RFR: 8242440: use separate, destroyable JavaVM instances per libgraal compiler thread [v3]
Vladimir Kozlov
kvn at openjdk.java.net
Wed Apr 20 22:23:30 UTC 2022
On Tue, 19 Apr 2022 10:17:14 GMT, Doug Simon <dnsimon at openjdk.org> wrote:
>> Currently, libgraal runs in a single, permanent JavaVM instance loaded from libjvmcicompiler.so. This poses 2 problems:
>>
>> 1. The memory used for libgraal is never released, even when the libgraal compiler queues are idle.
>> 2. Since all libgraal HotSpot compiler threads share the same SVM heap, libgraal compilation is effectively paused when a GC occurs (SVM currently has a stop the world collector). The more compiler threads there are, the more often this happens.
>>
>> This PR implements the following solution to these problems:
>> 1. Create a new JavaVM instance per libgraal compiler thread (i.e. a separate SVM isolate per thread). This prevents GC in one libgraal thread from impacting another libgraal thread.
>> 2. Destroy the JavaVM instances when a libgraal compiler thread becomes idle and create a new one when subsequent compilation requests arrive.
>>
>> Most of the changes are in JVMCI specific files. The most significant change to shared code is the addition of the `JavaThread:: _libjvmci_runtime` field. This is required as any thread (not just a `CompilerThread`) can be associated with a specific `JVMCIRuntime` object.
>
> Doug Simon has updated the pull request incrementally with one additional commit since the last revision:
>
> removed (incorrect) volatile modifier from JVMCIRuntime::_shared_library_javavm
src/hotspot/share/jvmci/jvmciCompiler.cpp line 172:
> 170: // Don't detach JVMCI compiler threads from their JVMCI
> 171: // runtime during the VM startup grace period
> 172: if (runtime != nullptr && delay > 0 && tty->time_stamp().milliseconds() > DEFAULT_COMPILER_IDLE_DELAY) {
Can it be more coordinated by some status instead of time? Time depends on machine you run and does not guarantee right case. I could be done in separate changes.
-------------
PR: https://git.openjdk.java.net/jdk/pull/8262
More information about the hotspot-dev
mailing list