RFR: 8252842: Extend jmap to support parallel heap dump

[Lin Zang]

On Wed, 10 Feb 2021 02:00:11 GMT, Hamlin Li <mli at openjdk.org> wrote:

>>>  IMO, there can be two ways which may help solve the compatibility issue of introducing "parallel" option:
>>> 
>>>     * unblock socket with timeout. In this way, old Jmap could work with new JDK, but it maybe complicated because of different implementation for different OS.
>>> 
>>>     * Two times communication as you mentioned,  it sounds a reasonable solution, I will work out a prototype and then discuss with you .
>>> 
>> A 3rd would be to create a new `dumpheap` command for the Attach API, maybe `dumpheap_parallel`. If you try using it with an old JVM, it should immediately return an error:
>> 
>>         st.print("Operation %s not recognized!", op->name());
>>         res = JNI_ERR;
>> If it does, you can fallback to the regular `dumpheap` command. If an old JVM attaches and uses `dumpheap`, you can choose to make it parallel by default if you wish. A slightly different variant of this idea is have something like `dumpheap_extraargs` or `dumpheap_ext`, leaving room to add even more args in the future (possibly a single argument with a list of args as you suggested before)
>> 
>> BTW, one thing to keep in mind that whatever we do with `dumpheap`, the jcmd version will support fine control over parallelism. So for that reason it wouldn't be so bad to just have `dumpheap` and `inspectheap` always use some default parallelism, and allow tuning it with the jcmd versions. But since the `inspectheap` parallel support already went into 16, that might be hard to undo.
>
>> BTW, one thing to keep in mind that whatever we do with `dumpheap`, the jcmd version will support fine control over parallelism. So for that reason it wouldn't be so bad to just have `dumpheap` and `inspectheap` always use some default parallelism, and allow tuning it with the jcmd versions. But since the `inspectheap` parallel support already went into 16, that might be hard to undo.
> 
> Hi Chris, Lin,
> IMHO, 16 is not an LTS, maybe it's better for us to adjust the behavior of jmap -histo in 17 which is LTS? I just filed https://bugs.openjdk.java.net/browse/JDK-8261482 to address the issue if you both think it's OK to undo it.
> Thanks.

Hi Chris, 
I have update the PR to add a new dumpheapext command that could handle more arguments passed from JMap.java. 
And I have tested with "jmap -dump:file=dump.bin,noparallel [pid]", with this change, new jmap targeting on old hotspot would print: 

Exception in thread "main" com.sun.tools.attach.AttachOperationFailedException: Operation dumpheapext not recognized!
	at jdk.attach/sun.tools.attach.VirtualMachineImpl.execute(VirtualMachineImpl.java:227)
	at jdk.attach/sun.tools.attach.HotSpotVirtualMachine.executeCommand(HotSpotVirtualMachine.java:309)
	at jdk.jcmd/sun.tools.jmap.JMap.executeCommandForPid(JMap.java:133)
	at jdk.jcmd/sun.tools.jmap.JMap.dump(JMap.java:262)
	at jdk.jcmd/sun.tools.jmap.JMap.main(JMap.java:114)

IMH, one problem is that it doesn't give out the reason clearly that "dumpheapext" is actually caused by new option "noparallel". But I think having this kind of error message is acceptable. Otherwise it seems the choice we left is remove the "parallel/noparallel" for Jmap. What do you think?

-------------

PR: https://git.openjdk.java.net/jdk/pull/2261