Support reading from .mo instead of .po

[maennchen]

Now that we cleaned up the compilation, I thought it would be interesting to compare the .po / .mo performance.

Script

https://github.com/elixir-gettext/gettext/blob/846cac137f7764101d734ff8f8c9ed30c4d64dda/performance_test.exs

Gettext Files

I've used the translation catalog of hygeia, which is one of the largest OS ones I know of.

Output

Expand Details

Expo Parse - .po is 11.3 times slower and uses 8.5 times more memory
Gettext Compile - .po is 1.5 times slower and uses 1.5 times more memory

11:24:34.545 [info] Preparing .mo files

11:24:34.590 [info] Performance Test Expo Parse
Operating System: Linux
CPU Information: 11th Gen Intel(R) Core(TM) i7-1185G7 @ 3.00GHz
Number of Available Cores: 8
Available memory: 46.76 GB
Elixir 1.17.2
Erlang 27.0.1
JIT enabled: true

Benchmark suite executing with the following configuration:
warmup: 5 s
time: 1 min
memory time: 2 s
reduction time: 0 ns
parallel: 8
inputs: none specified
Estimated total run time: 2 min 14 s

Benchmarking expo_po_parse ...
Benchmarking expo_mo_parse ...
Calculating statistics...
Formatting results...

Name                    ips        average  deviation         median         99th %
expo_mo_parse         84.78       11.79 ms    ±15.96%       11.56 ms       17.48 ms
expo_po_parse          7.52      132.95 ms    ±22.55%      125.54 ms      251.71 ms

Comparison: 
expo_mo_parse         84.78
expo_po_parse          7.52 - 11.27x slower +121.15 ms

Memory usage statistics:

Name                  average  deviation         median         99th %
expo_mo_parse         3.96 MB     ±0.00%        3.96 MB        3.96 MB
expo_po_parse        33.67 MB     ±0.00%       33.67 MB       33.67 MB

Comparison: 
expo_mo_parse         3.96 MB
expo_po_parse        33.67 MB - 8.51x memory usage +29.71 MB

11:26:48.993 [info] Performance Test Gettext Backend Compile
Operating System: Linux
CPU Information: 11th Gen Intel(R) Core(TM) i7-1185G7 @ 3.00GHz
Number of Available Cores: 8
Available memory: 46.76 GB
Elixir 1.17.2
Erlang 27.0.1
JIT enabled: true

Benchmark suite executing with the following configuration:
warmup: 5 s
time: 1 min
memory time: 2 s
reduction time: 0 ns
parallel: 8
inputs: none specified
Estimated total run time: 2 min 14 s

Benchmarking gettext_po_compile ...
Benchmarking gettext_mo_compile ...
Calculating statistics...
Formatting results...

Name                         ips        average  deviation         median         99th %
gettext_mo_compile         0.123         8.10 s     ±2.45%         8.12 s         8.63 s
gettext_po_compile        0.0818        12.23 s     ±2.29%        12.25 s        12.73 s

Comparison: 
gettext_mo_compile         0.123
gettext_po_compile        0.0818 - 1.51x slower +4.13 s

Memory usage statistics:

Name                       average  deviation         median         99th %
gettext_mo_compile       118.33 MB     ±0.00%      118.33 MB      118.33 MB
gettext_po_compile       173.44 MB     ±0.02%      173.44 MB      173.47 MB

Comparison: 
gettext_mo_compile       118.33 MB
gettext_po_compile       173.44 MB - 1.47x memory usage +55.11 MB

Based on those numbers, we should consider supporting to compile from .mo instead of .po if the user requests it.

EDIT: elixir-gettext/expo#141 will improve it slightly to 1.45 times slower and 1.28 times more memory for the backend compile, but still doesn't come close.

[maennchen]

Updated Numbers from Expo Metadata Stripping (#398)

Expand Details

Expo Parse - .po is 11.7 times slower and uses 8.5 times more memory
Expo Parse - .po with metadata stripping is 5.0 times slower and uses 2.8 times more memory
Gettext Compile - .po is 1.4 times slower and uses 1.3 times more memory

08:37:11.016 [info] Preparing .mo files

08:37:11.068 [info] Performance Test Expo Parse
Operating System: Linux
CPU Information: 11th Gen Intel(R) Core(TM) i7-1185G7 @ 3.00GHz
Number of Available Cores: 8
Available memory: 46.76 GB
Elixir 1.17.2
Erlang 27.0.1
JIT enabled: true

Benchmark suite executing with the following configuration:
warmup: 5 s
time: 1 min
memory time: 2 s
reduction time: 0 ns
parallel: 8
inputs: none specified
Estimated total run time: 3 min 21 s

Benchmarking expo_po_parse ...
Benchmarking expo_po_runtime_only_parse ...
Benchmarking expo_mo_parse ...
Calculating statistics...
Formatting results...

Name                                 ips        average  deviation         median         99th %
expo_mo_parse                      73.26       13.65 ms    ±26.63%       12.86 ms       28.36 ms
expo_po_runtime_only_parse         14.71       67.96 ms    ±18.68%       65.42 ms      108.59 ms
expo_po_parse                       6.24      160.27 ms    ±22.91%      151.16 ms      275.76 ms

Comparison: 
expo_mo_parse                      73.26
expo_po_runtime_only_parse         14.71 - 4.98x slower +54.31 ms
expo_po_parse                       6.24 - 11.74x slower +146.62 ms

Memory usage statistics:

Name                               average  deviation         median         99th %
expo_mo_parse                      3.96 MB     ±0.00%        3.96 MB        3.96 MB
expo_po_runtime_only_parse        11.20 MB     ±0.00%       11.20 MB       11.20 MB
expo_po_parse                     33.67 MB     ±0.00%       33.67 MB       33.67 MB

Comparison: 
expo_mo_parse                      3.96 MB
expo_po_runtime_only_parse        11.20 MB - 2.83x memory usage +7.24 MB
expo_po_parse                     33.67 MB - 8.51x memory usage +29.71 MB

08:40:32.577 [info] Performance Test Gettext Backend Compile
Operating System: Linux
CPU Information: 11th Gen Intel(R) Core(TM) i7-1185G7 @ 3.00GHz
Number of Available Cores: 8
Available memory: 46.76 GB
Elixir 1.17.2
Erlang 27.0.1
JIT enabled: true

Benchmark suite executing with the following configuration:
warmup: 5 s
time: 1 min
memory time: 2 s
reduction time: 0 ns
parallel: 8
inputs: none specified
Estimated total run time: 2 min 14 s

Benchmarking gettext_po_compile ...
Benchmarking gettext_mo_compile ...
Calculating statistics...
Formatting results...

Name                         ips        average  deviation         median         99th %
gettext_mo_compile         0.116         8.66 s     ±3.22%         8.69 s         9.16 s
gettext_po_compile        0.0801        12.48 s     ±4.56%        12.43 s        13.98 s

Comparison: 
gettext_mo_compile         0.116
gettext_po_compile        0.0801 - 1.44x slower +3.83 s

Memory usage statistics:

Name                       average  deviation         median         99th %
gettext_mo_compile       118.33 MB     ±0.00%      118.33 MB      118.33 MB
gettext_po_compile       151.73 MB     ±0.00%      151.73 MB      151.73 MB

Comparison: 
gettext_mo_compile       118.33 MB
gettext_po_compile       151.73 MB - 1.28x memory usage +33.39 MB