brunoborges 1 hour ago

Lots of details that teams miss when tuning the JVM on Kubernetes:

1. Default heap size varies: it can be 50% of the available memory for any amount of available memory up to 256mb; ~128mb for anywhere between 256 and 512; and then 25% for anything above 512mb;

2. The JVM will still do parallel threads even when CPU limit is 1,000 or below , which triggers CPU throttling much faster. Threads such as GC (even when it is SerialGC), ForkJoin pools, Database connection pools, HTTP server, and so on.

3. The JVM behaves MUCH better with 2,000 or more milicores, and a well sized heap. Having very small containers and dozens of replicas is worse than having slightly bigger containers (cpu/memory limits) and fewer replicas, even with a different ratio (e.g. 6 replicas will perform worse than 3 replicas, even when aggregated CPU and memory limits are the same).

I've been advocating on this for years [1].

[1] https://www.youtube.com/watch?v=wApqCjHWF8Q

Fizzadar 4 hours ago

I remember switching on G1GC for Elasticsearch back in 2014 and it was a total game changer for cluster stability (ES had plenty of other warts back then to counter the improvement). Cool to see it land as the default everywhere over a decade later.

  • cogman10 2 hours ago

    It's definitely gotten a LOT better since it was introduced. Java 17, in particular, really made G1 great.

    But if you've not already looked at it for your usages, I suggest people check out both the parallel collector and ZGC. ZGC in particular has been getting a lot of very nice features. Semi recently getting fast heap freeing (to the OS). The parallel collector for small heaps can generally beat G1 pause times.

    Where G1 really shines over the parallel collector is the fact that it regularly collects old gen. If your app has a lot of caches, then G1 is very good at keeping the VM from ending up with very long pauses.

  • pjmlp 1 hour ago

    It is not everywhere, because there are other JVM implementations, with their own GC algorithms.

ahmedfromtunis 4 hours ago

I love that they also have learn.java and inside.java.

They should make get.java point to the download page. Oh and docs.java would be very cool!

Sid_Lmao 3 days ago

Easy-to-miss detail: before JDK 27, the JVM only picked G1 with 2+ CPUs and ~1.8 GB of memory; otherwise, it silently used SerialGC. So a lot of 1- CPU Kubernetes pods have been on serial all along. Worth checking with -XX:+PrintCommandLineFlags before upgrading, since GC behaviour will change.

  • citrin_ru 5 hours ago

    Isn't the reason for this that on a single core SerialGC is better (lower CPU overhead at least)?

    • ch4s3 4 hours ago

      I'd imagine so. The kind of bookkeeping they're talking about here would negatively effect performance on a single core.

    • antonhag 4 hours ago

      Lower overhead yes - but G1 should have shorter pauses. And shorter pauses is probably worth it unless you run only background jobs in your jvms.

      • cogman10 2 hours ago

        Lower CPU and Memory overhead.

        Serial will be your best choice if your app is constantly using 100% of that core or if you need to minimize the amount of memory overhead.

        G1 is a better choice if you can be flexible with your memory and if you have a lot of periods in your app where it's below 100% CPU usage (maybe a lot of IO).

    • brunoborges 1 hour ago

      A 1,000 millicore CPU limit on Kubernetes does not mean the JVM can only use a single core. But that is indeed how the JVM interprets the limit: as a single core being available.

  • seabrookmx 2 hours ago

    > So a lot of 1- CPU Kubernetes pods have been on serial all along

    I don't think this is true.

    Setting your CPU limit to 1 in Kubernetes does not mean the container sees a single logical processor. It only means the kernel allows it to use 1 core worth of CPU _time_.

    If your k8s node is an 8 core node, Java will still see 8 logical processors. It just means you can only use 1/8th of each, or all of one, or some combination in-between.

    I've never seen a k8s node running on a single core VM, as most of the cloud providers only use these for their free tier or shared core VM's for the lightest of light workloads.

    • rileymichael 1 hour ago

      > Setting your CPU limit to 1 in Kubernetes does not mean the container sees a single logical processor

      that's incorrect, the jvm has been container aware for nearly a decade (https://bugs.openjdk.org/browse/JDK-8146115)

      you are correct about the underlying scheduling. limits don't pin you to a core, you get quotas across all available

  • brunoborges 1 hour ago

    This is correct. I have been advocating about this for years now [1].

    I have also seen customers wasting resources across Kubernetes because their developers would eventually tune JVMs with lowercase `-xmx` flag, which the JVM ignores, and thus by default, the JVM only sets heap size to 25% by default (for => 512mb of available memory).

    [1] https://www.youtube.com/watch?v=wApqCjHWF8Q

  • pjmlp 1 hour ago

    Open JDK, not The JVM.

    There are other implementations.

    • hashmash 18 minutes ago

      Not really, other than OpenJ9. In my experience, it's way behind OpenJDK in terms of performance and stability.

      • pjmlp 8 minutes ago

        Then you never embedded development with real time JVMs, for example.

cyberax 2 hours ago

And still no automatic heap sizing... Maybe next year?

  • karussell 2 hours ago

    What do you mean? G1 definitely has automatic heap sizing, i.e. Xmx just defines the maximum it is allowed to use.

    edit: looks like you probably mean sth like this: https://openjdk.org/jeps/8359211

    • cyberax 2 hours ago

      Indeed.

      One great feature in Go (or Python, or Node JS) is that you don't have to care about setting the -Xmx. The runtime will just use more RAM as needed, and you can use normal mechanisms like cgroups to limit the maximum size.

pancakeguy 1 hour ago

garbage collection will almost certainly be a relic of the past surprisingly soon. Fundamentally it exchanges efficiency for readability & developer experience, to things that do not matter when using AI to write code.

  • ogogmad 1 hour ago

    If true, it's sad. So much work and so much accumulated wisdom. It's like seeing a cathedral built by thousands of hands over dozens of years collapse in an earthquake.

    • jitix 1 hour ago

      Disagree, we need to move ahead with progress.

      Hobby and domain specific languages may still use GC but getting rid of the concept of GC is the right direction, along with removing different ways of doing things in a programming language. Go was right when it forced developers to use only pattern for loops, one pattern for error handling, etc.

      • titzer 56 minutes ago

        Memory management is a resource issue that should be orthogonal to the logic of code. Moving away from GC is not the right direction for writing smaller, more concise, and more correct programs.

        I don't think that making programs a lot more complicated because they have to prove a very specific memory management discipline is the right way to go.

        • jitix 35 minutes ago

          Programming languages are just an intermediate representation of the solution, just like bytecode or assembly.

          With AI we don't need expressive programming languages for humans to read and understand at line-level details. Instead, we need stronger guardrails and fewer building blocks so that AI generates consistent solutions.

          GC became poplar only because it was hard for programmers to keep track of memory management in languages like C and C++. And back then I was a huge proponent of GC based languages.

          This is just a step along the same path that Go introduced. Many programmers switching from other languages complain about Go's restrictive and seemingly verbose syntax but once the benefits are realized the tradeoffs become worth it.

  • IncreasePosts 1 hour ago

    Doesn't this assume infinite abundance of compute? For most problems, using a GC is probably easier for AI than to track everything manually(just like with humans). Similarly, we could probably ask AI to output everything in optimized assembler code, but it will almost certainly burn more tokens getting to a solution than if we allow it to write in python or some other HLL

  • xxpor 1 hour ago

    Disagree. Maybe the naive business app use case will, but there are plenty of places where you want to implement effective GC regardless. Anywhere you batch frees to avoid calling it in a hot path is a GC. So it's not _always_ an efficiency sacrifice. Or, a different perspective is you always have to ask "efficient with what resource?" CPU time efficiency vs memory efficiency for example (an interesting case where the _money efficiency_ has obviously moved severely in the past year!).

  • sealWithIt 1 hour ago

    On the one hand, I agree with you; on the other, I also find reasonable that lower cognitive load may keep justifying the existence of GC-based languages even for AI.

  • titzer 58 minutes ago

    No, manual memory management, with or without static ownership, is not inherently more efficient than GC. Bump pointer allocation is crazy fast. GC work is proportional to live data, not dead data. GC can do compaction and locality optimizations. Manual deallocation is proportional to dead data, unless you are doing bulk deallocation with regions. Stack allocation is crazy fast. It's a little beyond what state-of-the-art compilers do by default, but it's just pointer analysis that will also get better with AI optimizing and improving compilers.

    • paulddraper 29 minutes ago

      GC is fast.

      GC is not memory efficient.

      If a primary concern is optimizing memory usage, you will avoid GC languages.

  • jeffbee 53 minutes ago

    Can't agree. GC language platforms are more efficient in many ways. Look at the hoops that C allocators try to jump through just to get decent THP coverage which a compacting GC gets essentially for free.

  • quotemstr 32 minutes ago

    Why would you think that malloc/free style memory management is more efficient? Both functions can be quite expensive, and they work synchronously. GCes give you compaction, compressed pointers, and bump pointer young generation operation as well as offloading allocation machinery to threads other than your business logic one. It's manual memory management that's going the way of the dinosaur, and the current interest in manual memory management is a fad.

  • hashmash 31 minutes ago

    Implementing efficient concurrent algorithms without a tracing collector is quite difficult, and it often requires the use of hazard pointers or epochs. Both act like mini collectors, and they also require special tuning.

    Unless you want AI generated code to be single threaded and not take advantage of multiple cores, some form of collector is still required. If you want to avoid heap fragmentation issues, a compacting collector is generally recommended.

_verandaguy 5 hours ago

Tangential, but I really dislike that the `.java` gTLD was allowed to come about given it's also a reasonably popular file extension.

The `.zip` gTLD set a really bad precedent.

  • lmz 5 hours ago

    .zip? Not .com?

    • dbt00 3 hours ago

      The potential for confusion in 1985 wasn't really on anyone's radar.