[prev in list] [next in list] [prev in thread] [next in thread]
List: openjdk-hotspot-runtime-dev
Subject: Re: RFR(L) 8153224 Monitor deflation prolong safepoints (CR6/v2.06/9-for-jdk14)
From: "Daniel D. Daugherty" <daniel.daugherty () oracle ! com>
Date: 2019-08-28 21:02:16
Message-ID: e3ff1c7f-9220-a1a6-e3e7-00dcc24a9bcf () oracle ! com
[Download RAW message or body]
Greetings,
The Async Monitor Deflation project has rebased to JDK14 so it's time
for our first code review in that new context!!
I've been focused on changing the monitor list management code to be
lock-free in order to make SPECjbb2015 happier. Of course with a change
like that, it takes a while to chase down all the new and wonderful
races. At this point, I have the code back to the same stability that
I had with CR5/v2.05/8-for-jdk13.
To lay the ground work for this round of review, I pushed the following
two fixes to jdk/jdk earlier today:
JDK-8230184 rename, whitespace, indent and comments changes in
preparation
for lock free Monitor lists
https://bugs.openjdk.java.net/browse/JDK-8230184
JDK-8230317 serviceability/sa/ClhsdbPrintStatics.java fails after
8230184
https://bugs.openjdk.java.net/browse/JDK-8230317
I have attached the list of fixes from CR5 to CR6 instead of putting
in the main body of this email.
Main bug URL:
JDK-8153224 Monitor deflation prolong safepoints
https://bugs.openjdk.java.net/browse/JDK-8153224
The project is currently baselined on jdk-14+11 plus the fixes for
JDK-8230184 and JDK-8230317.
Here's the full webrev URL for those folks that want to see all of the
current Async Monitor Deflation code in one go (v2.06 full):
http://cr.openjdk.java.net/~dcubed/8153224-webrev/9-for-jdk14.v2.06.full/
The primary focus of this review cycle is on the lock-free Monitor List
management changes so here's a webrev for just that patch (v2.06c):
http://cr.openjdk.java.net/~dcubed/8153224-webrev/9-for-jdk14.v2.06c.inc/
The secondary focus of this review cycle is on the bug fixes that have
been made since CR5/v2.05/8-for-jdk13 so here's a webrev for just that
patch (v2.06b):
http://cr.openjdk.java.net/~dcubed/8153224-webrev/9-for-jdk14.v2.06b.inc/
The third and final bucket for this review cycle is the rename, whitespace,
indent and comments changes made in preparation for lock free Monitor list
management. Almost all of that was extracted into JDK-8230184 for the
baseline so this bucket now has just a few comment changes relative to
CR5/v2.05/8-for-jdk13. Here's a webrev for the remainder (v2.06a):
http://cr.openjdk.java.net/~dcubed/8153224-webrev/9-for-jdk14.v2.06a.inc/
Some folks might want to see just what has changed since the last review
cycle so here's a webrev for that (v2.06 inc):
http://cr.openjdk.java.net/~dcubed/8153224-webrev/9-for-jdk14.v2.06.inc/
Last, but not least, some folks might want to see the code before the
addition of lock-free Monitor List management so here's a webrev for
that (v2.00 -> v2.05):
http://cr.openjdk.java.net/~dcubed/8153224-webrev/9-for-jdk14.v2.05.inc/
The OpenJDK wiki will need minor updates to match the CR6 changes:
https://wiki.openjdk.java.net/display/HotSpot/Async+Monitor+Deflation
but that should only be changes to describe per-thread list async monitor
deflation being done by the ServiceThread.
(I did update the OpenJDK wiki for the CR5 changes back on 2019.08.14)
This version of the patch has been thru Mach5 tier[1-8] testing on
Oracle's usual set of platforms. It has also been through my usual set
of stress testing on Linux-X64, macOSX and Solaris-X64.
I did a bunch of SPECjbb2015 testing in Oracle's Aurora Performance lab
using using their tuned SPECjbb2015 Linux-X64 G1 configs. This was using
this patch baselined on jdk-13+31 (for stability):
hbIR hbIR
(max attempted) (settled) max-jOPS critical-jOPS runtime
--------------- --------- -------- ------------- -------
34282.00 28837.20 27905.20 19817.40 3658.10 base
34965.70 29798.80 27814.90 19959.00 3514.60 v2.06d
34282.00 29100.70 28042.50 19577.00 3701.90
v2.06d_off
34282.00 29218.50 27562.80 19397.30 3657.60
v2.06d_ocache
34965.70 29838.30 26512.40 19170.60 3569.90 v2.05
34282.00 28926.10 27734.00 19835.10 3588.40 v2.05_off
The "off" configs are with -XX:-AsyncDeflateIdleMonitors specified and
the "ocache" config is with 128 byte cache line sizes instead of 64 byte
cache lines sizes. "v2.06d" is the last set of changes that I made before
those changes were distributed into the "v2.06a", "v2.06b" and "v2.06c"
buckets for this review recycle.
Thanks, in advance, for any questions, comments or suggestions.
Dan
On 7/11/19 3:49 PM, Daniel D. Daugherty wrote:
> Greetings,
>
> I've been focused on chasing down and fixing the rare test failures
> that only pop up rarely. So this round is primarily fixes for races
> with a few additional fixes that came from Karen's review of CR4.
> Thanks Karen!
>
> I have attached the list of fixes from CR4 to CR5 instead of putting
> in the main body of this email.
>
> Main bug URL:
>
> JDK-8153224 Monitor deflation prolong safepoints
> https://bugs.openjdk.java.net/browse/JDK-8153224
>
> The project is currently baselined on jdk-13+29. This will likely be
> the last JDK13 baseline for this project and I'll roll to the JDK14
> (jdk/jdk) repo soon...
>
> Here's the full webrev URL:
>
> http://cr.openjdk.java.net/~dcubed/8153224-webrev/8-for-jdk13.full/
>
> Here's the incremental webrev URL:
>
> http://cr.openjdk.java.net/~dcubed/8153224-webrev/8-for-jdk13.inc/
>
> I have not yet checked the OpenJDK wiki to see if it needs any updates
> to match the CR5 changes:
>
> https://wiki.openjdk.java.net/display/HotSpot/Async+Monitor+Deflation
>
> (I did update the OpenJDK wiki for the CR4 changes back on 2019.06.26)
>
> This version of the patch has been thru Mach5 tier[1-3] testing on
> Oracle's usual set of platforms. Mach5 tier[4-6] is running now and
> Mach5 tier[78] will follow. I'll kick off the usual stress testing
> on Linux-X64, macOSX and Solaris-X64 as those machines become available.
> Since I haven't made any performance changes in this round, I'll only
> be running SPECjbb2015 to gather the latest monitorinflation logs.
>
> Next up:
>
> - We're still seeing 4-5% lower performance with SPECjbb2015 on
> Linux-X64 and we've determined that some of that comes from
> contention on the gListLock. So I'm going to investigate removing
> the gListLock. Yes, another lock free set of changes is coming!
> - Of course, going lock free often causes new races and new failures
> so that's a good reason for make those changes isolated in their
> own round (and not holding up CR5/v2.05/8-for-jdk13 anymore).
> - I finally have a potential fix for the Win* failure with
> gc/g1/humongousObjects/TestHumongousClassLoader.java
> but I haven't run it through Mach5 yet so it'll be in the next round.
> - Some RTM tests were recently re-enabled in Mach5 and I'm seeing some
> monitor related failures there. I suspect that I need to go take a
> look at the C2 RTM macro assembler code and look for things that might
> conflict if Async Monitor Deflation. If you're interested in that kind
> of issue, then see the macroAssembler_x86.cpp sanity check that I
> added in this round!
>
> Thanks, in advance, for any questions, comments or suggestions.
>
> Dan
>
>
> On 5/26/19 8:30 PM, Daniel D. Daugherty wrote:
>> Greetings,
>>
>> I have a fix for an issue that came up during performance testing.
>> Many thanks to Robbin for diagnosing the issue in his SPECjbb2015
>> experiments.
>>
>> Here's the list of changes from CR3 to CR4. The list is a bit
>> verbose due to the complexity of the issue, but the changes
>> themselves are not that big.
>>
>> Functional:
>> - Change SafepointSynchronize::is_cleanup_needed() from calling
>> ObjectSynchronizer::is_cleanup_needed() to calling
>> ObjectSynchronizer::is_safepoint_deflation_needed():
>> - is_safepoint_deflation_needed() returns the result of
>> monitors_used_above_threshold() for safepoint based
>> monitor deflation (!AsyncDeflateIdleMonitors).
>> - For AsyncDeflateIdleMonitors, it only returns true if
>> there is a special deflation request, e.g., System.gc()
>> - This solves a bug where there are a bunch of Cleanup
>> safepoints that simply request async deflation which
>> keeps the async JavaThreads from making progress on
>> their async deflation work.
>> - Add AsyncDeflationInterval diagnostic option. Description:
>> Async deflate idle monitors every so many milliseconds when
>> MonitorUsedDeflationThreshold is exceeded (0 is off).
>> - Replace ObjectSynchronizer::gOmShouldDeflateIdleMonitors() with
>> ObjectSynchronizer::is_async_deflation_needed():
>> - is_async_deflation_needed() returns true when
>> is_async_cleanup_requested() is true or when
>> monitors_used_above_threshold() is true (but no more often than
>> AsyncDeflationInterval).
>> - if AsyncDeflateIdleMonitors Service_lock->wait() now waits for
>> at most GuaranteedSafepointInterval millis:
>> - This allows is_async_deflation_needed() to be checked at
>> the same interval as GuaranteedSafepointInterval.
>> (default is 1000 millis/1 second)
>> - Once is_async_deflation_needed() has returned true, it
>> generally cannot return true for AsyncDeflationInterval.
>> This is to prevent async deflation from swamping the
>> ServiceThread.
>> - The ServiceThread still handles async deflation of the global
>> in-use list and now it also marks JavaThreads for async deflation
>> of their in-use lists.
>> - The ServiceThread will check for async deflation work every
>> GuaranteedSafepointInterval.
>> - A safepoint can still cause the ServiceThread to check for
>> async deflation work via is_async_deflation_requested.
>> - Refactor code from ObjectSynchronizer::is_cleanup_needed() into
>> monitors_used_above_threshold() and remove is_cleanup_needed().
>> - In addition to System.gc(), the VM_Exit VM op and the final
>> VMThread safepoint now set the is_special_deflation_requested
>> flag to reduce the in-use monitor population that is reported by
>> ObjectSynchronizer::log_in_use_monitor_details() at VM exit.
>>
>> Test update:
>> - test/hotspot/gtest/oops/test_markOop.cpp is updated to work with
>> AsyncDeflateIdleMonitors.
>>
>> Collateral:
>> - Add/clarify/update some logging messages.
>>
>> Cleanup:
>> - Updated comments based on Karen's code review.
>> - Change 'special cleanup' -> 'special deflation' and
>> 'async cleanup' -> 'async deflation'.
>> - comment and function name changes
>> - Clarify MonitorUsedDeflationThreshold description;
>>
>>
>> Main bug URL:
>>
>> JDK-8153224 Monitor deflation prolong safepoints
>> https://bugs.openjdk.java.net/browse/JDK-8153224
>>
>> The project is currently baselined on jdk-13+22.
>>
>> Here's the full webrev URL:
>>
>> http://cr.openjdk.java.net/~dcubed/8153224-webrev/7-for-jdk13.full/
>>
>> Here's the incremental webrev URL:
>>
>> http://cr.openjdk.java.net/~dcubed/8153224-webrev/7-for-jdk13.inc/
>>
>> I have not updated the OpenJDK wiki to reflect the CR4 changes:
>>
>> https://wiki.openjdk.java.net/display/HotSpot/Async+Monitor+Deflation
>>
>> The wiki doesn't say a whole lot about the async deflation invocation
>> mechanism so I have to figure out how to add that content.
>>
>> This version of the patch has been thru Mach5 tier[1-8] testing on
>> Oracle's usual set of platforms. My Solaris-X64 stress kit run is
>> running now. Kitchensink8H on product, fastdebug, and slowdebug bits
>> are running on Linux-X64, MacOSX and Solaris-X64. I still have to run
>> my stress kit on Linux-X64. I still have to run the SPECjbb2015
>> baseline and CR4 runs on Linux-X64, MacOSX and Solaris-X64.
>>
>> Thanks, in advance, for any questions, comments or suggestions.
>>
>> Dan
>>
>> On 5/6/19 11:52 AM, Daniel D. Daugherty wrote:
>>> Greetings,
>>>
>>> I had some discussions with Karen about a race that was in the
>>> ObjectMonitor::enter() code in CR2/v2.02/5-for-jdk13. This race was
>>> theoretical and I had no test failures due to it. The fix is pretty
>>> simple: remove the special case code for async deflation in the
>>> ObjectMonitor::enter() function and rely solely on the ref_count
>>> for ObjectMonitor::enter() protection.
>>>
>>> During those discussions Karen also floated the idea of using the
>>> ref_count field instead of the contentions field for the Async
>>> Monitor Deflation protocol. I decided to go ahead and code up that
>>> change and I have run it through the usual stress and Mach5 testing
>>> with no issues. It's also known as v2.03 (for those for with the
>>> patches) and as webrev/6-for-jdk13 (for those with webrev URLs).
>>> Sorry for all the names...
>>>
>>> Main bug URL:
>>>
>>> JDK-8153224 Monitor deflation prolong safepoints
>>> https://bugs.openjdk.java.net/browse/JDK-8153224
>>>
>>> The project is currently baselined on jdk-13+18.
>>>
>>> Here's the full webrev URL:
>>>
>>> http://cr.openjdk.java.net/~dcubed/8153224-webrev/6-for-jdk13.full/
>>>
>>> Here's the incremental webrev URL:
>>>
>>> http://cr.openjdk.java.net/~dcubed/8153224-webrev/6-for-jdk13.inc/
>>>
>>> I have also updated the OpenJDK wiki to reflect the CR3 changes:
>>>
>>> https://wiki.openjdk.java.net/display/HotSpot/Async+Monitor+Deflation
>>>
>>> This version of the patch has been thru Mach5 tier[1-8] testing on
>>> Oracle's usual set of platforms. My Solaris-X64 stress kit run had
>>> no issues. Kitchensink8H on product, fastdebug, and slowdebug bits
>>> had no failures on Linux-X64; MacOSX fastdebug and slowdebug and
>>> Solaris-X64 release had the usual "Too large time diff" complaints.
>>> 12 hour Inflate2 runs on product, fastdebug and slowdebug bits on
>>> Linux-X64, MacOSX and Solaris-X64 had no failures. My Linux-X64
>>> stress kit is running right now.
>>>
>>> I've done the SPECjbb2015 baseline and CR3 runs. I need to gather
>>> the results and analyze them.
>>>
>>> Thanks, in advance, for any questions, comments or suggestions.
>>>
>>> Dan
>>>
>>>
>>> On 4/25/19 12:38 PM, Daniel D. Daugherty wrote:
>>>> Greetings,
>>>>
>>>> I have a small but important bug fix for the Async Monitor Deflation
>>>> project ready to go. It's also known as v2.02 (for those for with the
>>>> patches) and as webrev/5-for-jdk13 (for those with webrev URLs). Sorry
>>>> for all the names...
>>>>
>>>> JDK-8222295 was pushed to jdk/jdk two days ago so that baseline patch
>>>> is out of our hair.
>>>>
>>>> Main bug URL:
>>>>
>>>> JDK-8153224 Monitor deflation prolong safepoints
>>>> https://bugs.openjdk.java.net/browse/JDK-8153224
>>>>
>>>> The project is currently baselined on jdk-13+17.
>>>>
>>>> Here's the full webrev URL:
>>>>
>>>> http://cr.openjdk.java.net/~dcubed/8153224-webrev/5-for-jdk13.full/
>>>>
>>>> Here's the incremental webrev URL (JDK-8153224):
>>>>
>>>> http://cr.openjdk.java.net/~dcubed/8153224-webrev/5-for-jdk13.inc/
>>>>
>>>> I still have to update the OpenJDK wiki to reflect the CR2 changes:
>>>>
>>>> https://wiki.openjdk.java.net/display/HotSpot/Async+Monitor+Deflation
>>>>
>>>> This version of the patch has been thru Mach5 tier[1-6] testing on
>>>> Oracle's usual set of platforms. Mach5 tier[7-8] is running now.
>>>> My stress kit is running on Solaris-X64 now. Kitchensink8H is running
>>>> now on product, fastdebug, and slowdebug bits on Linux-X64, MacOSX
>>>> and Solaris-X64. 12 hour Inflate2 runs are running now on product,
>>>> fastdebug and slowdebug bits on Linux-X64, MacOSX and Solaris-X64.
>>>> I'll start my my stress kit on Linux-X64 sometime on Sunday (after
>>>> my jdk-13+18 stress run is done).
>>>>
>>>> I'll do SPECjbb2015 baseline and CR2 runs after all the stress
>>>> testing is done.
>>>>
>>>> Thanks, in advance, for any questions, comments or suggestions.
>>>>
>>>> Dan
>>>>
>>>>
>>>> On 4/19/19 11:58 AM, Daniel D. Daugherty wrote:
>>>>> Greetings,
>>>>>
>>>>> I finally have CR1 for the Async Monitor Deflation project ready to
>>>>> go. It's also known as v2.01 (for those for with the patches) and as
>>>>> webrev/4-for-jdk13 (for those with webrev URLs). Sorry for all the
>>>>> names...
>>>>>
>>>>> Main bug URL:
>>>>>
>>>>> JDK-8153224 Monitor deflation prolong safepoints
>>>>> https://bugs.openjdk.java.net/browse/JDK-8153224
>>>>>
>>>>> Baseline bug fixes URL:
>>>>>
>>>>> JDK-8222295 more baseline cleanups from Async Monitor
>>>>> Deflation project
>>>>> https://bugs.openjdk.java.net/browse/JDK-8222295
>>>>>
>>>>> The project is currently baselined on jdk-13+15.
>>>>>
>>>>> Here's the webrev for the latest baseline changes (JDK-8222295):
>>>>>
>>>>> http://cr.openjdk.java.net/~dcubed/8153224-webrev/4-for-jdk13.8222295
>>>>>
>>>>> Here's the full webrev URL (JDK-8153224 only):
>>>>>
>>>>> http://cr.openjdk.java.net/~dcubed/8153224-webrev/4-for-jdk13.full/
>>>>>
>>>>> Here's the incremental webrev URL (JDK-8153224):
>>>>>
>>>>> http://cr.openjdk.java.net/~dcubed/8153224-webrev/4-for-jdk13.inc/
>>>>>
>>>>> So I'm looking for reviews for both JDK-8222295 and the latest
>>>>> version
>>>>> of JDK-8153224...
>>>>>
>>>>> I still have to update the OpenJDK wiki to reflect the CR changes:
>>>>>
>>>>> https://wiki.openjdk.java.net/display/HotSpot/Async+Monitor+Deflation
>>>>>
>>>>> This version of the patch has been thru Mach5 tier[1-3] testing on
>>>>> Oracle's usual set of platforms. Mach5 tier[4-6] is running now and
>>>>> Mach5 tier[78] will be run later today. My stress kit on Solaris-X64
>>>>> is running now. Linux-X64 stress testing will start on Sunday. I'm
>>>>> planning to do Kitchensink runs, SPECjbb2015 runs and my monitor
>>>>> inflation stress tests on Linux-X64, MacOSX and Solaris-X64.
>>>>>
>>>>> Thanks, in advance, for any questions, comments or suggestions.
>>>>>
>>>>> Dan
>>>>>
>>>>>
>>>>> On 3/24/19 9:57 AM, Daniel D. Daugherty wrote:
>>>>>> Greetings,
>>>>>>
>>>>>> Welcome to the OpenJDK review thread for my port of Carsten's
>>>>>> work on:
>>>>>>
>>>>>> JDK-8153224 Monitor deflation prolong safepoints
>>>>>> https://bugs.openjdk.java.net/browse/JDK-8153224
>>>>>>
>>>>>> Here's a link to the OpenJDK wiki that describes my port:
>>>>>>
>>>>>> https://wiki.openjdk.java.net/display/HotSpot/Async+Monitor+Deflation
>>>>>>
>>>>>>
>>>>>> Here's the webrev URL:
>>>>>>
>>>>>> http://cr.openjdk.java.net/~dcubed/8153224-webrev/3-for-jdk13/
>>>>>>
>>>>>> Here's a link to Carsten's original webrev:
>>>>>>
>>>>>> http://cr.openjdk.java.net/~cvarming/monitor_deflate_conc/0/
>>>>>>
>>>>>> Earlier versions of this patch have been through several rounds of
>>>>>> preliminary review. Many thanks to Carsten, Coleen, Robbin, and
>>>>>> Roman for their preliminary code review comments. A very special
>>>>>> thanks to Robbin and Roman for building and testing the patch in
>>>>>> their own environments (including specJBB2015).
>>>>>>
>>>>>> This version of the patch has been thru Mach5 tier[1-8] testing on
>>>>>> Oracle's usual set of platforms. Earlier versions have been run
>>>>>> through my stress kit on my Linux-X64 and Solaris-X64 servers
>>>>>> (product, fastdebug, slowdebug).Earlier versions have run
>>>>>> Kitchensink
>>>>>> for 12 hours on MacOSX, Linux-X64 and Solaris-X64 (product,
>>>>>> fastdebug
>>>>>> and slowdebug). Earlier versions have run my monitor inflation
>>>>>> stress
>>>>>> tests for 12 hours on MacOSX, Linux-X64 and Solaris-X64 (product,
>>>>>> fastdebug and slowdebug).
>>>>>>
>>>>>> All of the testing done on earlier versions will be redone on the
>>>>>> latest version of the patch.
>>>>>>
>>>>>> Thanks, in advance, for any questions, comments or suggestions.
>>>>>>
>>>>>> Dan
>>>>>>
>>>>>> P.S.
>>>>>> One subtest in gc/g1/humongousObjects/TestHumongousClassLoader.java
>>>>>> is currently failing in -Xcomp mode on Win* only. I've been trying
>>>>>> to characterize/analyze this failure for more than a week now. At
>>>>>> this point I'm convinced that Async Monitor Deflation is aggravating
>>>>>> an existing bug. However, I plan to have a better handle on that
>>>>>> failure before these bits are pushed to the jdk/jdk repo.
>>>>>>
>>>>>
>>>>>
>>>>
>>>>
>>>
>>>
>>
>>
>
["CR5-to-CR6-changes" (text/plain)]
Functional:
- Monitor list management is now lock-free and refactored into
separate functions. Almost all of the new functions are static
in synchronizer.cpp to reduce visibility.
- Several functions required major algorithm updates to be both
async and lock-free:
- ObjectSynchronizer::om_release()
- ObjectSynchronizer::om_flush()
- ObjectSynchronizer::deflate_monitor_list_using_JT()
- These three functions will need the most review scrutiny!
- ServiceThread now does the per-thread async monitor deflation
instead of a thread cleaning up after itself.
- Change ObjectMonitorHandle::save_om_ptr() to only return true
when ref_count >= 1. This prevents races with an async deflation
that is bailing out in assert() and guarantee() calls.
- ObjectSynchronizer::monitors_iterate() should use ref_count < 0
as the bailout trigger instead of _owner == DEFLATER_MARKER to
avoid bug where a deflation bailout leads to a missed visit.
Test update:
- WB_G1StartMarkCycle() and WB_FullGC() now request special cleanups
as needed.
- test/jdk/tools/jlink/multireleasejar/JLinkMultiReleaseJarTest.java
work around is backed out now that the ServiceThread is doing
per-thread async monitor deflation.
Collateral:
- Add OM_CACHE_LINE_SIZE so that ObjectMonitor cache line sizes
can be experimented with independently of DEFAULT_CACHE_LINE_SIZE;
for SPARC and X64 configs that use 128 for DEFAULT_CACHE_LINE_SIZE,
we are experimenting with 64.
- ObjectMonitor cache line changes:
- _allocation_state field now shares the _header field cache line.
- _ref_count field is now on its own cache line
- _next_om field is now on its own cache line
- See the comment additions/updates in objectMonitor.hpp.
Cleanup:
- Reduced visibility of a number of things in ObjectSynchronizer
from class visible to static in synchronizer.cpp.
Temporary:
- -XX:[+-]CheckMonitorLists diagnostic option (default false) and
supporting code is in place during the stress testing phase to
chase down any issues with the lock-free monitor list management
code. I have not seen a failure due to the lock-free monitor list
code for several weeks.
- unmark_next() is currently unused. If we don't make any changes
to use it during this code review cycle, then it can be deleted.
[prev in list] [next in list] [prev in thread] [next in thread]
Configure |
About |
News |
Add a list |
Sponsored by KoreLogic