I found claude and GPT very helpful on this, because java have a very sofisticated monitoring harness. Just ask the agent to connect to the running application (on kubernetes or whatever) on prod and do a java flight recording then analyze allocations.
I managed to improve some applications of ours from several garbage collections per second to several minutes between collections. That _really_ improves p99.
Take a peek at jolokia if you've never used it :) Attach it as a JVM agent to _any_ process to expose it's JMX attributes as http calls, and it even gives you a neat little console to log into.
If you want to collect these, telegraf has a jolokia plugin. It's an incredible combination!
Nothing much smarter than run the agent in the project folder, make sure it has means to interact with production (like configured kubectl) and ask it to jcmd the hell out of it.
I think it said something very meaningful and actionable (admittedly at a high level) about keeping the hot path deterministic: low allocations (including guidelines for GC frequency targets), don't use vthreads (instead use one thread on a pcore that never sleeps, to avoid amdahl), caution about techiques that work at p99 and fail at 99.999... This is not typical for "enterprise" Java, and is an interesting article for those who wish their systems were more deterministic (but may not be aware of the techniques or their cost). This is not a "how to", but that's fine. It's still good content.
Yes, disappointing given that Chronicle actually does have legit expertise here. (Chronicle Map is not very well-known even in the Java space but it's by far the best larger-than-memory Map available.)
Had to write a router software once in Java back in 2014. The past experience in me believed that its not a good choice to use something with stop the world GCs for routers. The goal was to have everything under 1ms (max 2ms in rare cases). It has http server on head and websocket connections on the tail end. Turns out the class instantiations and JSON parsing were a bottleneck. Thread safe resources were valuable, the libraries which did not had thread safe resources were re-written to support what we needed. The http server was re-written barebones, the JSON parsing was re-written with barebones. And it worked.
Yes. & in the end my experience has been that you can write low latency code with gc, but only by expending more effort than it'd take to use manual memory management
First of all it depends on the language, what GC models it supports, and what are the performance SLAs.
Secondly if we are talking about something like real time audio, or sending the maximum amount of triangles per second, then there is zero memory management, the code should be allocation free across all the critical path, regardless of the language.
How about not using Java? Then you can have low latency.
Average go, rust, c++ and c will outperform amazing java programs, and the former will also be way way more easy to run, troubleshoot, interpret logs from.
I need a UI which runs well on Windows, MacOS, and Linux, without having to build three different ones. Swing is still easily the best, most consistent, and most native-feeling cross-platform environment. It's much better than QT and GTK in most respects. And Java also runs elegantly on a little platform you may know as Android. I have high hopes for go and rust. But until they have mature UIs, they're out (for me).
C and C++ are dangerous languages filled with security failings and footguns, and no modern app should be written in them.
It's been my experience that well-written low-level Java code runs at about 75% the speed of good C code. (Of course lazy coders write in cushy Java which is much slower). When written efficiently, Java's biggest slowdown lies in array access (C and C++ array access is fast because it is very, very unsafe). But Java makes up for this in having a GC which will coalesce related objects into the same page and so take advantage of cache coherency effects in ways malloc and free cannot possibly do. I have some allocation-heavy algorithms in Java which are, as a result, significantly faster than well-written equivalents in C.
In heavily performance-engineered code having a GC coalesce memory is a pessimization in all cases.
I've done a lot of performance engineering in both C++ and Java. Every optimization available in Java also exists in C++ but the reverse is not true, which is why C++ is always faster. Every example I have ever seen of Java being faster than C++ was just poorly optimized code. The heuristic I use is that heavily optimized C++ is about twice as fast as heavily optimized Java at the limit. And this requires some non-idiomatic and ugly Java that isn't nice to maintain.
This doesn't necessarily make Java the wrong choice though. Few organizations prioritize absolute performance above all other things. There are practical tradeoffs.
May be true between late 90s to late mid 10s. Both Java and JVM has had enormous of work going into it. 2026 the JVM is pretty damn good pieces of engineering.
I've done some tests with current value classes in latest prototype with full optimisation through annotation. Apparently, everything was being compiled to assembly code with no gc calls. It seems this old school ideology of slow java is about to end in near future. I'm actually intrigued how it will compete with Rust which can't optimise code further while running, while JVM JIT has more info which it could aggressively further optimise, especially hot paths.
If low latency is your goal than you don't want JIT. JIT has two issues in low latency, first the first time through your code isn't compiled yet and so you get high latency. Second, it optimizes for the common case, which means when you hit an exception that exception will be higher latency because everything hits a branch miss.
Of course we are talking generals here. Sometimes the above is acceptable and Java/JIT is just fine. Sometimes it is unacceptable and you cannot use Java/JIT. Know your domain.
Of course in all cases (C, Rust, C++), you have to understand the system and what it is doing. Every language has a standard library that will do things that are not low latency on you. You have to know which library functions will do what, memory allocation and copies are both things that code often does without thought that are incompatible with low latency. No matter what you need to know what your language does that is against you.
Valid points, but this doesn't mean JIT doesn't work for relatively lower-latency coding.
To avoid the compilation etc. hit, common practice is to do some "warmups" before serving users. (Another reply has other ways to avoid this hit.)
Handling exceptions is higher latency, but they can/should be optimized out, so you're not hitting exceptions as part of your standard workflow (or even your 1% workflow).
> Average go, rust, c++ and c will outperform amazing java programs
Not true. Many benchmarks have shown otherwise. it is at least competitive in many areas.
> and the former will also be way way more easy to run, troubleshoot, interpret logs from
No language will save you from poor logging practices. If you log every debug log it's not Java's problem. No 1 says you have to log the full stack trace if that's your concern. You can configure / strip / do anything. Learn to use the stack.
I found claude and GPT very helpful on this, because java have a very sofisticated monitoring harness. Just ask the agent to connect to the running application (on kubernetes or whatever) on prod and do a java flight recording then analyze allocations.
I managed to improve some applications of ours from several garbage collections per second to several minutes between collections. That _really_ improves p99.
I work on OpenJDK, so I may be biased, but I’ve also found that JFR works really well with LLMs.
Then ask Codex, or whatever AI tool you use, to analyze report.txt for issues and use all.txt to dig deeper, if needed.> but I’ve also found that JFR works really well with LLMs
There's also async profiler. LLMs just use whatever. Lots of choices.
Take a peek at jolokia if you've never used it :) Attach it as a JVM agent to _any_ process to expose it's JMX attributes as http calls, and it even gives you a neat little console to log into.
If you want to collect these, telegraf has a jolokia plugin. It's an incredible combination!
That sounds like a great idea. What kind of prompt do you use?
Nothing much smarter than run the agent in the project folder, make sure it has means to interact with production (like configured kubectl) and ask it to jcmd the hell out of it.
I get that blog posts often advertise a company's products, but this one had absolutely zero content other than advertising.
This is a high-level article; for lower-level details, I also posted.
https://blog.vanillajava.blog/2026/06/testing-java-memory-ma...
https://blog.vanillajava.blog/2026/06/why-you-should-tun-cod...
Suggestions welcome, it can be hard to know what to include or not.
I think it said something very meaningful and actionable (admittedly at a high level) about keeping the hot path deterministic: low allocations (including guidelines for GC frequency targets), don't use vthreads (instead use one thread on a pcore that never sleeps, to avoid amdahl), caution about techiques that work at p99 and fail at 99.999... This is not typical for "enterprise" Java, and is an interesting article for those who wish their systems were more deterministic (but may not be aware of the techniques or their cost). This is not a "how to", but that's fine. It's still good content.
Yes, disappointing given that Chronicle actually does have legit expertise here. (Chronicle Map is not very well-known even in the Java space but it's by far the best larger-than-memory Map available.)
Had to write a router software once in Java back in 2014. The past experience in me believed that its not a good choice to use something with stop the world GCs for routers. The goal was to have everything under 1ms (max 2ms in rare cases). It has http server on head and websocket connections on the tail end. Turns out the class instantiations and JSON parsing were a bottleneck. Thread safe resources were valuable, the libraries which did not had thread safe resources were re-written to support what we needed. The http server was re-written barebones, the JSON parsing was re-written with barebones. And it worked.
Even C requires discipline to write low latency code, if you think otherwise, you never used a profiler.
Yes. & in the end my experience has been that you can write low latency code with gc, but only by expending more effort than it'd take to use manual memory management
First of all it depends on the language, what GC models it supports, and what are the performance SLAs.
Secondly if we are talking about something like real time audio, or sending the maximum amount of triangles per second, then there is zero memory management, the code should be allocation free across all the critical path, regardless of the language.
And Compiler Explorer also
On mobile the whole page text is aligned center which makes it really hard to read.
page doesn't load "En attente de la réponse de chronicle.software."
+ low latency anything requires discipline. if you lose 5ms you can't get it back.
How about not using Java? Then you can have low latency.
Average go, rust, c++ and c will outperform amazing java programs, and the former will also be way way more easy to run, troubleshoot, interpret logs from.
Java is usch garbage in every stack.
I need a UI which runs well on Windows, MacOS, and Linux, without having to build three different ones. Swing is still easily the best, most consistent, and most native-feeling cross-platform environment. It's much better than QT and GTK in most respects. And Java also runs elegantly on a little platform you may know as Android. I have high hopes for go and rust. But until they have mature UIs, they're out (for me).
C and C++ are dangerous languages filled with security failings and footguns, and no modern app should be written in them.
It's been my experience that well-written low-level Java code runs at about 75% the speed of good C code. (Of course lazy coders write in cushy Java which is much slower). When written efficiently, Java's biggest slowdown lies in array access (C and C++ array access is fast because it is very, very unsafe). But Java makes up for this in having a GC which will coalesce related objects into the same page and so take advantage of cache coherency effects in ways malloc and free cannot possibly do. I have some allocation-heavy algorithms in Java which are, as a result, significantly faster than well-written equivalents in C.
In heavily performance-engineered code having a GC coalesce memory is a pessimization in all cases.
I've done a lot of performance engineering in both C++ and Java. Every optimization available in Java also exists in C++ but the reverse is not true, which is why C++ is always faster. Every example I have ever seen of Java being faster than C++ was just poorly optimized code. The heuristic I use is that heavily optimized C++ is about twice as fast as heavily optimized Java at the limit. And this requires some non-idiomatic and ugly Java that isn't nice to maintain.
This doesn't necessarily make Java the wrong choice though. Few organizations prioritize absolute performance above all other things. There are practical tradeoffs.
>Java is usch garbage in every stack.
May be true between late 90s to late mid 10s. Both Java and JVM has had enormous of work going into it. 2026 the JVM is pretty damn good pieces of engineering.
I've done some tests with current value classes in latest prototype with full optimisation through annotation. Apparently, everything was being compiled to assembly code with no gc calls. It seems this old school ideology of slow java is about to end in near future. I'm actually intrigued how it will compete with Rust which can't optimise code further while running, while JVM JIT has more info which it could aggressively further optimise, especially hot paths.
Rust? OK.
C++ or go? Then you'll have to take a very closer look, because the java JIT is wonderful. A masterpiece of several hands, actually.
If low latency is your goal than you don't want JIT. JIT has two issues in low latency, first the first time through your code isn't compiled yet and so you get high latency. Second, it optimizes for the common case, which means when you hit an exception that exception will be higher latency because everything hits a branch miss.
Of course we are talking generals here. Sometimes the above is acceptable and Java/JIT is just fine. Sometimes it is unacceptable and you cannot use Java/JIT. Know your domain.
Of course in all cases (C, Rust, C++), you have to understand the system and what it is doing. Every language has a standard library that will do things that are not low latency on you. You have to know which library functions will do what, memory allocation and copies are both things that code often does without thought that are incompatible with low latency. No matter what you need to know what your language does that is against you.
Valid points, but this doesn't mean JIT doesn't work for relatively lower-latency coding.
To avoid the compilation etc. hit, common practice is to do some "warmups" before serving users. (Another reply has other ways to avoid this hit.)
Handling exceptions is higher latency, but they can/should be optimized out, so you're not hitting exceptions as part of your standard workflow (or even your 1% workflow).
> If low latency is your goal than you don't want JIT. JIT has two issues in low latency...
There's startup "AOT cache" via Leyden that speeds up startup. Isn't native speed up it's quite a big boost.
Then there's GraalVM that does give you a native image. Real AOT.
It is a matter of skill.
yes, until you need debugging.
> Average go, rust, c++ and c will outperform amazing java programs
Not true. Many benchmarks have shown otherwise. it is at least competitive in many areas.
> and the former will also be way way more easy to run, troubleshoot, interpret logs from
No language will save you from poor logging practices. If you log every debug log it's not Java's problem. No 1 says you have to log the full stack trace if that's your concern. You can configure / strip / do anything. Learn to use the stack.
yeah...no, try low latency when running out of memory, or with more CPU-bound threads than cores...discipline vs almost impossible