While working on things related to JDK-8246402, I wrote a JMH that constantly loads new classes and creates a few instances of them in each iteration.
On my WS

16-b26: 451.955 ± 1.864 ops/s
16-b27: 394.893 ± 5.071 ops/s

Using perf top, I can see from b27 the test spends 5-7% of its time in CompilerOracle::has_option_value<double>
 even when the compilation has quieted down after the first few seconds of running.
I think this is related to https://bugs.openjdk.java.net/browse/JDK-8256508

   9.11% [kernel] [k] read_hpet
   6.96% libjvm.so [.] CompilerOracle::has_option_value<double>
   4.01% libjvm.so [.] steal_work
   2.93% [kernel] [k] do_syscall_64
   2.31% libc-2.23.so [.] vfprintf


The micro is here:

https://github.com/ericcaspole/jmh-jdk-microbenchmarks/blob/dyn-load-and-new/micros-jdk8/src/main/java/org/openjdk/bench/java/lang/DynamicLoading.java

Run it like:
$ numactl --membind=1 --cpunodebind=1 /opt/jdk-16-b27/bin/java -jar target/micros-jdk8-1.0-SNAPSHOT.jar DynamicLoading.loadAndSeveralNewInstance -f 1 -i 8 -p instances=25 -p numberOfClasses=10 -jvmArgs "-XX:+PrintCompilation -XX:+UseParallelGC"

All that being said I dont know if other benchmarks see any regression from tihs one.