zero build slow

Severin Gehwolf sgehwolf at redhat.com
Thu Nov 23 09:42:19 UTC 2017


Hi Thomas,

On Wed, 2017-11-22 at 17:26 +0100, Thomas Stüfe wrote:
> Hi guys,
> 
> thanks for your help.
> 
> I think the reason is the combination of zero+slowdebug and the fact that
> we run our own binaries during build (at least this is true for jmod).
> 
> I am building on Ubuntu 16.4, "normal" machine (i7, 16G ram, ssd). Normally
> my builds take ~5-10 minutes.
> 
> My configure line in this case:
> 
> CONFIGURE_COMMAND_LINE:=--with-boot-jdk=/shared/projects/openjdk/jdks/openjdk9
> --with-debug-level=slowdebug --with-jvm-variants=zero
> --with-native-debug-symbols=internal
> 
> Wanted slowdebug to test something. Boot jdk is downloaded from
> adoptopenjdk (I was lazy).
> 
> When building (make images), top is dominated by jmod, which is taken from
> the output directory, not the bootstrap vm:
> 
> 19710 thomas    30  10 1176348  63480  15800 S 100,0  0,4   2:02.64
> /shared/projects/openjdk/jdk-hs/output-zero/jdk/bin/jmod -J-XX:+UseSerialGC
> -J-Xms32M -J-Xmx512M -J-XX:TieredStopAtLevel=1 create --module-version
> 10-internal --target-platform linux-amd64 --module-path /sha+
> 
> Note that my normal (non-zero) slowdebug builds are done in ~5 minutes, but
> this thing here is now running for 30 minutes and not done yet. Hence my
> assumption, that we need to run jmod from the build directory and that a
> zero jmod in slowdebug mode is terrible.
> 
> It probably makes sense, when I run jmod interpreted and the C++ code is
> not optimized... Hence my original question, whether I can force the build
> to use a different jmod. This would also help in cases when I am developing
> and my hotspot is crashy - currently, the build does not go thru in this
> case.
> 
> I will retry with fastdebug to see how much faster this is.

That matches my experience.

FWIW, despite it being slow, running jmod using the just built JDK
proved to be a reasonable test case for Zero issues. I've seen asserts
being triggered or segfaults to happen during some runs indicating that
there were issues. If you don't care about that, Erik pointed you at
the solution :)

Cheers,
Severin



More information about the build-dev mailing list