A simple instrumentation tells that most of the time we are installing just a few classes, with total work under <400 us. This is not enough to get decent parallel improvements. In fact, it is close to the break-even against the sequential version. Therefore, the use of parallelStream is questionable. The fix is very simple:

$ hg diff
diff -r ad7a46083364 src/jdk.scripting.nashorn/share/classes/jdk/nashorn/internal/runtime/Context.java
--- a/src/jdk.scripting.nashorn/share/classes/jdk/nashorn/internal/runtime/Context.java Mon Aug 25 20:26:53 2014 +0530
+++ b/src/jdk.scripting.nashorn/share/classes/jdk/nashorn/internal/runtime/Context.java Mon Aug 25 22:44:54 2014 +0400
@@ -166,7 +166,7 @@
             // do these in parallel, this significantly reduces class installation overhead
             // however - it still means that every thread needs a separate doPrivileged
             final Global global = currentGlobal.get();
- classes.parallelStream().forEach(
+ classes.stream().forEach(
                 new Consumer<Class<?>>() {
                     @Override
                     public void accept(final Class<?> clazz) {

On the current jdk9/dev and this simple benchmark:

$ for I in `seq 1 5`; do ~/trunks/jdk9-dev/build/linux-x86_64-normal-server-release/images/j2sdk-image/bin/java -jar ~/trunks/jdk9-dev/build/linux-x86_64-normal-server-release/images/j2sdk-image/jre/lib/ext/nashorn.jar -Dnashorn.typeInfo.disabled=false --class-cache-size=0 --persistent-code-cache=false -scripting --log=time test/script/basic/compile-octane.js -- --iterations 1 2>&1 | grep "Class Installation"; done

...yields a nice improvement.

Before (parallel):
 [time] 'Class Installation' 4616 ms
 [time] 'Class Installation' 4563 ms
 [time] 'Class Installation' 4712 ms
 [time] 'Class Installation' 4637 ms
 [time] 'Class Installation' 4668 ms

After (sequential):
 [time] 'Class Installation' 2831 ms
 [time] 'Class Installation' 2755 ms
 [time] 'Class Installation' 2885 ms
 [time] 'Class Installation' 2816 ms
 [time] 'Class Installation' 2943 ms

One might consider a threshold for going to parallel version, but that threshold should be figured out experimentally.