RFR(S) 8218751 Do not store original classfiles inside the CDS archive

Jiangli Zhou jianglizhou at google.com
Tue Feb 19 04:12:02 UTC 2019 

On Mon, Feb 18, 2019 at 11:33 AM Ioi Lam <ioi.lam at oracle.com> wrote:

>
>
> On 2/15/19 7:11 PM, Jiangli Zhou wrote:
>
> On Fri, Feb 15, 2019 at 6:56 PM Ioi Lam <ioi.lam at oracle.com> wrote:
>
>> On 2/15/19 4:54 PM, Jiangli Zhou wrote:
>>
>> Hi,
>>
>> To answer your use case question, one of the case reported last year in
>> OpenJDK was JRebel (please go back to the hotspot-dev mail list 2018
>> Oct. archive). That is an existing example as I've tried to point out in my
>> earlier email.
>>
>> First of all, if you read the email carefully, CDS was not even in their
>> usual test matrix. Also, if you're modifying classes on the fly, you're
>> slowing down start-up big time. A few ms spent in decoding the classfile
>> data will be completely drown out by the overhead of the classfile parsing
>> and patching code in the JVMTI agent.
>>
>> So no, JRebel will not start up a few % faster if CDS stores the
>> classfile data. It will most likely be not measurable.
>>
>
> My understanding of the case is that their users generated a CDS archive
> and used it together with JRebel. That's why the issue was unnoticed until
> their users reported. There are use cases out there in the community that
> we simply are not aware of, so it would not be wise to assume there is no
> usage.
>
> Just to confirm one thing -- you do not dispute my claim that this
> "optimization" has no benefits even for those that actually use CDS and
> CFLH together. Correct?
>

That's actually the part that I have a different opinion. With the proposed
change in the current webrev, the performance hit is across the board for
loading shared classes regardless if there is a class being
redefined/retransformed, as long as the can_generate_all_class_hook_events
capability is enabled.


>
> We have plenty of nice-to-have harmless optimizations in HotSpot that
> probably weren't vigorously validated. However, this is not one of them.
>
> Here we have clear evidence of harm (50% footprint increase), with no
> evidence of benefit (theoretical or real-life). This code was checked into
> HotSpot without proper validation. That was wrong, and that's why I am
> taking it out now.
>
> It's easy to prove me wrong. Just supply a proper benchmark that shows
> benefits, and I will change my mind. That's the minimal standard for an
> optimization that has harmful side effects. I have created JDK-8219255 for
> tracking this.
>

Looks like there is a deadlock in the discussion. To summarize and help
move this forward. Here are your reasons for dropping the 'od' space:

   - static footprint increase due to 'od' space
   - no benefit
   - maintenance of the code

Regarding static footprint, the change proposed in the current webrev is a
short-term and temporary solution. For a long-term solution, the original
class files can be removed from the final image created by jlink when jlink
can work with CDS to produce a single runtime image.

On the benefit side, the 'od' space provides both startup speedup and
runtime memory saving (when can_generate_all_class_hook_events capability
is enabled), which I've already pointed those out in earlier emails.

For maintenance, it is also not an issue with the support from the
community (I'll for sure to continue contributing to it). What's more
important is to continue keeping CDS/AppCDS moving in the right direction
by providing more performance & memory benefits and making it easier to use.

So I think providing a command-line option, -XX:EnableOptionalDataSpace
that allows user to enable/disable the 'od' space at both dump time and
runtime is the best choice that can work with the short-term fix proposed
in the current webrev.

Thanks,
Jiangli

Thanks,
Jiangli


> Thanks
> - Ioi
>
>
> Thanks,
> Jiangli
>
>
>> The issue with the current change is that it's only targeted for reducing
>> the static footprint (static footprint reduction could be achieved with
>> alternative approaches such as file compression). The removal of the
>> 'od' space and the benefits (both the startup and runtime footprint) is not
>> backed up by a clear requirement here and the change should not go in as is.
>>
>> As part of the JVM development, we constantly review and remove
>> unnecessary code to keep the JVM healthy. That's a clear requirement and
>> mandate on Oracle as the major contributor to the JDK.
>>
>> In this case, we are removing a premature optimization in the original
>> design. The classfile data stored in the CDS archive are rarely used
>> (99.99+%), and when they are used, the speed optimization is irrelevant
>> because the user of this data (JVMTI agent) is so much slower. So everyone
>> is taking a footprint hit with no benefit to anyone.
>> Thanks
>>
>> - Ioi
>>
>>
>> Just to summarize the drawbacks of the current change:
>>
>>    - Runtime footprint increase (causes more memory fragmentation even
>>    the memory is freed after use)
>>    - Startup time regression when can_generate_all_class_hook_events
>>    capability is enabled
>>    - May cause issue with future optimization
>>
>> Thanks,
>> Jiangli
>>
>>
>>
>>
>>
>> On Fri, Feb 15, 2019 at 2:07 PM Ioi Lam <ioi.lam at oracle.com> wrote:
>>
>>> Hi Jiangli,
>>>
>>> Thanks for volunteering on this. I think it's best to do that (optional
>>> support for storing uncompressed classfile data into the CDS archive) as a
>>> separate bug than the current issue.
>>>
>>> If you decide to provide a patch for that, I think it would be best to
>>> find actual users or use cases that would find it beneficial. Otherwise it
>>> will appear to be a solution looking for a problem.
>>>
>>> Thanks
>>>
>>> - Ioi
>>>
>>>
>>> On 2/15/19 10:35 AM, Jiangli Zhou wrote:
>>>
>>> I'm willing to continue contributing to the support and maintaining of
>>> CDS/AppCDS. Often startup improvements with measurable gain (>4% in this
>>> case) are non-trivial. The support for 'od' is not overly complex comparing
>>> to the performance gain. In a rosier picture if Jlink/AOT/CDS can work
>>> together in harmony to create a single image in the future, class files
>>> (and even JAR files) can be eliminated and no class file data can be loaded
>>> at runtime from a JAR file. We should consider all aspects and not make a
>>> decision lightly.
>>>
>>> For the dynamic archiving, this is not a blocking issue. I can step in
>>> an help with making 'od' optional if needed.
>>>
>>> Thanks,
>>> Jiangli
>>>
>>> On Thu, Feb 14, 2019 at 7:53 PM Ioi Lam <ioi.lam at oracle.com> wrote:
>>>
>>>> That means we have to add a new -XX option to enable this, and need to
>>>> add extra test cases and maintenance.
>>>>
>>>> We are talking about only single digit % benefits in very narrow use
>>>> cases, and it's not clear whether performance is important in such cases
>>>> (or if people would bother to use this -XX flag to gain a few %). Without
>>>> user input, it's hard to justify keep spending time on an optimization that
>>>> literally no one had asked for.
>>>>
>>>> I just don't think we have time to keep maintaining this. If someone is
>>>> willing to step up and provide support for this, feel free.
>>>>
>>>> Thanks
>>>>
>>>> - Ioi
>>>>
>>>>
>>>> On 2/14/19 1:20 PM, Jiangli Zhou wrote:
>>>>
>>>> Based on the issues raised in OpenJDK last year, there appears to be
>>>> existing use cases with AppCDS + CLFH. When designing the platform, we
>>>> should think from the user perspective. Making 'od' optional is not complex
>>>> and gives the control to the Java users. It also avoids potential waste of
>>>> effort for removing&putting back the optimization.
>>>>
>>>> Thanks,
>>>> Jiangli
>>>>
>>>> On Thu, Feb 14, 2019 at 11:46 AM Ioi Lam <ioi.lam at oracle.com> wrote:
>>>>
>>>>> Such optional support increase the complexity of the VM. My preference
>>>>> would be to wait until there's an actual need for this.
>>>>>
>>>>> The JDK is release under 6 months cadence, so if the removal turns out
>>>>> to be a bad idea, we can reinstate the code in the next release. Or,
>>>>> someone can just make their own JDK build with a simple anti-delta of this
>>>>> patch.
>>>>>
>>>>> If we are too reluctant to remove anything, the JDK will eventually
>>>>> become a hopeless mess.
>>>>>
>>>>> Thanks
>>>>>
>>>>> - Ioi
>>>>> On 2/14/19 10:43 AM, Jiangli Zhou wrote:
>>>>>
>>>>> Sorry for the slow turnaround. Hopefully it will get better after this
>>>>> week. As there is no enough user data/requirement to determine which
>>>>> optimization direction is more important in this case, it might be
>>>>> reasonable to make the 'optional data' space truly optional, which can be
>>>>> controlled by a command-line option. If the performance is more important
>>>>> (in case should_post_class_file_load_hook is required for a particular use
>>>>> case), user can enable dumping out the 'od' space with archived class file
>>>>> data, otherwise the data can be loaded at runtime. What's your thought on
>>>>> this?
>>>>>
>>>>> Thanks,
>>>>> Jiangli
>>>>>
>>>>> On Wed, Feb 13, 2019 at 12:03 PM Ioi Lam <ioi.lam at oracle.com> wrote:
>>>>>
>>>>>>
>>>>>> On 2/13/19 11:24 AM, Jiangli Zhou wrote:
>>>>>>
>>>>>> Hi Ioi,
>>>>>>
>>>>>> Thanks for the additional information and performance data. Please
>>>>>> see more comments below.
>>>>>>
>>>>>> On Wed, Feb 13, 2019 at 5:33 AM Ioi Lam <ioi.lam at oracle.com> wrote:
>>>>>>
>>>>>>> Hi Jiangli,
>>>>>>>
>>>>>>> The main reason for doing this is to reduce the size of the CDS
>>>>>>> file.
>>>>>>> For example, when archiving Eclipse IDE with JDK-8215311 (Dynamic
>>>>>>> Class
>>>>>>> Metadata Archive), the CDS archive is reduced from 100MB to about
>>>>>>> 67MB.
>>>>>>>
>>>>>>> We have not heard any requirement for high performance with
>>>>>>> CDS+ClassFileLoadHook. It doesn't seem right for everyone to take a
>>>>>>> 50%
>>>>>>> file size penalty for an optimization that no one has asked for.
>>>>>>>
>>>>>>
>>>>>> Jut to clarify, it is ~30% not 50% (possibly a typo?), correct?
>>>>>>
>>>>>>
>>>>>> Depending on how you look at it. Going from 67MB to 100MB is an
>>>>>> increase of 33MB which is 50%.
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>>
>>>>>>> Here are some perf numbers (running "java -version" with an agent
>>>>>>> that
>>>>>>> installs a CFLH that doesn't nothing). Yes, it's somewhat slower as
>>>>>>> expected, but it doesn't seem to be catastrophic. Using CDS is
>>>>>>> nevertheless much faster than without CDS anyway.
>>>>>>>
>>>>>>> After No CFLH         0.0476 seconds time elapsed  ( +- 0.20% )
>>>>>>> After with CFLH       0.0513 seconds time elapsed  ( +- 0.18% )
>>>>>>> 7.773%
>>>>>>> slower
>>>>>>>
>>>>>>> Before no CFLH        0.0472 seconds time elapsed  ( +- 0.21% )
>>>>>>> Before with CFLH      0.0492 seconds time elapsed  ( +- 0.18% )
>>>>>>> 4.237%
>>>>>>> slower
>>>>>>>
>>>>>>
>>>>>> Comparing 'After with CFLH' (0.0513 seconds) with 'Before with CFLH'
>>>>>> (0.0492 seconds), there is about  4.27% performance degradation observed
>>>>>> from your data.
>>>>>>
>>>>>> Another disadvantage of this proposed change is the increasing of
>>>>>> runtime memory. When the class file data is stored in the shared archive
>>>>>> file and mapped as RO (read-only), it can be shared among multiple JVM
>>>>>> processes. For the default CDS archive case, it's shared by all JVM
>>>>>> processes running simultaneously. With the proposed change in the webrev,
>>>>>> the memory for the class file data is allocated at runtime. It moves the
>>>>>> cost from static footprint to runtime memory footprint, which is a higher
>>>>>> price because it needs to be multiplied by the number of running JVM
>>>>>> processes.
>>>>>>
>>>>>>
>>>>>> With this patch, the decoded classdata is freed after the CFLH is
>>>>>> called, so there's no resident memory cost.
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>> CDS Off with CFLH     0.0869 seconds time elapsed  ( +- 0.18% )
>>>>>>> CDS Off without CFLH  0.0852 seconds time elapsed  ( +- 0.14% )
>>>>>>> 1.995%
>>>>>>> slower
>>>>>>>
>>>>>>> I am not sure if there's a performance critical case for CFLH.
>>>>>>
>>>>>>
>>>>>> I have the similar question. It would be good to go back and
>>>>>> re-investigate the original requirements for supporting CFLH use cases.
>>>>>>
>>>>>>
>>>>>> I think when we added CFLH support for CDS, the goal was "to not lose
>>>>>> all benefit of CDS when CFLH is enabled". I think we still achieve that
>>>>>> after this patch. There really was no requirement "to make CFLH blazing
>>>>>> fast at all cost".
>>>>>>
>>>>>> Thanks
>>>>>>
>>>>>> - Ioi
>>>>>>
>>>>>>
>>>>>>
>>>>>>> I suppose
>>>>>>> if performance is critical you would rewrite the classes statically
>>>>>>> and
>>>>>>> rebuild your app, instead of patching its bytecodes at runtime.
>>>>>>> However,
>>>>>>> if there were indeed a performance critical case, I think it's
>>>>>>> better to
>>>>>>> change JVMTI to allow a 2-level filtering for CFLH:
>>>>>>>
>>>>>>> I believe the overwhelming use case, where performance is critical,
>>>>>>> the
>>>>>>> CFLH will just patch a small number of class files. I can't fathom
>>>>>>> any
>>>>>>> use case where someone wants to patch EVERY loaded class.
>>>>>>>
>>>>>>> The current CFLH passes both the name of the class, as well as the
>>>>>>> classfile data, in a single call. This forces CDS to decode the
>>>>>>> classfile data for every hook call. However, in most cases, the CFLH
>>>>>>> will just examine the class name, and do nothing unless the name
>>>>>>> matches
>>>>>>> a certain pattern, so we end up wasting the decoding effort.
>>>>>>>
>>>>>>> My suggested improvement is to add a new filtering call in JVMTI
>>>>>>> that
>>>>>>> passes only the name. If the CFLH wants to patch the class, it will
>>>>>>> then
>>>>>>> request the classfile data, at which point CDS will decode it from
>>>>>>> the
>>>>>>> modules file or JAR file.
>>>>>>>
>>>>>>
>>>>>> That probably is the right direction.
>>>>>>
>>>>>> Thanks,
>>>>>> Jiangli
>>>>>>
>>>>>>
>>>>>>>
>>>>>>> Thanks
>>>>>>>
>>>>>>> - Ioi
>>>>>>>
>>>>>>>
>>>>>>> On 2/12/19 5:18 PM, Jiangli Zhou wrote:
>>>>>>> > Hi Ioi,
>>>>>>> >
>>>>>>> > I'd like to understand the performance impact with this change. Do
>>>>>>> you have
>>>>>>> > any performance numbers  when
>>>>>>> JvmtiExport::should_post_class_file_load_hook()
>>>>>>> > is required? This is a performance vs footprint trade-off. For
>>>>>>> some users,
>>>>>>> > performance is more important than static footprint.
>>>>>>> >
>>>>>>> > Could you also please provide some background/motivation for this
>>>>>>> change?
>>>>>>> >
>>>>>>> > Thanks,
>>>>>>> > Jiangli
>>>>>>> >
>>>>>>> > On Mon, Feb 11, 2019 at 9:24 AM Ioi Lam <ioi.lam at oracle.com>
>>>>>>> wrote:
>>>>>>> >
>>>>>>> >>
>>>>>>> http://cr.openjdk.java.net/~iklam/jdk13/8218751-dont-store-classfiles-in-cds.v01/
>>>>>>> >> https://bugs.openjdk.java.net/browse/JDK-8218751
>>>>>>> >>
>>>>>>> >> For JVMTI ClassFileLoadHook support, the CDS archive currently
>>>>>>> stores
>>>>>>> >> the original classfile data of all archived classes.
>>>>>>> >>
>>>>>>> >> However, this consists of over 30% of the archive size. Because
>>>>>>> all
>>>>>>> >> original classfile data are already available in other files
>>>>>>> (such as the
>>>>>>> >> JDK lib/modules file, or JAR files in the classpath), we can
>>>>>>> simply read
>>>>>>> >> from these locations when needed by JVMTI.
>>>>>>> >>
>>>>>>> >> For the default CDS archive (included as part of the JDK
>>>>>>> distribution),
>>>>>>> >> the size is reduced from about 18.5MB to 12.1MB on Linux/x64.
>>>>>>> >>
>>>>>>> >> Thanks
>>>>>>> >>
>>>>>>> >> - Ioi
>>>>>>> >>
>>>>>>> >>
>>>>>>>
>>>>>>
>

More information about the hotspot-runtime-dev mailing list