Re: Octave on Windows takes a long time to start & run.m reloads Of packages

[PhilipNienhuis]

PhilipNienhuis wrote
> John W. Eaton wrote:
>> On 2/2/20 1:18 PM, PhilipNienhuis wrote:
>>> While running __run_test_suite__ with a freshly cross-built 
>>> Octave-6.0.0 on
>>> Windows, I saw a FAIL in run.m.
>>> That FAIL has been there for quite some time, didn't bother, "test
>>> run.m"
>>> works fine when run outside of __run_test_suite__.
>>>
>>> But because since one or two days Octave on Windows suddenly takes a
>>> very
>>> long time to load I added some tic/toc statements in
>>> chk_spreadsheet_support.m, invoked by __init_io__.m, in turn invoked by
>>> PKG_ADD in the io package, to find out if loading Java classes for
>>> LibreOffice would take so long. I saw no strange things there, finding &
>>> loading loading the .jars takes altogether ~5-6 seconds of in total 25
>>> seconds for he gUI to appear until the prompt is shown.
>> 
>> Can you bisect and determine which changeset caused the difference in 
>> behavior?
>> 
>> If it is something that happened recently, maybe it is an unintended 
>> side effect of this change:
>> 
>>    http://hg.savannah.gnu.org/hgweb/octave/rev/262cdfc6faf9
>> 
>> which made the load_path class use the absolute canonical directory name 
>> as the key for the private function map?
> 
> Sure, I'll do. This started to happen very recently, just a few days ago.
> It is most conspicuously in the chk_spreadsheet_support.m function in 
> the io package that helps to load Java classes or spreadsheet I/O. I'll 
> start with running it through the profiler.

Some profiling results from a version built Jan 29 (still fast) and Jan 31
(slow), in attached .png.
(sorry for that, it's a bit hard to copy tables into Nabble, otherwise it
would be html.)
In between those dates the slowdown took place.
I also tried with Octave-5.1.0, that is still a bit faster than dev Octave
of the end of last January.
<https://octave.1599824.n4.nabble.com/file/t248596/chk_spr-supp_timings.png> 
It turns out that especially string functions take much longer to execute
(some in the order of 50 times slower). That could be due to the functions
themselves but it can be equally probably due to the variable content they
are applied to. I still have to sort out the latter, TTBOMK most of it is
javaclasspath contents, path stuff and some calls to dir.m.
chk_spreadsheet_support.m is a fairly complex function and I have little
time until Friday so I have to see how far I get.

FTR, if I do not load the io package, starting up Octave (incl. mapping,
geometry/matgeom, windows, signal/control and linear-algebra packages) seems
about as fast as it has always been the last years.

Philip




--
Sent from: https://octave.1599824.n4.nabble.com/Octave-Maintainers-f1638794.html