Please implement client switch in 64-bit server JDK 14 builds
Vladimir Kozlov
vladimir.kozlov at oracle.com
Fri Jul 19 21:56:05 UTC 2019
It does not reduce static size of JVM itself - it is still Server VM. You have to rebuild it as other suggested if you
want to reduce its size.
The default memory allocated during execution is reduced with this flag as you can see in my first link. Such as
ReservedCodeCacheSize (compiled code), MetaspaceSize (classes metadata). Default Java heap size is also reduced but, as
I understand from your previous comments, you want to use your own -Xmx value. These 3 values are main settings which
control memory usage by VM.
I would suggest to do experiments and see.
Vladimir
On 7/19/19 2:35 PM, Ty Young wrote:
> Yes that works nicely. Are there any other hotspot switches to reduce memory usage or is that it?
>
>
> On 7/19/19 12:25 PM, Vladimir Kozlov wrote:
>> Ty, you can try -XX:+NeverActAsServerClassMachine flag which sets configuration similar to old Client VM (C1 JIT +
>> SerialGC):
>>
>> http://hg.openjdk.java.net/jdk/jdk/file/014decdb5086/src/hotspot/share/compiler/compilerDefinitions.cpp#l116
>> http://hg.openjdk.java.net/jdk/jdk/file/014decdb5086/src/hotspot/share/gc/shared/gcConfig.cpp#l109
>>
>> Vladimir
>>
>> On 7/18/19 8:28 PM, David Holmes wrote:
>>> Hi Ty,
>>>
>>> I'm moving this discussion to hotspot-dev as it's more appropriate.
>>>
>>> On 19/07/2019 12:46 pm, Ty Young wrote:
>>>> Hi,
>>>>
>>>>
>>>> I'm requesting that the long unimplemented "client" java switch be implemented in Java 14.
>>>
>>> Background: the client VM is historically only supported on 32-bit platforms explicitly, so the memory issues you are
>>> seeing are a combination of factors based on the ergonomic selections made by the VM during startup. The "client VM"
>>> is predominantly a 32-bit JVM that only supports the C1 JIT-compiler. The "server VM" in contrast supports the C2
>>> JIT-compiler. For a while now this distinction has blurred because the JIT uses tiered-compilation so that it starts
>>> by acting similar to the C1 compiler (for faster startup) and progresses into a mode that acts like C2 (for
>>> throughput optimisation). Though there are flags you can set to get it to act just like C1 or just like C2.
>>>
>>> Whether a machine is considered "server class" only partially relates to this. The startup ergonomics for a "server
>>> class" machine will configure subsystems to use more memory than a "non-server class" machine. Again these days (and
>>> for a while) we do not use this classification when starting the JVM. Various ergonomic selections are made based on
>>> the default settings for a range of components (mainly GC and JIT) together with the characteristics of the actual
>>> runtime environment (available memory and processors etc).
>>>
>>> The JVM is highly tunable in this regard, but of course it needs to have a reasonable out-of-the-box configuration -
>>> and that has evolved over the years, but is, at least for 64-bit systems, skewed towards server-style systems. So we
>>> cannot please everybody with the out-of-box default configuration. It's been suggested in the past that perhaps we
>>> should support a number of different initial configurations to make it easy(er) to adapt to specific user
>>> requirements, but this quickly breaks down as you can't get consensus on what those settings should be, and anyone
>>> who really cares will do their own tuning anyway.
>>>
>>> I can't go through your email point by point in detail sorry. Perhaps others can focus on specific memory issues. In
>>> particular if JavaFX is a source of problems then that will need to be discussed with the JavaFX folk.
>>>
>>> A very strong "business case" would need to be made for the community to look at supporting something like "-client"
>>> in the current OpenJDK.
>>>
>>> Cheers,
>>> David
>>>
>>>>
>>>> (Note: this entire request is based on the assumption that a JVM with -client is equivalent to a client JVM variant.
>>>> If this is wrong, I apologies. There isn't much documentation to go on.)
>>>>
>>>>
>>>> Since there aren't many google results or any kind of mention of this feature/ability even existing, i'll give an
>>>> explanation to the best of my knowledge and personal observations:
>>>>
>>>>
>>>> A "client" JVM variant is geared towards graphical end-user applications. According to a URL link found in the man
>>>> entry for java[1] this supposedly results in faster startups. While this *may* be true, a much larger and more
>>>> important benefit is a massive committed memory reduction in the range of about 25% to 50% when running a JavaFX
>>>> application. At minimum with similar heap sizes, that is a 75 MB memory savings at 300MB (a somewhat typical peak
>>>> usage with JavaFX applications) with a typical server JVM. That's huge.
>>>>
>>>>
>>>> The downside to this however is that at most, the maximum amount of (committed?) memory that a client JVM variant
>>>> can use is somewhere around 300MB by default. For the intended purpose of the client JVM switch/variant this is
>>>> *probably* fine. Server JVM variants only seems to allocate more memory to boost performance, which really isn’t
>>>> that much of a difference with the intended use case of the client JVM switch/variant… especially considering the
>>>> more appealing memory savings.
>>>>
>>>>
>>>> So why should this be implemented?
>>>>
>>>>
>>>> The answer is simple: using more memory then is necessary is bad, angers users, and frustrates developers who want
>>>> to be responsible by not wanting to eat up their users's memory[2] when it isn't needed.
>>>>
>>>>
>>>> Even if you've have never heard anyone complain about Java's memory usage, you've most likely heard someone complain
>>>> about a similar cross-platform software: Electron. People hate Electron applications for their absurd memory usage
>>>> and will actively avoid them by using alternatives if possible.
>>>>
>>>>
>>>> For reference, Etcher, an Electron application that allows users to easily create bootable USB drives on Windows,
>>>> Linux, and probably Mac OS uses around 298 MB just at launch on Linux. Electron is both comparable in both
>>>> goals(cross-platform solutions, JavaFX vs. Electron) and in memory usage.
>>>>
>>>>
>>>> Java may not be a native language and there may be *some* unavoidable penalty for that but being wasteful and
>>>> consuming resources where not necessary is, well, unnecessary. This can help reduce the amount of memory a java
>>>> application uses significantly when used.
>>>>
>>>>
>>>> With that all said, since JEPs include risks/impact/problems, it's best to mention some that come to mind:
>>>>
>>>>
>>>> Because of the default lower memory limit, applications which go beyond this will fail. The easiest and best
>>>> workaround would be to simply make the client JVM switch/variant opt-in. This would allow all existing Java
>>>> applications to continue to work as expected.
>>>>
>>>>
>>>> The only other issue that I can think of is people launching applications with -client without knowing the
>>>> limitations of it and filing bogus bug reports to app developers. This can be mitigated with better documentation
>>>> and awareness in places like the man page for Java. Since no one seems to really have used or knew about it before
>>>> it's more likely end developers that will be passing the switch to their applications via scripts then end users
>>>> will be.
>>>>
>>>>
>>>> All in all, this is pretty safe as long as server JVM switch/variant remains the default. Maybe others can think of
>>>> other risks/impacts/problems.
>>>>
>>>>
>>>> And finally addressing the two questions/comments I imagine someone at some point are going to ask/say:
>>>>
>>>>
>>>> Why not just compile a client JVM variant from source and use jLink?
>>>>
>>>>
>>>> and/or
>>>>
>>>>
>>>> If heap and garbage collection is healthy, who cares?
>>>>
>>>>
>>>> For the first one, yes, this is a route that could be taken. It has some problems however, namely:
>>>>
>>>>
>>>> - You have to be the developer or have source code access to use jLink.
>>>>
>>>>
>>>> - jLink -from my understanding- requires a **fully** modular Java application. Some used libraries may not be
>>>> modular yet.
>>>>
>>>>
>>>> - A full JDK source code compile is required - something that is really easy to do under Linux but might not be
>>>> under Windows and takes considerable CPU power to do. No one that I’m aware of (on Linux anyway) provides client JVM
>>>> variant builds. Presumably This is because the server JVM variant is the most versatile.
>>>>
>>>>
>>>> and as for the second: just because there is say, 5.8GB out of 8GB available doesn't mean you should or have the
>>>> right to use it as you see fit. People do more than use Java applications. If you are running a web browser with
>>>> lots of tabs open, a Java application could realistically cause major system stuttering as memory is moved to
>>>> swap/pagefile. While I used 300MB above as an easy realistic example, i've seen JavaFX applications consume as much
>>>> as 700MB and even 1GB committed memory. Just opening Scene Builder and playing around with the GUI consumes 400MB
>>>> easily on a server JVM variant(Oracle JDK/JRE 10 to be exact). While memory usage may never be as good as native,
>>>> the current amount of memory being consumed is insane and any normal user with standard amount of memory(6-8GB)
>>>> *will* feel this. Adding this switch could potentially help a lot here and give Java a slight edge over similar
>>>> software solutions.
>>>>
>>>>
>>>> Can this feature please be implemented? Likewise, could the documentation on what a "client" JVM and other JVM
>>>> variants be updated and improved?
>>>>
>>>>
>>>> [1] https://docs.oracle.com/javase/8/docs/technotes/guides/vm/server-class.html
>>>>
>>>>
>>>> [2] https://stackoverflow.com/questions/13692206/high-java-memory-usage-even-for-small-programs
>>>>
More information about the hotspot-dev
mailing list