The Jestream Win32 benchmarks for read, write and parse text have dropped to level worse then 1.1.7 (our worst baseline.)  We have an unfortunate gap in our measurements due to other bug issues but we know it happend in the 12-19 to 12-28 time frame.

Its is worth noting that the copy benchmark seems un-effected. We guss that the key difference is that all the other benchamrks involve soem form of parsign of the bianry stream of bits while copy just mvoes bits from one file to another.

Measurement data is viewable using our Java Webstart (1.0) acessible applet at
http://javaweb.eng/engineering/jdk-performance/reports/NT4/jetstream.jnlp

* Added 1/9/01 eileeny

performance is also affected on linux platform.  With -server flag enabled, performance is worse that jdk1.2.2 without a jit.  copy benchmark is also affected.  see results at:
http://javaweb.eng/engineering/jdk-performance/engineering/linux/jetstream-comparison.jnlp

to recreate the bug:
java -cp /net/perf-lnx/export/performance perfteam.jetstream.JetStream -testfile IOTests.txt

* Added 1/10/01 zzhang

similiar performance regression has also been found on Solaris/sparc, please
check:  http://javaweb/engineering/jdk-performance/reports/solaris/

N/A

This is a combination of several problems, the most serious of which is the
fact that the HotSpot compilers do not yet inline or optimize sun.misc.Unsafe
operations (4411596, 4411600, 4411601).  Further investigation is in progress.

--   xxxxx@xxxxx   2001/2/4


The redesign of the Charset API (4503732) brought the JetStream numbers down to
more reasonable levels (see comments section for measurements).  Further
improvement is possible but the drop no longer qualifies as "precipitous",
hence I'm closing this bug.

--   xxxxx@xxxxx   2001/10/2