On Fri, 2 Jun 2023 20:32:14 GMT, Doug Simon <[email protected]> wrote:
> This PR improves the startup time for libgraal by speeding up how
> `VM.savedProps` is copied into libgraal. This data structure is now
> serialized to a native buffer directly from C++ and the native buffer is then
> directly decoded by libgraal.
>
> ## Times
>
> The basic benchmarking below shows that this change brings the time for a nop
> Java app with eager libgraal initialization (2) down to almost the same time
> as lazy libgraal initialization (1). The latter typically means no libgraal
> initialization happens as a top tier JIT compilation is never scheduled in
> such a short running app.
>
>
> public class Nop {
> public static void main(String[] args) {}
> }
>
>
> (1) Baseline (no options):
>
>> for i in (seq 10); java Nop; end
> 0.05 real 0.04 user 0.01 sys
> 0.04 real 0.03 user 0.01 sys
> 0.04 real 0.03 user 0.01 sys
> 0.04 real 0.03 user 0.01 sys
> 0.03 real 0.03 user 0.00 sys
> 0.04 real 0.03 user 0.01 sys
> 0.04 real 0.03 user 0.00 sys
> 0.03 real 0.03 user 0.00 sys
> 0.04 real 0.03 user 0.01 sys
> 0.03 real 0.03 user 0.00 sys
>
>
> (2) Eagerly initialize libgraal (with PR):
>
>> for i in (seq 10); /usr/bin/time java -XX:+EagerJVMCI Nop; end
> 0.06 real 0.04 user 0.01 sys
> 0.05 real 0.03 user 0.01 sys
> 0.05 real 0.03 user 0.01 sys
> 0.05 real 0.03 user 0.01 sys
> 0.05 real 0.03 user 0.01 sys
> 0.05 real 0.03 user 0.01 sys
> 0.05 real 0.03 user 0.01 sys
> 0.05 real 0.03 user 0.01 sys
> 0.05 real 0.03 user 0.01 sys
> 0.05 real 0.03 user 0.01 sys
>
>
> (3) Eagerly initialize libgraal (without PR):
>
>> for i in (seq 10); /usr/bin/time java -XX:+EagerJVMCI Nop; end
> 0.11 real 0.08 user 0.02 sys
> 0.08 real 0.06 user 0.01 sys
> 0.08 real 0.07 user 0.01 sys
> 0.10 real 0.07 user 0.01 sys
> 0.08 real 0.06 user 0.01 sys
> 0.10 real 0.07 user 0.01 sys
> 0.08 real 0.07 user 0.01 sys
> 0.08 real 0.07 user 0.01 sys
> 0.08 real ...
I don't really love the hard code parsing of the HashMap. What properties are
actually required for JVMCI? It seems to me that the contents of
Arguments::system_properties() should contain all the properties we want to
advertise to JVMCI. That would have avoid having to decode them after they've
been converted into Java objects.
-------------
PR Comment: https://git.openjdk.org/jdk/pull/14291#issuecomment-1577305531
