On Fri, 4 Dec 2020 17:34:33 GMT, Erik Joelsson <er...@openjdk.org> wrote:

>> Ok, after looking at `set -x` output WSL seems to have a problem with the 
>> space in the path `/mnt/c/Program Files/Java/jdk-15` e.g. if I move the jdk 
>> to `/mnt/h/jdk-15` it works. Might be a missing fixup somewhere? (FWIW, that 
>> Program Files dir does have an 8dot3 name).
>> 
>> With that I can build `images` on both Cygwin and WSL. Build times are about 
>> 29 mins and 18 mins respectively.
>
> Do you see any significant difference in build times compared to before the 
> patch?

@erikj79 The big difference is the shell script fixpath, as opposed to a native 
fixpath.exe. I've gone to great pains to make sure the "exec" pathway in 
fixpath.sh contains no external calls, but only native bash constructs, to 
avoid forks, which are expensive at Windows. With that said, fixpath.sh is 
still slower than fixpath.exe. A typical invocation with fixpath.sh is about 30 
ms slower than fixpath.exe;  that is, in average 0.079 s instead of  0.047 s. 
(This is on Cygwin; on WSL fixpath.sh executes basically instantaneously.) This 
do translate to longer build times, but I have a hard time getting a proper 
value. The fluctuation in build times is so large so it's hard to say with 
certainty, but e.g. on GitHub Actions the build time increased from 53-57 
minutes typically, to 58-62 minutes typically.  On my personal machine, I could 
get more reliable and repeatable results. There the time taken to build hotspot 
increased from 4 min 45 s to 4 min 53 s.

All these values are from running on Cygwin. The build times on WSL noticeably 
faster. (As a follow-up bug, I intend to migrate the GitHub Action script to 
running on WSL instead of Cygwin.) On machine, the build time decreased from 
over. 15 minutes, to under 13, when going from Cygwin to WSL.

-------------

PR: https://git.openjdk.java.net/jdk/pull/1597

Reply via email to