AxiDraw software release 3.0 made a significant change in how documents were processed as part of the plotting process:
Historically, this software [...] has parsed the SVG document essentially one element at a time, plotting each element, and then going onto the next element in the document. [...] This distributes the "up front" computation time for parsing different types of SVG elements across the duration of the plot.
Starting with version 3.0, a copy of the SVG document is now fully processed to create an intermediate "digest" prior to plotting [...]. This method adds overhead in terms of memory use and initial computation time [...]
The net effect is that some users upgrading from AxiDraw < 3.0 experienced that the software felt much slower even in cases where the total computation + plot time was unchanged. Since that release, we have become of some use cases -- with very large SVG files -- for which the initial delay before plotting was taking a very long time.
In AxiDraw 3.8, we have tried to make the software faster, particularly in that initial computation stage, and particularly for those very large SVG files. Rather than saying that the software is (say) "20% faster", or "up to 97% faster", we are documenting some specific measurements in this issue that highlight the actual changes, and demonstrate which types of things are now faster.
Test files and conditions
Five files were used for testing, some of them directly from users, some of them created from scratch to test particular corner cases. (These files do include user data and not made publicly available.)
File
Size
Length
Notes
test.1svg
277 MB
1225 m
1.6 million polyline elements. Many short disordered segments that together form paths. Tests joining & optimization.
test2.svg
280 MB
760 m
Only 80 paths, but each is very long. Tests path parsing & supersampling
test3.svg
6.3 MB
15 m
Moderate size user file; 5500 short disjoint paths
test4.svg
12.2 MB
389 m
Extremely high point density, tests supersampling.
test5.svg
32.8 MB
885 m
Heavily nested transformed objects, high point density.
Default settings were used on all files, except that the model was set to 5 (SE/A1 size; some of the test files are large) for all tests and the gap joining threshold was reduced to 0.001 for all tests on file test1.svg (only).
All times measured are in seconds; lower is better, and these should be regarded as rough measurements only.*
(*Time measurements were made on a MacBook Pro computer with an Apple M1 Pro CPU, on Python 3.10, while the computer was otherwise being used as a normal computer. Some runs were verified with a second or third test to verify the rough time results. These results provide a useful, if rough, comparison, but we advise that no weight should be given past 2 significant digits in any result.)
Significant performance gains for initial processing of the SVG file, particularly for large input files.
Parsing and processing SVG, saving plot digest. Effectively measures time between initiating a plot and when the AxiDraw begins drawing (or, begins drawing when resuming a plot). Run with different options for plot optimization:
Full optimization, with joining and then reordering for speed: axicli test*.svg -L5 --digest 2 -T --reorder 2
(-L5: model SE/A1. -T: Enable time reports. --digest 2: Only process file; don't print)
File
Strict 3.7
Strict 3.8
Join 3.7
Join 3.8
Full 3.7
Full 3.8
test1.svg
78
34
161
122
161
121
test2.svg
1068
181
1085
178
1078
178
test3.svg
25
2.8
25
3.2
25
3.5
test4.svg
34
6.4
34
6.7
34
6.8
test5.svg
50
11
53
13
51
12
File parsing and processing times are significantly lower in AxiDraw 3.8; initial processing prior to optimizations is 50-80% faster, with little or no change in optimization times. Overall initial processing, including optimization time, is 25 - 85% faster, depending on the nature of the file.
Resume -> Home command is significantly faster
Starting with the files file paused about 50% through, and then using the resume home command:
time axicli test*.svg --mode res_home -L5 -v
File
3.7
3.8
3.7 (Plob)
3.8 (Plob)
test1.svg
83
12
12.2
5.57
test2.svg
1104
30
1.53
0.55
test3.svg
25
0.23
0.19
0.15
test4.svg
34
0.31
0.31
0.17
test5.svg
51
0.62
0.50
0.23
83 -> 12 s is not bad, but 1104 -> 30 s is a very good improvement. The tests designated "(Plob)" are ones that began with a Plob digest file, which means that most of the file parsing and processing was already handled prior to starting the job.
Multi-page previews are faster
Testing with optimization strictly disabled, previewing 1 page or 10, with otherwise default settings:
axicli test*.svg -L5 --reorder 4 -vT
axicli test*.svg -L5 --reorder 4 -vT --copies 10
File
1pp 3.7
1pp 3.8
10pp 3.7
10pp 3.8
slope 3.7
slope 3.8
test1.svg
134
86
1411
84
141.9
~0
test2.svg
1104
195
11407
193
1144.8
~0
test3.svg
26
3.45
265
3.45
26.6
~0
test4.svg
40
8.16
411
8.13
41.2
~0
test5.svg
61
14
630
14
63.2
~0
The slope columns give the amount of additional time, in seconds) to preview each additional page. While the baseline times to preview a page are shorter due to faster processing (see the tables earlier on this page), the added time per page is essentially zero here. That's because we added a trivial and obvious (in retrospect...) optimization: multi-page previews now simply multiply the time and distance for a single page, if the random_start option is disabled.
(* Random start does not significantly affect the execution time for a single page here.)
File
1pp 3.7
1pp 3.8
10pp 3.7
10pp 3.8
slope 3.7
slope 3.8
test1.svg
134
86
1383
558
138.8
52.4
test2.svg
1104
195
11235
423
1125.7
25.3
test3.svg
26
3.45
267
8.94
26.8
0.6
test4.svg
40
8.16
403
25
40.3
1.9
test5.svg
61
14
382
45
35.7
3.4
With random_start enabled, each individual page is fully previewed (the motion profile for the entire document is computed, so that there is a full record of the time and distance). The additional time per page (differential or marginal, depending if you come from science vs economics) is not zero, but it is significantly lower in AxiDraw 3.8 than AxiDraw 3.7.
Motivation
AxiDraw software release 3.0 made a significant change in how documents were processed as part of the plotting process:
The net effect is that some users upgrading from AxiDraw < 3.0 experienced that the software felt much slower even in cases where the total computation + plot time was unchanged. Since that release, we have become of some use cases -- with very large SVG files -- for which the initial delay before plotting was taking a very long time.
In AxiDraw 3.8, we have tried to make the software faster, particularly in that initial computation stage, and particularly for those very large SVG files. Rather than saying that the software is (say) "20% faster", or "up to 97% faster", we are documenting some specific measurements in this issue that highlight the actual changes, and demonstrate which types of things are now faster.
Test files and conditions
Five files were used for testing, some of them directly from users, some of them created from scratch to test particular corner cases. (These files do include user data and not made publicly available.)
Default settings were used on all files, except that the model was set to 5 (SE/A1 size; some of the test files are large) for all tests and the gap joining threshold was reduced to 0.001 for all tests on file test1.svg (only).
All times measured are in seconds; lower is better, and these should be regarded as rough measurements only.*
(*Time measurements were made on a MacBook Pro computer with an Apple M1 Pro CPU, on Python 3.10, while the computer was otherwise being used as a normal computer. Some runs were verified with a second or third test to verify the rough time results. These results provide a useful, if rough, comparison, but we advise that no weight should be given past 2 significant digits in any result.)
Significant performance gains for initial processing of the SVG file, particularly for large input files.
Parsing and processing SVG, saving plot digest. Effectively measures time between initiating a plot and when the AxiDraw begins drawing (or, begins drawing when resuming a plot). Run with different options for plot optimization:
axicli test*.svg -L5 --digest 2 -T --reorder 4axicli test*.svg -L5 --digest 2 -T --reorder 0axicli test*.svg -L5 --digest 2 -T --reorder 2(-L5: model SE/A1. -T: Enable time reports. --digest 2: Only process file; don't print)
File parsing and processing times are significantly lower in AxiDraw 3.8; initial processing prior to optimizations is 50-80% faster, with little or no change in optimization times. Overall initial processing, including optimization time, is 25 - 85% faster, depending on the nature of the file.
Resume -> Home command is significantly faster
Starting with the files file paused about 50% through, and then using the resume home command:
time axicli test*.svg --mode res_home -L5 -v83 -> 12 s is not bad, but 1104 -> 30 s is a very good improvement. The tests designated "(Plob)" are ones that began with a Plob digest file, which means that most of the file parsing and processing was already handled prior to starting the job.
Multi-page previews are faster
Testing with optimization strictly disabled, previewing 1 page or 10, with otherwise default settings:
axicli test*.svg -L5 --reorder 4 -vTaxicli test*.svg -L5 --reorder 4 -vT --copies 10The slope columns give the amount of additional time, in seconds) to preview each additional page. While the baseline times to preview a page are shorter due to faster processing (see the tables earlier on this page), the added time per page is essentially zero here. That's because we added a trivial and obvious (in retrospect...) optimization: multi-page previews now simply multiply the time and distance for a single page, if the
random_startoption is disabled.OK, but what if
random_startoption is enabled?axicli test*.svg -L5 --reorder 4 -vT*axicli test*.svg -L5 --reorder 4 -vT --random_start --copies 10(* Random start does not significantly affect the execution time for a single page here.)
With
random_startenabled, each individual page is fully previewed (the motion profile for the entire document is computed, so that there is a full record of the time and distance). The additional time per page (differential or marginal, depending if you come from science vs economics) is not zero, but it is significantly lower in AxiDraw 3.8 than AxiDraw 3.7.