RFC: regarding metaspace(metadata?) dump
Thomas Stüfe
thomas.stuefe at gmail.com
Thu Jan 12 09:55:22 UTC 2023
Hi Yi,
On Thu, Jan 12, 2023 at 9:29 AM Yi Yang <qingfeng.yy at alibaba-inc.com> wrote:
> Thanks Ioi and Thomas for your valuable thoughts! I list some pros about
> metaspace dump.
>
> 1. Standardization. The format of metadata dump is standard and
> well-formed. It could be seamlessly integrated with DevOps/Diagnose/APM
> platforms, while SA is interactively and the output of jcmd is not
> well-formed and parser-unfriendly, and its content is subject to change.
> You can't expect DevOps platform to use SA/coredump/gdb conveniently and
> automatically in production environment.
>
I'm not convinced that json solves the standardization problem. These tools
are for hotspot devs mainly, so they expose implementation details that are
frequently changing. That runs the risk of breakage, regardless of the
format. Json will give you better parsing (at the expense of developer
experience), but their content will change. Your renderer needs to make
sense of the data to do something useful beyond printing the raw json soup.
So the renderer can break if the implementation changes. But it is not
maintained by us, in contrast to our jcmds, so we cannot easily fix it. We
see these kinds of problems even with JFR, and that we maintain ourselves.
If standardization is the goal we take seriously, we'd need to carefully
choose the information we expose, limiting them to those that are stable.
That limits the usefulness of the command. It also puts the onus of
backward compatibility on us. Regression tests and all that. If we don't
take standardization seriously, why do it at all? Why worsen the experience
for hotspot devs to get machine-readable output?
We could add json output as an option beside the normal human-readable
output. I could see this to be useful. But it increases code complexity. It
would also be good to have this consistently for all jcmds, or all that
produce table-like output.
>
> 2. Functionality. MetaspaceDumpBeforeFullGC could generate a small dump
> for further debugging, it works as well as heap dump.
>
Sounds useful.
>
> 3. JVM Metadata. Codecache dump, method counter, method data, they are
> unexplored scopes, reconstruct human-readable representation of compiled
> method.
>
> 4. Complexity. In my humble opinion, all of this stuff such as
> by-chunktype/show-loaders/VM.classloader_stats/VM.classloader_hierarchy/etc
> could be done in other place. VM is eligible to provides a standard and
> rich raw metadata output, third-party parser and UI render display them in
> their way instead of continuously adding new
> filter/grouping/hierarcy/VM.method_counter features when we do want to know
> them. Heap dump is a good example, it dumps all objects/symbols/etc to a
> binary file, and third-party tools orchestrate them to
> histogram/thread/classloader/domtree.
>
Heap dump is a good example of the limits of this approach, since the sheer
size of data can make them cumbersome and expensive to get. There is a
niche for tools that quickly tell you what you need, without a lot of fuss,
runtime costs, and intermediate steps. And that help you along the way of
analysis.
One pro of doing rendering in hotspot instead of dumping the whole thing
and letting tool providers figure it out is that the backend developers
know their backend best. They write analysis tools for themselves (e.g.
VM.metaspace was carefully designed to answer the common Metaspace OOMEs
quickly). They also can get a quick turnaround. Of course, there is a limit
somewhere, which is why we have JFR and JMC.
Tool providers need to understand the backend in order to intelligently
display internal status. Then, they need to agree with hotspot devs on the
data to be provided, then need to render them in an intelligent way. E.g.
if you were to provide a "waste" section equally useful as that one from
VM.metaspace, we would need to expose a lot of internals to the tool
provider. The interface would be broad and very brittle.
> Basically, I know the content of metaspace dump file has overlap with some
> existing tools such as SA/jcmd/coredump/gdb/HeapDump, as Thomas commented
> inlinely, I don’t think metaspace dump can troubleshoot many problems that
> it is the only solution. I think metadata dump is more about providing a
> standardized and parser-friendly framework, so that users at all levels can
> inspect JVM metadata information they care about. In addition, something
> like M(etaspace)AT could orchestrate dump content with
> filter/grouping/hierarchy options.
>
> > Also you mentioned that "Internally we implemented a metaspace dump that
> generates human-readable text". Can you share how this tool was implemented?
>
> That's not surprising, it iterates CLD/Classes/etc and dumps basic
> information about metaspace, a demo metaspace dump can be found at
> https://gist.github.com/y1yang0/683d8a58dd946b3e9180682863df55ea
>
>
In that case, it misses an important part of the picture that VM.metaspace
provides (Fragmentation and dead-memory analysis).
---
I am not opposed to adding json format output to jcmd, but not at the
expense of the current human-readable reports. There are still a lot of
questions are unanswered. Lets see what others think.
A jcmd that dumps metadata could be useful as an addition to existing
commands if we can keep the overlap small.
Cheers, Thomas
------------------------------------------------------------------
> From:Thomas Stüfe <thomas.stuefe at gmail.com>
> Send Time:2023 Jan. 12 (Thu.) 15:06
> To:"YANG, Yi" <qingfeng.yy at alibaba-inc.com>
> Cc:HotSpot Open Source Developers <hotspot-dev at openjdk.java.net>;
> hotspot-runtime-dev <hotspot-runtime-dev at openjdk.java.net>; hotspot-dev <
> hotspot-dev at openjdk.org>
> Subject:Re: RFC: regarding metaspace(metadata?) dump
>
> Hi Yi,
>
> A lot of what you try to do already exists. For example, we also have
> `VM.metaspace`. This is a quite powerful command to analyze
> Metaspace-related issues, especially for fragmentation and other wastages.
> Generally speaking, it is the tool you use to look at the underpinnings of
> metaspace, the allocator, while tools like `VM.classloaders`,
> `VM.classloader_stats` and `VM.classes` look at things "from above", e.g.
> walk the CLDG. All these tools have already a bit of overlap.
>
> For analyzing OOMEs, you need several tools, since it can be caused by
> multiple issues. E.g. tools that walk the CLDG don't see fragmentation, or
> unclaimed metaspace for dead loaders.
>
> Please find more remarks inline.
>
> On Wed, Jan 11, 2023 at 1:56 PM Yi Yang <qingfeng.yy at alibaba-inc.com>
> wrote:
> Hi,
>
> Internally, we often receive feedback from users and ask for help on
> metaspace-related issues, for example
> 1. Users are eager to know which GroovyClassLoader loads which classes,
> why they are not unloaded,
> and why they are leading to Metaspace OOME.
>
> There are several tools to do this, for example:
>
> `VM.metaspace show-loaders show-classes`
> `VM.classloaders show-classes`
>
> both show you loaded classes by loader, and the former also shows you
> metaspace stats needed to understand OOMEs. None of these tools shows you
> why loaders are kept alive, but for that you need heap- and
> GC-root-analysis. This quickly enters the territory of Eclipse MAT and
> similar tools, where having a text-based tool alone gets cumbersome.
>
> 2. They want to know the class structure of dynamically generated classes
> in some scenarios such as
> deserialization
>
> Interesting. This seems to be a very specific query; not sure how general
> the need for this is. `VM.classes -verbose` shows a part of the story.
>
> 3. Finding memory leaking about duplicated classes
>
> Again,
>
> `VM.metaspace show-loaders show-classes`
> `VM.classloaders show-classes`
>
> but also `VM.classes` and `VM.classloader_stats` are your friends here.
>
>
> ...
> Internally we implemented a metaspace dump that generates human-readable
> text, it looks something like this:
>
> [Basic Information]
> Dump Reason : JCMD
> MaxMetaspaceSize : 18446744073709547520 B
> CompressedClassSpaceSize : 1073741824 B
> Class Space Used : 309992 B
> Class Space Capacity : 395264 B
> ...
> [Class Loader Data]
> ClassLoaderData : loader = 0x000000008024f928, loader_klass =
> 0x0000000800010098, loader_klass_name =
> sun/misc/Launcher$AppClassLoader, label = N/A
> Class Used Chunks :
> * Chunk : [0x0000000800060000, 0x0000000800060230, 0x0000000800060800)
> NonClass Used Chunks :
> * Chunk : [0x00007fd8379c1000, 0x00007fd8379c1350, 0x00007fd8379c2000)
> Klasses :
> Klass : 0x0000000800060028, name = Test, size = 520 B
> ConstantPool : 0x00007fd8379c1050, size = 296 B
> ...
>
>
> `VM.metaspace` shows you the chunk composition of arenas if needed.
>
> E.g. : `VM.metaspace by-chunktype show-loaders`
>
> ```
> Usage per loader:
>
>
>
> 1: CLD 0x00007f72fc29b820: "app" instance of
> jdk.internal.loader.ClassLoaders$AppClassLoader
>
>
> Loaded classes:
>
>
>
> 1: de.stuefe.repros.MiscUtils$$Lambda$1/0x0000000801001448
>
> 2: de.stuefe.repros.MiscUtils
>
> 3: de.stuefe.repros.Simple2
>
>
>
> 4: de.stuefe.repros.Simple
>
> 5: de.stuefe.repros.SimpleBase
>
> 6: de.stuefe.repros.I2
> 7: de.stuefe.repros.I1
> -total-: 7 classes
> Non-Class:
> Usage by chunk level:
> 4m chunks: (none)
> 2m chunks: (none)
> 1m chunks: (none)
> 512k chunks: (none)
> 256k chunks: (none)
> 128k chunks: (none)
> 64k chunks: (none)
> 32k chunks: (none)
> 16k chunks: (none)
> 8k chunks: 1 chunk, 8,00 KB capacity, 8,00 KB (100%)
> committed, 8,00 KB (100%) used, 0 bytes ( 0%) free, 0 bytes (
> 0%) waste
> 4k chunks: 1 chunk, 4,00 KB capacity, 4,00 KB (100%)
> committed, 256 bytes ( 6%) used, 3,75 KB ( 94%) free, 0 bytes (
> 0%) waste
> 2k chunks: (none)
> 1k chunks: (none)
> -total-: 2 chunks, 12,00 KB capacity, 12,00 KB
> (100%) committed, 8,25 KB ( 69%) used, 3,75 KB ( 31%) free, 0
> bytes ( 0%) waste
> deallocated: 1 blocks with 24 bytes
>
> Class:
> Usage by chunk level:
> .... and so forth
> ```
>
> but for analyzing potential fragmentation issues (which have been rare
> since JEP 387) the "Waste" section at the end of the printout is much more
> helpful, e.g.:
>
> ```
> Waste (unused committed space):(percentages refer to total committed size
> 384,00 KB):
> Waste in chunks in use: 0 bytes ( 0%)
> Free in chunks in use: 85,01 KB ( 22%)
> In free chunks: 0 bytes ( 0%)
> Deallocated from chunks in use: 928 bytes ( <1%) (3 blocks)
> -total-: 85,91 KB ( 22%)
> ```
>
> It has been working effectively for several years and has helped many
> users solve metaspace-related problems.
> But a more user-friendly way is that JDK can inherently support this
> capability. We hope that format of the metaspace
> dump file can take both flexibility and compatibility into account, and
> the content of dump file should be detailed
> enough to meet the needs of both application developers and lower-level
> developers.
>
> Based on above considerations, I think using JSON as its file format is an
> appropriate solution(But XML or binary
> format are still not excluded as candidates). Specifically, in earlier
> thoughts, I thought the format of the metaspace
> file could be as follows(pretty printed)
>
> https://gist.github.com/y1yang0/ab3034b6381b8a9d215602c89af4e9c3
>
> Using the JSON format, we can flexibly add new fields without breaking
> compatibility. It is debatable as to which data
> to write. We can reach a consensus that third-party parsers(Metaspace
> Analyzer Tool) can at least reconstruct Java
> source code from the dump file. Based on this, we can write more useful
> information for low-level troubleshooting
> or debugging. (e.g. the init_state of InstanceKlass).
> In addition, we can even output the native code and associated information
> with regard to Method, third-party parser
> can reconstruct the human-readable assembly representation of the compiled
> method based on dump file. To some extent,
> we have implemented code cache dump by the way. For this reason, I'm not
> sure if the title of the RFC proposal should
> be called metaspace dump, maybe metadata dump? It looks more like a
> metadata-dump framework.
>
> Do you have any thoughts about metaspace/metadata dump? Looking forward to
> hearing your feedback, any comments are invaluable!
>
> Best regards,
> Yi Yang
>
>
> Analyzing the structure of generated classes sounds interesting, and could
> help with analyzing issues with bytecode instrumentation tools.
>
> For analyzing general metaspace OOMEs we are already covered quite well.
> Not perfect, but your proposal does intersect with existing tools a lot. To
> keep code complexity down, I'd rather avoid adding duplicate features.
>
> Cheers, Thomas
>
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.openjdk.org/pipermail/hotspot-runtime-dev/attachments/20230112/36290698/attachment-0001.htm>
More information about the hotspot-runtime-dev
mailing list