On Tue, 20 Feb 2024 07:45:59 GMT, Thomas Stuefe <stu...@openjdk.org> wrote:
>> DeCapo Benchmark Results (3 runs) : >> >> Before : >> ===== DaCapo 9.12 h2 PASSED in 281402 msec ===== >> ===== DaCapo 9.12 h2 PASSED in 269818 msec ===== >> ===== DaCapo 9.12 h2 PASSED in 279279 msec ===== >> >> After: >> ===== DaCapo 9.12 h2 PASSED in 279192 msec ===== >> ===== DaCapo 9.12 h2 PASSED in 269769 msec ===== >> ===== DaCapo 9.12 h2 PASSED in 271577 msec ===== >> >> Environmental variables LDR_CNTRL and MALLOCOPTIONS has caused >> test/jdk/java/lang/ProcessBuilder/Basic.java failure. >> Additional environmental variables has to removed from >> removeAixExpectedVars(). >> >> JBS Issue : [JDK-8252802](https://bugs.openjdk.org/browse/JDK-8252802) > > I don't think this works as intended. > > IIRC, both variables must be set *before process invocation*. Setting them > inside the launcher will only affect child processes. The dacapo results > posted seem to support this, they seem random-ish to me. > > Note that we already use 64KB pages for all large memory regions (everything > that goes through `os::reserve_memory`). So, while the value of LDRCNTRL is > not nil, it is very diminished. Mostly just the C-Heap would be affected. > > But I would not change defaults for options like these anyway, especially not > hard-coded. These knobs have far-ranging implications. If we want that, we > need to investigate carefully. > > (for example: would using 64KB pages increase average heap size? Possibly, > depending on the implementation. But AIX still has this problem where the > heap can only live in the break and can bump against low-hanging regions. So, > AIX is especially vulnerable against any changes that increase average heap > usage) Thanks @tstuefe for the response. Closing the PR since it is not advisable to do. ------------- PR Comment: https://git.openjdk.org/jdk/pull/17906#issuecomment-1953721220