sure I've frozen all the relevant modules
• make sure all startup extension modules are built-in
• re-run the comparison on the benchmarking machine (I'm comfortable with the results I'm seeing locally.)
• create a BPO issue and an upstream PR
• move generated .h files to a new "Python/frozen" directory (like we do for argument clinic)
• detect if dev build without accessing filesystem
• add -X frozen-modules=[on|off] (defaulting to "off")
• sort out Windows
  • Exception in HTMLParser for special JavaScript code python/cpython#45188
• stop tracking frozen .h files in the repo (auto-regen them in the Windows build process)
  • bpo-45020: Drop the frozen .h files from the repo. python/cpython#28375
  • bpo-45020: Drop the frozen .h files from the repo. python/cpython#28392
• freeze os, site, and codecs
  • blocked by bpo-45186 (or maybe just by removing the .h files)
  • bpo-45020: Freeze os, site, and codecs. python/cpython#28398
• remember the stdlib dir in getpath.c (bpo-45211)
• set __file__ (and __path__) on frozen [stdlib] modules
  • see bpo-21736 and bpo-45020
  • update FrozenImporter
    • bpo-21736: Set __file__ on frozen stdlib modules. python/cpython#28656
  • fix up frozen modules imported before importlib._bootstrap._install() is called
    • bpo-21736: Set __file__ on frozen stdlib modules. python/cpython#28656
  • set co_filename during unmarshal of frozen modules (for tracebacks, etc.) (bpo-45652)
• (partially?) freeze encodings package (bpo-45653)
  • blocked by noise added to "make" output (see comment)
  • ? only include common encodings? (partial pkg would need proper __path__ bpo-21736)
  • bpo-45653: partially freeze encodings package python/cpython#29331
• default to "on"
  • blocked by bpo-45211
  • "off" if actually running out of the source tree
    • bpo-45020: Default to using frozen modules unless running from source tree. python/cpython#28940
  • ? "off" if Py_DEBUG
    • bpo-45020: Default to using frozen modules only on non-debug builds. python/cpython#28590
  • ? always "on" if not a PGO build, even if running out of the source tree
  • ? emit a warning if running out of the source tree
• change -X frozen_modules to -X frozen_stdlib and "use_frozen" in import.c (see gh-28633)
  • do not apply "-X frozen_modules" to custom frozen modules (bpo-45396)
    • bpo-45396: Always import custom frozen modules. python/cpython#28776
    • bpo-45395: Make custom frozen modules additions instead of replacements. python/cpython#28778
  • use frozen stdlib modules even if built with custom frozen modules (bpo-45395)
    • bpo-45396: Always import custom frozen modules. python/cpython#28776
    • bpo-45395: Make custom frozen modules additions instead of replacements. python/cpython#28778
• default to "off" on Windows when running in source tree (bpo-45651)
• do no freeze os.path (bpo-45272)
  • bpo-45272: os.path should not be a frozen module python/cpython#29329
• make sure Tools/freeze is still working properly (bpo-45096, bpo-45395, ...)
• add a devguide entry about frozen stdlib modules (python/devguide#756)
• double-check performance improvement
• test test test #117

after:

• freeze stdlib modules imported during ./python -m ... (bpo-45654)
• freeze argparse (and dependencies) (bpo-45660)
• ? freeze other commonly used modules (bpo-45661)
• ? do not freeze platform-specific modules when not applicable (bpo-45273)
• ? move away from a linear search in find_frozen() (see bpo-45213)
  • for now a comment noting the concern (and a reference to the issue) is sufficient
• improve errors related to frozen modules (bpo-45379)
  • bpo-45379: raise import error on FROZEN_BAD_NAME python/cpython#29180
• ? add a build option to specify the default for "-X frozen_modules"
• ? record any default derived from the build options (and set of PyConfig
• Makefile cleanup
  • ? use a common make recipe for freezing modules
    • less generated code in Makefile.pre.in and simplifies Tools/scripts/freeze_modules.py a little
    • bpo-45020: Use a common recipe for freezing modules. python/cpython#28583
  • ? do not rebuild frozen.o if no frozen modules were actually rebuilt
    • bpo-45020: Make "make all" output less noisy. python/cpython#28554
  • ? make "make all" less noisy (dot per file
    • bpo-45020: Make "make all" output less noisy. python/cpython#28554
• use _imp.frozen_module_names() in Tools/scripts/generate_stdlib_module_names.py (bpo-45189)
• update Tools/freeze to depend on Toos/scripts/freeze_modules.py? (bpo-45096)
• ? support more of the FileLoader API on the frozen loader?
  • add FrozenImporter.get_filename() (bpo-45659)
  • implement FrozenImporter.get_source() (bpo-45658)

investigate:

• create a flamegraph for "after"
• figure out where all the unexpected syscalls are coming from (e.g. 25 stats, down from 80)
• get a clear picture of where the other syscalls come from (e.g. "rt_sigaction", "mmap")

related:

• inconsistent marshal output (https://bugs.python.org/issue45186 and https://bugs.python.org/issue34093)
  • gh-78274: Produce consistent output from marshal.dumps(). python/cpython#28379
• missing documentation for ModuleSpec.loader_state (https://bugs.python.org/issue45364)

---

Candidate solutions to the editing-stdlib-.py-files problem, from bpo-45020:

• use a command-line flag to opt-out of frozen modules?
• use a build flag to opt out (e.g. a configure flag or a new Py_NO_FROZEN or even Py_DEBUG)?
• ignore frozen modules if it's a dev build?
• (note: importlib._bootstrap and importlib._bootstrap_external must always be frozen, regardless of a flag)
• accommodate users of an installed Python that sometimes edit stdlib modules while debugging code?
• always emit a warning if a frozen module is ignored (in favor of the source module)?

[gvanrossum]

It would be nice to be able to put numbers on how substantial the I/O overhead really is.

[ericsnowcurrently]

I've found the following .py modules get imported during startup (some only without -S) and frozen them:

without site (python -S):

• abc
• codecs
• encodings.* # we need all of them since the one we use is determined at runtime
• io

with site:

• _collections_abc
• _sitebuiltins
• genericpath
• os
• posixpath (or ntpath)
• site
• stat

See my branch: https://github.com/ericsnowcurrently/cpython/tree/frozen-modules-cleanup

before adding stdlib modules

# repeated 100 times and averaged
$ strace -c ./python -c pass
% time     seconds  usecs/call     calls    errors syscall
------ ----------- ----------- --------- --------- ----------------
 20.43    0.001504          22        68         0 rt_sigaction
 17.32    0.001275          15        80         8 stat
  9.62    0.000708          15        47         0 fstat
  7.46    0.000549          15        35         0 read
  7.35    0.000541          17        31         2 openat
  6.66    0.000490          15        32         0 close
  6.20    0.000456          15        30         3 lseek
  5.62    0.000414          19        21         0 mmap
  4.21    0.000310          16        19        17 ioctl
  3.77    0.000277          23        12         0 mprotect
  2.46    0.000181          15        12         0 getdents
  1.70    0.000125          15         8         0 brk
  1.42    0.000105          14         7         7 access
  1.02    0.000075          25         3         0 fcntl
  0.60    0.000044          14         3         0 dup
  0.46    0.000034          33         1         0 getrandom
  0.41    0.000030          30         1         0 gettid
  0.39    0.000029           9         3         0 munmap
  0.31    0.000023          23         1         0 futex
  0.30    0.000022          21         1         0 arch_prctl
  0.29    0.000022          21         1         0 prlimit64
  0.29    0.000021          21         1         0 set_robust_list
  0.29    0.000021          21         1         0 set_tid_address
  0.29    0.000021          21         1         0 rt_sigprocmask
  0.23    0.000017          16         1         0 sysinfo
  0.19    0.000014          14         1         0 getcwd
  0.19    0.000014          13         1         1 readlink
  0.16    0.000012          11         1         0 getgid
  0.15    0.000011          10         1         0 getegid
  0.12    0.000009           8         1         0 geteuid
  0.11    0.000008           8         1         0 getuid
  0.00    0.000000           0         1         0 execve
------ ----------- ----------- --------- --------- ----------------
100.00    0.007362          17       427        38 total

$ time ./python -c pass

real    0m0.024s
user    0m0.020s
sys     0m0.004s
$ time ./python -c pass

real    0m0.029s
user    0m0.025s
sys     0m0.004s
$ time ./python -c pass

real    0m0.028s
user    0m0.027s
sys     0m0.001s

after adding stdlib modules

# repeated 100 times and averaged
$ strace -c ./python -c pass
% time     seconds  usecs/call     calls    errors syscall
------ ----------- ----------- --------- --------- ----------------
 15.22    0.000731          10        68         0 rt_sigaction
 10.97    0.000527          25        21         0 mmap
 10.35    0.000497          19        25         5 stat
  9.74    0.000468          27        17         2 openat
  8.47    0.000407          20        20         0 fstat
  7.97    0.000383          21        18         0 close
  5.27    0.000253          21        12         0 mprotect
  4.99    0.000240          26         9         0 read
  4.88    0.000234          23        10         0 getdents
  3.28    0.000158          19         8         0 brk
  2.64    0.000127          18         7         7 access
  2.10    0.000101          16         6         4 ioctl
  2.04    0.000098          48         2         0 fcntl
  1.97    0.000095          23         4         3 lseek
  1.04    0.000050          16         3         0 dup
  0.75    0.000036          35         1         0 gettid
  0.72    0.000035          34         1         0 sysinfo
  0.70    0.000034          33         1         0 set_robust_list
  0.67    0.000032          32         1         0 getrandom
  0.63    0.000030          30         1         0 set_tid_address
  0.63    0.000030          10         3         0 munmap
  0.62    0.000030          29         1         0 futex
  0.62    0.000030          29         1         0 rt_sigprocmask
  0.60    0.000029          28         1         0 prlimit64
  0.56    0.000027          27         1         1 readlink
  0.55    0.000026          26         1         0 getcwd
  0.51    0.000025          24         1         0 getgid
  0.44    0.000021          21         1         0 getegid
  0.44    0.000021          21         1         0 getuid
  0.40    0.000019          19         1         0 arch_prctl
  0.21    0.000010          10         1         0 geteuid
  0.00    0.000000           0         1         0 execve
------ ----------- ----------- --------- --------- ----------------
100.00    0.004806          19       249        22 total

$ time ./python -c pass

real    0m0.025s
user    0m0.025s
sys     0m0.001s
$ time ./python -c pass

real    0m0.026s
user    0m0.021s
sys     0m0.005s
$ time ./python -c pass

real    0m0.023s
user    0m0.022s
sys     0m0.001s

(Note that I ran this comparison in a VM, so the pure numbers aren't so reliable.)

Overall, It seems like there is a real, positive impact on performance (roughly 10%?), though the difference isn't huge (and the VM makes it a little cloudy).

[ericsnowcurrently]

I re-ran the comparison 100 times and averaged the results:

before freezing those stdlib modules:

$ time ./python -c pass

real:    0m0.023
user:    0m0.020
sys:     0m0.003

after freezing those stdlib modules:

$ time ./python -c pass

real:    0m0.020
user:    0m0.018
sys:     0m0.002

So I'm seeing a 15% improvement, even after factoring in disk caching. (again, running in a VM) I'm still going to check those other "todo" items above.

[ericsnowcurrently]

@gvanrossum, if I exclude encodings.* from freezing, the results are basically the same. (It's actually slower: 21ms instead of 20ms.)

[ericsnowcurrently]

I've verified that all the startup modules are either builtin or frozen:

$ ./python -c 'import sys^Mfor name in sorted(sys.modules): print(f"{name:30} {sys.modules[name]}")'
__main__                       <module '__main__' (built-in)>
_abc                           <module '_abc' (built-in)>
_codecs                        <module '_codecs' (built-in)>
_collections_abc               <module '_collections_abc' (frozen)>
_frozen_importlib              <module '_frozen_importlib' (frozen)>
_frozen_importlib_external     <module '_frozen_importlib_external' (frozen)>
_imp                           <module '_imp' (built-in)>
_io                            <module '_io' (built-in)>
_signal                        <module '_signal' (built-in)>
_sitebuiltins                  <module '_sitebuiltins' (frozen)>
_stat                          <module '_stat' (built-in)>
_thread                        <module '_thread' (built-in)>
_warnings                      <module '_warnings' (built-in)>
_weakref                       <module '_weakref' (built-in)>
abc                            <module 'abc' (frozen)>
builtins                       <module 'builtins' (built-in)>
codecs                         <module 'codecs' (frozen)>
encodings                      <module 'encodings' (frozen)>
encodings.aliases              <module 'encodings.aliases' (frozen)>
encodings.utf_8                <module 'encodings.utf_8' (frozen)>
genericpath                    <module 'genericpath' (frozen)>
io                             <module 'io' (frozen)>
marshal                        <module 'marshal' (built-in)>
os                             <module 'os' (frozen)>
os.path                        <module 'posixpath' (frozen)>
posix                          <module 'posix' (built-in)>
posixpath                      <module 'posixpath' (frozen)>
site                           <module 'site' (frozen)>
stat                           <module 'stat' (frozen)>
sys                            <module 'sys' (built-in)>
time                           <module 'time' (built-in)>
zipimport                      <module 'zipimport' (frozen)>

[gvanrossum]

This is awesome. Maybe move the generated files to a subdir like clinic?

[ericsnowcurrently]

cleanup:

• https://bugs.python.org/issue45019
• bpo-45019: Add a tool to generate list of modules to include for frozen modules python/cpython#27980

frozen stdlib modules:

• https://bugs.python.org/issue45020
• bpo-45020: Freeze the modules imported during startup. python/cpython#28107

[markshannon]

One way to remove the regenerated file problem in PRs is to only generate one file.
One 1M line file that no one can read is no more or less incomprehensible that 100 10K line files that no one can read, but the GitHub PR experience is vastly better.

Another thing we can do is to only put two bytes per line not 16 in the output of the bytecode. Since we often add or remove a few instructions here and there, the diffs can be very large as subsequently lines get shifted.
With only two bytes (one instruction) per line the diffs should be much more localized preventing the size of the repo from ballooning. And drop the whitespace on the left to keep the file size down.

E.g.

    100,19,100,20,147,1,100,21,100,22,147,1,100,23,100,24,

would become

100,19,
100,20,
147,1,
100,21,
100,22,
147,1,
100,23,
100,24,

which is only 3 bytes larger, but produces much smaller diffs should one or two instructions be changed.

Forget readability of generated code. Minimizing the number of files and keeping diffs down are much more important to ease of development.

[gvanrossum]

But it’s marshal data, and the constants, names etc. are not aligned on even bytes. So maybe one byte per line?

[ericsnowcurrently]

Note that the generated .h files are no longer in the repo, so they don't show up in diffs or PRs any more. Thus we don't need to worry about strategies to mitigate the earlier annoyances.

One way to remove the regenerated file problem in PRs is to only generate one file.

Good point. Having a single .h file would also mean a single include in frozen.c (and we wouldn't have to generate the includes any more). So this may be worth doing even though .h files aren't in the repo.

Another thing we can do is to only put two bytes per line not 16 in the output of the bytecode. Since we often add or remove a few instructions here and there, the diffs can be very large as subsequently lines get shifted.

Good point, though there isn't much value to changing this now.

[ericsnowcurrently]

@markshannon, I've created an "always default to on" branch: https://github.com/ericsnowcurrently/cpython/tree/frozen-modules-always-default-on.