RFR: JDK-8236604: Optimize SystemDictionary::resolve_well_known_classes for CDS
Claes Redestad
claes.redestad at oracle.com
Tue Feb 25 05:41:54 UTC 2020
Hi,
before diving into a full review, I took your patch for a spin and
can confirm a speed-up around .5ms, along with about a 1% reduction of
#instructions and #branches on Hello World[1]. Sweet!
/Claes
[1]
Baseline:
53.028269 task-clock (msec) # 1.396 CPUs
utilized ( +- 0.34% )
244 context-switches # 0.005 M/sec
( +- 0.60% )
32 cpu-migrations # 0.607 K/sec
( +- 0.39% )
3,634 page-faults # 0.069 M/sec
( +- 0.03% )
138,590,680 cycles # 2.614 GHz
( +- 0.25% )
112,946,802 instructions # 0.81 insns per
cycle ( +- 0.06% )
22,428,330 branches # 422.950 M/sec
( +- 0.06% )
771,755 branch-misses # 3.44% of all
branches ( +- 0.13% )
0.037981200 seconds time elapsed
( +- 0.39% )
Patched:
52.537826 task-clock (msec) # 1.409 CPUs
utilized ( +- 0.26% )
244 context-switches # 0.005 M/sec
( +- 0.65% )
32 cpu-migrations # 0.607 K/sec
( +- 0.43% )
3,637 page-faults # 0.069 M/sec
( +- 0.04% )
137,129,495 cycles # 2.610 GHz
( +- 0.23% )
111,931,764 instructions # 0.82 insns per
cycle ( +- 0.07% )
22,215,346 branches # 422.845 M/sec
( +- 0.07% )
766,573 branch-misses # 3.45% of all
branches ( +- 0.13% )
0.037283845 seconds time elapsed
( +- 0.30% )
On 2020-02-25 06:02, Yumin Qi wrote:
> Hi,
>
> Please review fix:
>
> Bug: https://bugs.openjdk.java.net/browse/JDK-8236604
>
> Webrev:http://cr.openjdk.java.net/~minqi/8236604/webrev-00
>
> Description: Optimize well known classes initialization for CDS
> during jvm startup.
>
> When run with CDS, initialize well known classes (95 classes) will
> call resolve_or_fail thus go though the resolve functions, locks etc.
> The initialization of well-known classes happens in fact in very early
> stage, those lock can be avoided. The fix is serializing
> SystemDictionay::_well_known_klasses into CDS and in runtime restore
> them by avoiding call resolve_or_fail. Since Compile_lock has to be held
> for SystemDictionary::add_to_hierachy, the fix avoids calling it by
> copying the code from it in SytemDictionary::quick_resolve, but this way
> I have to remove two asserts which will assert on Compile_lock. The
> reminding usage of the lock was added as comments on the function
> declarations of Klass::append_to_sibling_list and
> InstanceKlass::add_implementor. The two functions calling paths have
> been checked to make sure they are not used in other places.
>
> Performance measured by take time before and after
> SystemDictionary::resolve_well_known_classes(manually modified
> orignal/new versions), since this is the most direct measurement and
> accurate(excluded in review code):
>
> + jlong s0 = os::javaTimeNanos();
> resolve_well_known_classes(CHECK);
>
> + jlong s1 = os::javaTimeNanos();
>
> // print out s1 - s0
>
> Run -version 2000 times for original/new versions respectively and
> took the averages. The saving is about 18% (2.9ms vs 2.4ms), it is not a
> big saving but still helps to improve the startup time.
>
> Testing: local jtreg test on linux-x86.
>
> pending mach5 hs-tier1-4
>
>
> Thanks
>
> Yumin
>
More information about the hotspot-runtime-dev
mailing list