Following up on why tier1 exhibits low paralellism when run on modern heavy-threaded desktop. On my TR 3970X, jdk:tier1 takes about 400 seconds, and there is a significantly large phase during which CPU load drops significantly.

One of the reasons is JDK-8058204 that put java/util/stream tests to run in exclusive mode. Dropping java/util/stream from there improves the test times significantly (and CPU loads now make sense too):

diff -r 216c6baa0564 test/jdk/TEST.ROOT
--- a/test/jdk/TEST.ROOT Sat Jun 20 15:11:19 2020 +0800
+++ b/test/jdk/TEST.ROOT Mon Jun 22 11:24:26 2020 +0200
@@ -27,5 +27,5 @@
 sun/tools/jstatd sun/tools/jcmd sun/tools/jhsdb sun/tools/jhsdb/heapconfig \
 sun/tools/jinfo sun/tools/jmap sun/tools/jps sun/tools/jstack sun/tools/jstat \
-com/sun/tools/attach sun/security/mscapi java/util/stream java/util/Arrays/largeMemory \
+com/sun/tools/attach sun/security/mscapi java/util/Arrays/largeMemory \
 java/util/BitSet/stream javax/rmi
 
$ time CONF=linux-x86_64-server-release make images run-test TEST=java/util/stream/

# before: ~17x parallelism; mostly due to parallel javac compiles
real 4m23.668s
user 75m47.699s
sys 2m50.316s

# after: ~43x parallelism, limited by test tails
real 1m32.027s
user 61m59.474s
sys 3m20.213s

In other words, we are losing roughly 180s out of 400s jdk:tier1 for this, about 45% of the test time. This is way beyond the overhead JDK-8058204 envisioned originally ("I doubt it would increase the overall test run time, perhaps by 1% to 3%."). Observe:

$ time CONF=linux-x86_64-server-release make images run-test TEST=jdk:tier1

# before: ~25x parallelism
real 6m37.627s
user 158m32.437s
sys 8m12.700s

# after: ~44x parallelism
real 3m43.018s
user 159m47.390s
sys 8m56.156s

This makes jdk:tier1 almost 1.8x faster! fastdebug has roughly the same relative speedup.