RFR(L): 8215624: add parallel heap inspection support for jmap histo(G1)(Internet mail)
Hohensee, Paul
hohensee at amazon.com
Wed Jul 29 21:21:19 UTC 2020
A submit repo run with this succeeded, so afaic you're good to go. Stefan, you reviewed the GC part before, it'd be great if you could ok the final version.
Thanks,
Paul
On 7/29/20, 5:02 AM, "linzang(臧琳)" <linzang at tencent.com> wrote:
Upload a new change at http://cr.openjdk.java.net/~lzang/jmap-8214535/8215624/webrev_10/
It fix an issue of windows fail :
####################################
In heapInspect.cpp
- size_t HeapInspection::populate_table(KlassInfoTable* cit, BoolObjectClosure *filter, uint parallel_thread_num) {
+ uint HeapInspection::populate_table(KlassInfoTable* cit, BoolObjectClosure *filter, uint parallel_thread_num) {
####################################
In heapInspect.hpp
- size_t populate_table(KlassInfoTable* cit, BoolObjectClosure* filter = NULL, uint parallel_thread_num = 1) NOT_SERVICES_RETURN_(0);
+ uint populate_table(KlassInfoTable* cit, BoolObjectClosure* filter = NULL, uint parallel_thread_num = 1) NOT_SERVICES_RETURN_(0);
####################################
BRs,
Lin
On 2020/7/27, 11:26 AM, "linzang(臧琳)" <linzang at tencent.com> wrote:
I update a new change at http://cr.openjdk.java.net/~lzang/jmap-8214535/8215624/webrev_09
It includes a tiny fix of build failure on windows:
####################################
In attachListener.cpp:
- uint parallel_thread_num = MAX(1, (uint)os::initial_active_processor_count() * 3 / 8);
+ uint parallel_thread_num = MAX2<uint>(1, (uint)os::initial_active_processor_count() * 3 / 8);
####################################
BRs,
Lin
On 2020/7/23, 11:56 AM, "linzang(臧琳)" <linzang at tencent.com> wrote:
Hi Paul,
Thanks for your help, that all looks good to me.
Just 2 minor changes:
• delete the final return in ParHeapInspectTask::work, you mentioned it but seems not include in the webrev :-)
• delete a unnecessary blank line in heapInspect.cpp at merge_entry()
#########################################################################
--- old/src/hotspot/share/memory/heapInspection.cpp 2020-07-23 11:23:29.281666456 +0800
+++ new/src/hotspot/share/memory/heapInspection.cpp 2020-07-23 11:23:29.017666447 +0800
@@ -251,7 +251,6 @@
_size_of_instances_in_words += cie->words();
return true;
}
-
return false;
}
@@ -568,7 +567,6 @@
Atomic::add(&_missed_count, missed_count);
} else {
Atomic::store(&_success, false);
- return;
}
}
#########################################################################
Here is the webrev http://cr.openjdk.java.net/~lzang/jmap-8214535/8215624/webrev_08/
BRs,
Lin
---------------------------------------------
From: "Hohensee, Paul" <hohensee at amazon.com>
Date: Thursday, July 23, 2020 at 6:48 AM
To: "linzang(臧琳)" <linzang at tencent.com>, Stefan Karlsson <stefan.karlsson at oracle.com>, "serguei.spitsyn at oracle.com" <serguei.spitsyn at oracle.com>, David Holmes <david.holmes at oracle.com>, serviceability-dev <serviceability-dev at openjdk.java.net>, "hotspot-gc-dev at openjdk.java.net" <hotspot-gc-dev at openjdk.java.net>
Subject: RE: RFR(L): 8215624: add parallel heap inspection support for jmap histo(G1)(Internet mail)
Just small things.
heapInspection.cpp:
In ParHeapInspectTask::work, remove the final return statement and fix the following ‘}’ indent. I.e., replace
+ Atomic::store(&_success, false);
+ return;
+ }
with
+ Atomic::store(&_success, false);
+ }
In HeapInspection::heap_inspection, missed_count should be a uint to match other missed_count declarations, and should be initialized to the result of populate_table() rather than separately to 0.
attachListener.cpp:
In heap_inspection, initial_processor_count returns an int, so cast its result to a uint.
Similarly, parse_uintx returns a uintx, so cast its result (num) to uint when assigning to parallel_thread_num.
BasicJMapTest.java:
I took the liberty of refactoring testHisto*/histoToFile/testDump*/dump to remove redundant interposition methods and make histoToFile and dump look as similar as possible.
Webrev with the above changes in
http://cr.openjdk.java.net/~phh/8214535/webrev.01/
Thanks,
Paul
On 7/15/20, 2:13 AM, "linzang(臧琳)" <linzang at tencent.com> wrote:
Upload a new webrev at http://cr.openjdk.java.net/~lzang/jmap-8214535/8215624/webrev_07/
It fix a potential issue that unexpected number of threads maybe calculated for "parallel" option of jmap -histo in container.
As shown at http://cr.openjdk.java.net/~lzang/jmap-8214535/8215624/webrev_07-delta/src/hotspot/share/services/attachListener.cpp.udiff.html
############### attachListener.cpp ####################
@@ -252,11 +252,11 @@
static jint heap_inspection(AttachOperation* op, outputStream* out) {
bool live_objects_only = true; // default is true to retain the behavior before this change is made
outputStream* os = out; // if path not specified or path is NULL, use out
fileStream* fs = NULL;
const char* arg0 = op->arg(0);
- uint parallel_thread_num = MAX(1, os::processor_count() * 3 / 8); // default is less than half of processors.
+ uint parallel_thread_num = MAX(1, os::initial_active_processor_count() * 3 / 8); // default is less than half of processors.
if (arg0 != NULL && (strlen(arg0) > 0)) {
if (strcmp(arg0, "-all") != 0 && strcmp(arg0, "-live") != 0) {
out->print_cr("Invalid argument to inspectheap operation: %s", arg0);
return JNI_ERR;
}
###################################################
Thanks.
BRs,
Lin
On 2020/7/9, 3:22 PM, "linzang(臧琳)" <linzang at tencent.com> wrote:
Hi Paul,
Thanks for reviewing!
>>
>> I'd move all the argument parsing code to JMap.java and just pass the results to Hotspot. Both histo() in JMap.java and code in attachListener.* parse the command line arguments, though the code in histo() doesn't parse the argument to "parallel". I'd upgrade the code in histo() to do a complete parse and pass the option values to executeCommandForPid as before: there would just be more of them now. That would allow you to eliminate all the parsing code in attachListener.cpp as well as the change to arguments.hpp.
>>
The reason I made the change in Jmap.java that compose all arguments as 1 string , instead of passing 3 argments, is to avoid the compatibility issue, as we discussed in http://mail.openjdk.java.net/pipermail/serviceability-dev/2019-February/thread.html#27240. The root cause of the compatibility issue is because max argument count in HotspotVirtualMachineImpl.java and attachlistener.cpp need to be enlarged (changes like http://hg.openjdk.java.net/jdk/jdk/rev/e7cf035682e3#l2.1) when jmap has more than 3 arguments. But if user use an old jcmd/jmap tool, it may stuck at socket read(), because the "max argument count" don't match.
I re-checked this change, the argument count of jmap histo is equal to 3 (live, file, parallel), so it can work normally even without the change of passing argument. But I think we have to face the problem if more arguments is added in jcmd alike tools later, not sure whether it should be sloved (or a workaround) in this changeset.
And here are the lastest webrev and delta:
http://cr.openjdk.java.net/~lzang/jmap-8214535/8215624/webrev_06/
http://cr.openjdk.java.net/~lzang/jmap-8214535/8215624/webrev_06-delta/
Cheers,
Lin
On 2020/7/7, 5:57 AM, "Hohensee, Paul" <hohensee at amazon.com> wrote:
I'd like to see this feature added. :)
The CSR looks good, as does the basic parallel inspection algorithm. Stefan's done the GC part, so I'll stick to the non-GC part (fwiw, the GC part lgtm).
I'd move all the argument parsing code to JMap.java and just pass the results to Hotspot. Both histo() in JMap.java and code in attachListener.* parse the command line arguments, though the code in histo() doesn't parse the argument to "parallel". I'd upgrade the code in histo() to do a complete parse and pass the option values to executeCommandForPid as before: there would just be more of them now. That would allow you to eliminate all the parsing code in attachListener.cpp as well as the change to arguments.hpp.
heapInspection.hpp:
_shared_miss_count (s/b _missed_count, see below) isn't a size, so it should be a uint instead of a size_t. Same with the new parallel_thread_num argument to heap_inspection() and populate_table().
Comment copy-edit:
+// Parallel heap inspection task. Parallel inspection can fail due to
+// a native OOM when allocating memory for TL-KlassInfoTable.
+// _success will be set false on an OOM, and serial inspection tried.
_shared_miss_count should be _missed_count to match the missed_count() getter, or rename missed_count() to be shared_miss_count(). Whichever way you go, the field type should match the getter result type: uint is reasonable.
heapInspection.cpp:
You might use ResourceMark twice in populate_table, separately for the parallel attempt and the serial code. If the parallel attempt fails and available memory is low, it would be good to clean up the memory used by the parallel attempt before doing the serial code.
Style nit in KlassInfoTable::merge_entry(). I'd line up the definitions of k and elt, so "k" is even with "elt". And, because it's two lines shorter, I'd replace
+ } else {
+ return false;
+ }
with
+ return false;
KlassInfoTableMergeClosure.is_success() should be just success() (i.e., no "is_" prefix) because it's a getter.
I'd reorganize the code in populate_table() to make it more clear, vis (I changed _shared_missed_count to _missed_count)
+ if (cit.allocation_failed()) {
+ // fail to allocate memory, stop parallel mode
+ Atomic::store(&_success, false);
+ return;
+ }
+ RecordInstanceClosure ric(&cit, _filter);
+ _poi->object_iterate(&ric, worker_id);
+ missed_count = ric.missed_count();
+ {
+ MutexLocker x(&_mutex);
+ merge_success = _shared_cit->merge(&cit);
+ }
+ if (merge_success) {
+ Atomic::add(&_missed_count, missed_count);
+ else {
+ Atomic::store(&_success, false);
+ }
Thanks,
Paul
On 6/29/20, 7:20 PM, "linzang(臧琳)" <linzang at tencent.com> wrote:
Dear All,
Sorry to bother again, I just want to make sure that is this change worth to be continue to work on? If decision is made to not. I think I can drop this work and stop asking for help reviewing...
Thanks for all your help about reviewing this previously.
BRs,
Lin
On 2020/5/9, 3:47 PM, "linzang(臧琳)" <linzang at tencent.com> wrote:
Dear All,
May I ask your help again for review the latest change? Thanks!
BRs,
Lin
On 2020/4/28, 1:54 PM, "linzang(臧琳)" <linzang at tencent.com> wrote:
Hi Stefan,
>> - Adding Atomic::load/store.
>> - Removing the time measurement in the run_task. I renamed G1's function
>> to run_task_timed. If we need this outside of G1, we can rethink the API
>> at that point.
>> - ZGC style cleanups
Thanks for revising the patch, they are all good to me, and I have made a tiny change based on it:
http://cr.openjdk.java.net/~lzang/jmap-8214535/8215624/webrev_04/
http://cr.openjdk.java.net/~lzang/jmap-8214535/8215624/webrev_04-delta/
it reduce the scope of mutex in ParHeapInspectTask, and delete unnecessary comments.
BRs,
Lin
On 2020/4/27, 4:34 PM, "Stefan Karlsson" <stefan.karlsson at oracle.com> wrote:
Hi Lin,
On 2020-04-26 05:10, linzang(臧琳) wrote:
> Hi Stefan and Paul，
> I have made a new patch based on your comments and Stefan's Poc code:
> Webrev: http://cr.openjdk.java.net/~lzang/jmap-8214535/8215624/webrev_03/
> Delta(based on Stefan's change:) : http://cr.openjdk.java.net/~lzang/jmap-8214535/8215624/webrev_03-delta/webrev_03-delta/
Thanks for providing a delta patch. It makes it much easier to look at,
and more likely for reviewers to continue reviewing.
I'm going to continue focusing on the GC parts, and leave the rest to
others to review.
>
> And Here are main changed I made and want to discuss with you:
> 1. changed"parallelThreadNum=" to "parallel=" for jmap -histo options.
> 2. Add logic to test where parallelHeapInspection is fail, in heapInspection.cpp
> This is because the parHeapInspectTask create thread local KlassInfoTable in it's work() method, and this may fail because of native OOM, in this case, the parallel should fail and serial heap inspection can be tried.
> One more thing I want discuss with you is about the member "_success" of parHeapInspectTask, when native OOM happenes, it is set to false. And since this "set" operation can be conducted in multiple threads, should it be atomic ops? IMO, this is not necessary because "_success" can only be set to false, and there is no way to change it from back to true after the ParHeapInspectTask instance is created, so it is save to be non-atomic, do you agree with that?
In these situations you should be using the Atomic::load/store
primitives. We're moving toward a later C++ standard were data races are
considered undefined behavior.
> 3. make CollectedHeap::run_task() be an abstract virtual func, so that every subclass of collectedHeap should support it, so later implementation of new collectedHeap will not miss the "parallel" features.
> The problem I want to discuss with you is about epsilonHeap and SerialHeap, as they may not need parallel heap iteration, so I only make task->work(0), in case the run_task() is invoked someway in future. Another way is to left run_task() unimplemented, which one do you think is better?
I don't have a strong opinion about this.
And also please help take a look at the zHeap, as there is a class
zTask that wrap the abstractGangTask, and the collectedHeap::run_task()
only accept AbstraceGangTask* as argument, so I made a delegate class
to adapt it , please see src/hotspot/share/gc/z/zHeap.cpp.
>
> There maybe other better ways to sovle the above problems, welcome for any comments, Thanks!
I've created a few cleanups and changes on top of your latest patch:
https://cr.openjdk.java.net/~stefank/8215624/webrev.02.delta
https://cr.openjdk.java.net/~stefank/8215624/webrev.02
- Adding Atomic::load/store.
- Removing the time measurement in the run_task. I renamed G1's function
to run_task_timed. If we need this outside of G1, we can rethink the API
at that point.
- ZGC style cleanups
Thanks,
StefanK
>
> BRs,
> Lin
>
> On 2020/4/23, 11:08 AM, "linzang(臧琳)" <linzang at tencent.com> wrote:
>
> Thanks Paul! I agree with using "parallel", will make the update in next patch, Thanks for help update the CSR.
>
> BRs,
> Lin
>
> On 2020/4/23, 4:42 AM, "Hohensee, Paul" <hohensee at amazon.com> wrote:
>
> For the interface, I'd use "parallel" instead of "parallelThreadNum". All the other options are lower case, and it's a lot easier to type "parallel". I took the liberty of updating the CSR. If you're ok with it, you might want to change variable names and such, plus of course JMap.usage.
>
> Thanks,
> Paul
>
> On 4/22/20, 2:29 AM, "serviceability-dev on behalf of linzang(臧琳)" <serviceability-dev-bounces at openjdk.java.net on behalf of linzang at tencent.com> wrote:
>
> Dear Stefan,
>
> Thanks a lot! I agree with you to decouple the heap inspection code with GC's.
> I will start from your POC code, may discuss with you later.
>
>
> BRs,
> Lin
>
> On 2020/4/22, 5:14 PM, "Stefan Karlsson" <stefan.karlsson at oracle.com> wrote:
>
> Hi Lin,
>
> I took a look at this earlier and saw that the heap inspection code is
> strongly coupled with the CollectedHeap and G1CollectedHeap. I'd prefer
> if we'd abstract this away, so that the GCs only provide a "parallel
> object iteration" interface, and the heap inspection code is kept elsewhere.
>
> I started experimenting with doing that, but other higher-priority (to
> me) tasks have had to take precedence.
>
> I've uploaded my work-in-progress / proof-of-concept:
> https://cr.openjdk.java.net/~stefank/8215624/webrev.01.delta/
> https://cr.openjdk.java.net/~stefank/8215624/webrev.01/
>
> The current code doesn't handle the lifecycle (deletion) of the
> ParallelObjectIterators. There's also code left unimplemented in around
> CollectedHeap::run_task. However, I think this could work as a basis to
> pull out the heap inspection code out of the GCs.
>
> Thanks,
> StefanK
>
> On 2020-04-22 02:21, linzang(臧琳) wrote:
> > Dear all,
> > May I ask you help to review? This RFR has been there for quite a while.
> > Thanks!
> >
> > BRs,
> > Lin
> >
> > > On 2020/3/16, 5:18 PM, "linzang(臧琳)" <linzang at tencent.com> wrote:>
> >
> >> Just update a new path, my preliminary measure show about 3.5x speedup of jmap histo on a nearly full 4GB G1 heap (8-core platform with parallel thread number set to 4).
> >> webrev: http://cr.openjdk.java.net/~lzang/jmap-8214535/8215624/webrev_02/
> >> bug: https://bugs.openjdk.java.net/browse/JDK-8215624
> >> CSR: https://bugs.openjdk.java.net/browse/JDK-8239290
> >> BRs,
> >> Lin
> >> > On 2020/3/2, 9:56 PM, "linzang(臧琳)" <linzang at tencent.com> wrote:
> >> >
> >> > Dear all,
> >> > Let me try to ease the reviewing work by some explanation :P
> >> > The patch's target is to speed up jmap -histo for heap iteration, from my experience it is necessary for large heap investigation. E.g in bigData scenario I have tried to conduct jmap -histo against 180GB heap, it does take quite a while.
> >> > And if my understanding is corrent, even the jmap -histo without "live" option does heap inspection with heap lock acquired. so it is very likely to block mutator thread in allocation-sensitive scenario. I would say the faster the heap inspection does, the shorter the mutator be blocked. This is parallel iteration for jmap is necessary.
> >> > I think the parallel heap inspection should be applied to all kind of heap. However, consider the heap layout are different for GCs, much time is required to understand all kinds of the heap layout to make the whole change. IMO, It is not wise to have a huge patch for the whole solution at once, and it is even harder to review it. So I plan to implement it incrementally, the first patch (this one) is going to confirm the implemention detail of how jmap accept the new option, passes it to attachListener of the jvm process and then how to make the parallel inspection closure be generic enough to make it easy to extend to different heap layout. And also how to implement the heap inspection in specific gc's heap. This patch use G1's heap as the begining.
> >> > This patch actually do several things:
> >> > 1. Add an option "parallelThreadNum=<N>" to jmap -histo, the default behavior is to set N to 0, means let's JVM decide how many threads to use for heap inspection. Set this option to 1 will disable parallel heap inspection. (more details in CSR: https://bugs.openjdk.java.net/browse/JDK-8239290)
> >> > 2. Make a change in how Jmap passing arguments, changes in http://cr.openjdk.java.net/~lzang/jmap-8214535/8215624/webrev_01/src/jdk.jcmd/share/classes/sun/tools/jmap/JMap.java.udiff.html, originally it pass options as separate arguments to attachListener, this patch change to that all options be compose to a single string. So the arg_count_max in attachListener.hpp do not need to be changed, and hence avoid the compatibility issue, as disscussed at https://mail.openjdk.java.net/pipermail/serviceability-dev/2019-March/027334.html
> >> > 3. Add an abstract class ParHeapInspectTask in heapInspection.hpp / heapInspection.cpp, It's work(uint worker_id) method prepares the data structure (KlassInfoTable) need for every parallel worker thread, and then call do_object_iterate_parallel() which is heap specific implementation. I also added some machenism in KlassInfoTable to support parallel iteration, such as merge().
> >> > 4. In specific heap (G1 in this patch), create a subclass of ParHeapInspectTask, implement the do_object_iterate_parallel() for parallel heap inspection. For G1, it simply invoke g1CollectedHeap's object_iterate_parallel().
> >> > 5. Add related test.
> >> > 6. it may be easy to extend this patch for other kinds of heap by creating subclass of ParHeapInspectTask and implement the do_object_iterate_parallel().
> >> >
> >> > Hope these info could help on code review and initate the discussion :-)
> >> > Thanks!
> >> >
> >> > BRs,
> >> > Lin
> >> > >On 2020/2/19, 9:40 AM, "linzang(臧琳)" <linzang at tencent.com> wrote:.
> >> > >
> >> > > Re-post this RFR with correct enhancement number to make it trackable.
> >> > > please ignore the previous wrong post. sorry for troubles.
> >> > >
> >> > > webrev: http://cr.openjdk.java.net/~lzang/jmap-8214535/8215624/webrev_01/
> >> > > Hi bug: https://bugs.openjdk.java.net/browse/JDK-8215624
> >> > > CSR: https://bugs.openjdk.java.net/browse/JDK-8239290
> >> > > --------------
> >> > > Lin
> >> > > >Hi Lin,
> > > > > >
> >> > > >Could you, please, re-post your RFR with the right enhancement number in
> >> > > >the message subject?
> >> > > >It will be more trackable this way.
> >> > > >
> >> > > >Thanks,
> >> > > >Serguei
> >> > > >
> >> > > >
> >> > > >On 2/17/20 10:29 PM, linzang(臧琳) wrote:
> >> > > >> Dear David,
> >> > > >> Thanks a lot!
> >> > > >> I have updated the refined code to http://cr.openjdk.java.net/~lzang/jmap-8214535/8215264/webrev_01/.
> >> > > >> IMHO the parallel heap inspection can be extended to all kinds of heap as long as the heap layout can support parallel iteration.
> >> > > >> Maybe we can firstly use this webrev to discuss how to implement it, because I am not sure my current implementation is an appropriate way to communicate with collectedHeap, then we can extend the solution to other kinds of heap.
> >> > > >>
> >> > > >> Thanks,
> >> > > >> --------------
> >> > > >> Lin
> >> > > >>> Hi Lin,
> >> > > >>>
> >> > > >>> Adding in hotspot-gc-dev as they need to see how this interacts with GC
> >> > > >>> worker threads, and whether it needs to be extended beyond G1.
> >> > > >>>
> >> > > >>> I happened to spot one nit when browsing:
> >> > > >>>
> >> > > >>> src/hotspot/share/gc/shared/collectedHeap.hpp
> >> > > >>>
> >> > > >>> + virtual bool run_par_heap_inspect_task(KlassInfoTable* cit,
> >> > > >>> + BoolObjectClosure* filter,
> >> > > >>> + size_t* missed_count,
> >> > > >>> + size_t thread_num) {
> >> > > >>> + return NULL;
> >> > > >>>
> >> > > >>> s/NULL/false/
> >> > > >>>
> >> > > >>> Cheers,
> >> > > >>> David
> > > > > >>>
> >> > > >>> On 18/02/2020 2:15 pm, linzang(臧琳) wrote:
> >> > > >>>> Dear All,
> >> > > >>>> May I ask your help to review the follow changes:
> >> > > >>>> webrev:
> >> > > >>>> http://cr.openjdk.java.net/~lzang/jmap-8214535/8215264/webrev_00/
> >> > > >>>> bug: https://bugs.openjdk.java.net/browse/JDK-8215624
> >> > > >>>> related CSR: https://bugs.openjdk.java.net/browse/JDK-8239290
> >> > > >>>> This patch enable parallel heap inspection of G1 for jmap histo.
> >> > > >>>> my simple test shown it can speed up 2x of jmap -histo with
> >> > > >>>> parallelThreadNum set to 2 for heap at ~500M on 4-core platform.
> >> > > >>>>
> >> > > >>>> ------------------------------------------------------------------------
> >> > > >>>> BRs,
> >> > > >>>> Lin
> >> > > >> >
> >> > > >
> >
> >
> >
>
>
>
>
>
>
More information about the hotspot-gc-dev
mailing list