It's like caching, in kind but not in type. Once you add it, people will stop trying to be parsimonious with resources and just reach for the cache every time. They'll just lean into it. In a hot minute you will discover you can't turn it off because people have lost their brains and the data flow of the app is through the cache and not through the call tree.
If you tune for allocation patterns that are in the code, then you are cementing those as continuing in perpetuity. Better to cut the fat first, so that you can tune for the necessary complexity instead of the accidental. That will be self-correcting because any new misuses will be taxed with higher performance regressions.
I worked on Java code at AWS for a few years and nobody tried to optimize allocations. Then I changed jobs and started working on a Java MPP database and my first code review was brutal. You were expected to avoid allocations as much as possible (mostly by using the SoA pattern everywhere). At that scale no GC could save you from excessive allocations.
I believe the whole string vs stringbuffer that later was made redundant by compiler contributed to that vision.
People started dismissing allocation discipline as a thing from the past because "that thing was solved a lot ago and the compiler now is smart enough".
Well, for string, yes, but not for arbitrary objects.
The most surprising allocation pressure I constantly run into is primitive boxing.
The JVM does heroics to try and avoid it as much as possible, but when you end up with some primitive boxing in a hotspot the amount of GC pressure that creates can be unreal.
The especially nasty thing about that boxing pressure is that when I still paid attention to the JVM, they had some sort of cache for boxing of numbers under about 128, which meant that if you did tests with toy inputs, identity checks would pass in code that needed an equality check instead, and benchmarks built the same way would see no GC pressure in those call trees but see quite a lot of pressure in production. I remember being really mad, almost to the point of a sense of betrayal, when I learned that. It's kind of a DieselGate but with more plausible deniability. You can't say it was deliberate, but if you already had trust concerns with that person it would move the dial farther into the red.
One of my least proud (most proud?) hacks when working with very large data sets is something like this
Map<Integer, Integer> intCache = new HashMap<>();
while (loading) {
Integer feild1 = intCache.computeIfAbsent(getField1(), (i)->i);
}
This is a terrible thing that shouldn't be as useful as it is to us... but it is really useful. We have a bunch of objects that can optionally have Integer values (hence a null is valid) but those int values are frequently the same.
This saves a bunch of memory and ultimately GC pressure as a result.
I do something similar for Java Time local dates. Financial data in particular has lots of redundant date info and benefits from being memoized. Converting to epoch millis also works.
There was a generation of date processing in Java when the class wasn't reentrant. The API was functionally flavored though, so people would instantiate one instance configured how they liked and reuse it across all calls. You would only see this problem under heavy load, such as in production, and it basically took encountering someone for whom it had already happened to spot what was going on. I think I found out from the Java issue tracker, thanks to some other dev filing an accurate bug report, but then I was that person for others at a full handful of other jobs afterward. Everyone was surprised, as one would be. They eventually fixed it but that was busted for a long long time.
Not really heap compression or special, it's just reusing a reference to an object already allocated on the heap.
Right now, if I do this
LocalDate a = LocalDate.of(2020, 1, 1);
LocalDate b = LocalDate.of(2020, 1, 1);
A and B reference 2 different object allocations on the heap even though they are the same date. a != b.
In Java, that can be pretty expensive even for an object as light as a LocalDate. By running the cache and doing
var cache = new HashMap<LocalDate, LocalDate>();
LocalDate a = cache.computeIfAbsent(LocalDate.of(2020, 1, 1), (i)->i);
LocalDate b = cache.computeIfAbsent(LocalDate.of(2020, 1, 1), (i)->i);
Now you have the situation where `a == b` and you immediately end up dropping the object allocation for b on the next GC.
The technique works best when you have a lot of repeated objects which are immutable. It is also only really needed because Valhalla isn't here. Once "value types" become a thing, then the representation for `LocalDate` inside the JVM can become just the fields and not a reference. The JVM is also free to do the sort of de-duplication optimization all on it's own for larger objects.
I mean FWIW this looks to me much like the sort of hacks I had to get into last time I worked with Java so I can't really blame you....
> Valhalla can't come soon enough for us.
This will be interesting to see for sure, I think it will raise the bar for competitors as well; .NET GC has lingered in some ways for a while in progressing [0][1], there is a long standing github discussion around a lower latency GC where there is a potential alternative [2] but nothing really signaling that it could be integrated in future as an option. Valhalla might finally put enough pressure on microsoft to do something about the lag in this space.
[0] - The inferred stackalloc stuff on the JIT level is awesome but I don't count that as improvement to actual GC
[1] - Pinned allocs took a long time and we still can't get aligned pinned allocs (have to manually pad instead)
I worked on projects where we indeed needed the manual stringbuffer refactor on hot paths in the code. It's not difficult work but boy is it tedious. Spent a lot of time with headphones on during those days. Only to work on another project a few years later where I paid for my sins by reversing the same change on someone else's codebase.
There's a version of Java, I can't recall which but I want to say 4? Maybe 3? Where someone rewrote parts of the Swing backend as native code to speed it up. Then Hotspot got good enough in the next version that the generated code was faster than the native code + FFI overhead. FFI was pretty high at the time. So they reverted the native code migration and went back to the old code.
In this post I look at a simple event to response latency benchmark, MarketDataSnapshot to NewOrderSingle at 50K/s for 30 minutes using JLBH to test Chronicle-FIX. The goal is to compare a system which is doing redundant work (in this case logging each message using SLF4J), compared with not logging (Chronicle-FIX records every message internally using Chronicle Queue) and how this changes the choice of Garbage Collector
Reading this gives me considerable pause - I can’t think of many classes within the codebase I work on that don’t have @Slf4j at the top…
Since there wasn’t a link to the source code in that post, can you help me understand this - for the SLF4J baseline is your logger impl a console appender, a file appender, or a network service like an OTel collector? Does any of that matter for GC context?
All common logging backends create a LogEvent or similar object for each logging call, and logging calls also typically construct new strings, which usually means a new StringBuilder object, its internal array (multiple ones if it grows), the final array it is copied to, and the String object that wraps that array.
These are typically short-lived objects and therefore cheap. Nevertheless, continually creating many such objects increases GC pressure, in particular if the logging happens in code that doesn't otherwise create many objects.
Cheap, not free, and even pretty simple to accidentally fool the GC on the lifetime of these objects.
Consider, for example, if you have a log message like this
logger.info("Hello {}", myOldObject);
if "myOldObject" is large enough or contains references to large things or has just been around for a while, it may be a part of OldGen at this point. And if that's the case, the LogEvent objects will end up automatically promoted to OldGen. Meaning the only time those can be be claimed is in an expensive major collection. The end result is that these things will ultimately fill up old gen and trigger more of the expensive old gen collections.
That's why it can be faster in some circumstances to write the more wordy
if (logger.isInfoEnabled()) {
logger.info("Hello {}", myOldObject.toString());
}
Nothing saves you, however, if your string being logged is too long. It can be autopromoted to old gen if you are trying to log a 10mb string.
> And if that's the case, the LogEvent objects will end up automatically promoted to OldGen.
Why do you think this would happen? There's no mechanism that makes young gen objects that reference old gen objects (or are referenced by old gen objects) get promoted faster. You have to survive a certain number of collections.
Ah shoot, you're right this wouldn't be what I'm thinking.
I was thinking of "GC Nepotism" [1]. That's the case where an object in old gen pointing to an object in new gen will automatically promote that new gen object into old gen. This can be particularly problematic with graph structures.
> All common logging backends create a LogEvent or similar object for each logging call, and logging calls also typically construct new strings, which usually means a new StringBuilder object, its internal array (multiple ones if it grows), the final array it is copied to, and the String object that wraps that array.
Which then gets discarded because that was a Log.verbose and your minimum log level in production is WARN.
Which is why many libraries have moved towards making your log message returned by a lambda. One constant lambda allocation (so, not a lot, an invokedynamic is absolutely fuck all.) that allows you to straight up skip allocating a full string that most likely is interpolating things and attempting to reach for context present on other threads is strictly better in 99.9% of the cases. The GC pressure is kept minimal and most importantly, constant.
> Which then gets discarded because that was a Log.verbose and your minimum log level in production is WARN.
This isn't true for the LogEvent or equivalent object, which only gets created after the log level is tested to be applicable by the logger implementation.
For call-site object allocation, you can wrap the logging call into an if statement that checks for the corresponding log level. The lambda allocation isn't constant if it captures anything from the surrounding scope, which will generally be the case for logging calls. (Unless by "constant" you mean that it's a single allocation per execution.)
> I don't like garbage. I don't like littering. My ideal is to eliminate the need for a garbage collector by not producing any garbage. That is now possible.
I prefer to see C++ as a failed experiment that just keeps going and going rather than garbage. The software industry learned a lot from it, both good and bad. But yea, I haven’t programmed in it since the late 1990s.
I'd phrase it differently, C++ was a set of power-to-performance trade-offs that were optimal in the 1990s.
Time has moved on.
More importantly, a typical 1990s C++ dev was likely someone who learned assembly, then C or C++. Meaning they already knew how to control hardware / memory allocation, and C++ was just a new set of abstraction tools. It was a step forward for them.
To modern devs, C++ is a step backwards. And a tough one at that.
Actually, C++ was rather poor in the 1990s if you ask me (albeit still very usable). Time has moved on - but so has the language. Its implementation tradeoffs were much better IMNSHO after 2011; but it wasn't there yet. And it still isn't! It has a lot of warts that have to stay for backwards compatibility (which is a design goal); and then, it has annoyances I can't believe are not yet addressed (like - where is my 'restrict' keyword, damn it?!)
Anyway, your view of the 1990s devs is incorrect. Almost no programmers who took up C++ learned assembly first (and few ever learned assembly). I believe most of them learned Pascal, or C, or scripting languages like Perl or Tcl or Unix shell scripts. Some may have learned Lisp or some ML variant as their first language, or Fortran 90. They didn't take a step back, they switched to a different set of language design goals and tradeoffs, and found it, well, serviceable.
It is indeed a bit peculiar that the language has had this much staying power. For C, it's much more understandable - because C is such a small and simple language (and one which, as you suggested, often feels like a bunch of syntactic sugar over PDP-7 assembly). But C++ is big, and has its baggage and warts and flaws. I think it's probably because it's been able to adapt and stretch just enough under the influence of trends in programming languages, for people not to ditch it for something new. Maybe Rust will change that; but - C++ might very well "eat its lunch".
> "because C is such a small and simple language (and one which, as you suggested, often feels like a bunch of syntactic sugar over PDP-7 assembly)."
I didn't describe such qualities of C in my previous post. At all.
And to be clear, I didn't mean "devs who learned how to program in the 1990s". I meant "devs who were active in the 1990s", which would have mostly been devs who came up in the 1970s and 1980s (where learning BASIC, then assembly was common).
I also made no claim about first languages in my previous statement. I only said there was a likely progression from assembly to C or C++ (for a C++ dev).
My point was, the talent pool in the 1990s would be able to handle the downsides of C++, and the upsides would have been a bonus. Whereas modern devs aren't used to the downsides, thus C++ is a step back for them.
What I should have added is that for devs in the 1990s, there would have been very few large legacy C++ codebases.
C++ has staying power because there was a surge in its use in the 1990s and early 2000s. Some of those people are now senior managers, who approve C++.
I'm using words using technology that would be much better built on C or lisp or smalltalk or any number of other languages. C++ is very popular because it was promoted in the right places, but is too abstract for low level tasks and too fiddly and fragile for high level tasks. It's like using a swiss army knife for construction when there's a whole box of better tools for every task.
Are you calling an entire generation of programmers idiots? Do you think they didn't know about C, Smalltalk or Lisp? C++ was the most popular language in the 90s because it was the best tool for a lot of jobs back then, people chose to use it willingly. Arguing that it only got popular because of PR is delusional, it had already gotten popular long before it got any corporation money behind it
I was part of that generation. C++ was the language that was taught in schools, and many many people were trying to get programming jobs that paid well, but they didn't really care about programming. They treated a programming language like a spoken language, and it never occurred to them how easy it was to learn another one. Many many corporate shops were C++ (or java) shops that wrote gang-of-four design pattern programs with no room for any creativity or experimentation, and a lot of that software is just designed to the moon and unnecessarily complicated. contrast that with the open source software of the time (which was more likely to be written in C then C++) and you'll see a clear quality difference.
It's like caching, in kind but not in type. Once you add it, people will stop trying to be parsimonious with resources and just reach for the cache every time. They'll just lean into it. In a hot minute you will discover you can't turn it off because people have lost their brains and the data flow of the app is through the cache and not through the call tree.
If you tune for allocation patterns that are in the code, then you are cementing those as continuing in perpetuity. Better to cut the fat first, so that you can tune for the necessary complexity instead of the accidental. That will be self-correcting because any new misuses will be taxed with higher performance regressions.
I worked on Java code at AWS for a few years and nobody tried to optimize allocations. Then I changed jobs and started working on a Java MPP database and my first code review was brutal. You were expected to avoid allocations as much as possible (mostly by using the SoA pattern everywhere). At that scale no GC could save you from excessive allocations.
I believe the whole string vs stringbuffer that later was made redundant by compiler contributed to that vision.
People started dismissing allocation discipline as a thing from the past because "that thing was solved a lot ago and the compiler now is smart enough".
Well, for string, yes, but not for arbitrary objects.
The most surprising allocation pressure I constantly run into is primitive boxing.
The JVM does heroics to try and avoid it as much as possible, but when you end up with some primitive boxing in a hotspot the amount of GC pressure that creates can be unreal.
The especially nasty thing about that boxing pressure is that when I still paid attention to the JVM, they had some sort of cache for boxing of numbers under about 128, which meant that if you did tests with toy inputs, identity checks would pass in code that needed an equality check instead, and benchmarks built the same way would see no GC pressure in those call trees but see quite a lot of pressure in production. I remember being really mad, almost to the point of a sense of betrayal, when I learned that. It's kind of a DieselGate but with more plausible deniability. You can't say it was deliberate, but if you already had trust concerns with that person it would move the dial farther into the red.
isn't that kinda how python preallocates integers -5 to 256? https://docs.python.org/3/c-api/long.html#c.PyLong_FromLong
Yep, and sometimes just a small code change can flip on boxing.
One of my least proud (most proud?) hacks when working with very large data sets is something like this
This is a terrible thing that shouldn't be as useful as it is to us... but it is really useful. We have a bunch of objects that can optionally have Integer values (hence a null is valid) but those int values are frequently the same.This saves a bunch of memory and ultimately GC pressure as a result.
Valhalla can't come soon enough for us.
I do something similar for Java Time local dates. Financial data in particular has lots of redundant date info and benefits from being memoized. Converting to epoch millis also works.
There was a generation of date processing in Java when the class wasn't reentrant. The API was functionally flavored though, so people would instantiate one instance configured how they liked and reuse it across all calls. You would only see this problem under heavy load, such as in production, and it basically took encountering someone for whom it had already happened to spot what was going on. I think I found out from the Java issue tracker, thanks to some other dev filing an accurate bug report, but then I was that person for others at a full handful of other jobs afterward. Everyone was surprised, as one would be. They eventually fixed it but that was busted for a long long time.
If I squint, is this a special kind of heap compression?
Not really heap compression or special, it's just reusing a reference to an object already allocated on the heap.
Right now, if I do this
A and B reference 2 different object allocations on the heap even though they are the same date. a != b.In Java, that can be pretty expensive even for an object as light as a LocalDate. By running the cache and doing
Now you have the situation where `a == b` and you immediately end up dropping the object allocation for b on the next GC.The technique works best when you have a lot of repeated objects which are immutable. It is also only really needed because Valhalla isn't here. Once "value types" become a thing, then the representation for `LocalDate` inside the JVM can become just the fields and not a reference. The JVM is also free to do the sort of de-duplication optimization all on it's own for larger objects.
Right. And this is safe now since Java Time classes are immutable. That’s key. Before, you had to worry about someone modifying the object.
Now I understand what Valhalla is (new JVM vesion!), I thought it was some kind of dark joke at first. Thanks for the explanation!
I mean FWIW this looks to me much like the sort of hacks I had to get into last time I worked with Java so I can't really blame you....
> Valhalla can't come soon enough for us.
This will be interesting to see for sure, I think it will raise the bar for competitors as well; .NET GC has lingered in some ways for a while in progressing [0][1], there is a long standing github discussion around a lower latency GC where there is a potential alternative [2] but nothing really signaling that it could be integrated in future as an option. Valhalla might finally put enough pressure on microsoft to do something about the lag in this space.
[0] - The inferred stackalloc stuff on the JIT level is awesome but I don't count that as improvement to actual GC
[1] - Pinned allocs took a long time and we still can't get aligned pinned allocs (have to manually pad instead)
[2] - https://github.com/dotnet/runtime/discussions/115627
can't wait for Project Valhalla, going into preview shortly
I worked on projects where we indeed needed the manual stringbuffer refactor on hot paths in the code. It's not difficult work but boy is it tedious. Spent a lot of time with headphones on during those days. Only to work on another project a few years later where I paid for my sins by reversing the same change on someone else's codebase.
There's a version of Java, I can't recall which but I want to say 4? Maybe 3? Where someone rewrote parts of the Swing backend as native code to speed it up. Then Hotspot got good enough in the next version that the generated code was faster than the native code + FFI overhead. FFI was pretty high at the time. So they reverted the native code migration and went back to the old code.
How does SoA help ?
In this post I look at a simple event to response latency benchmark, MarketDataSnapshot to NewOrderSingle at 50K/s for 30 minutes using JLBH to test Chronicle-FIX. The goal is to compare a system which is doing redundant work (in this case logging each message using SLF4J), compared with not logging (Chronicle-FIX records every message internally using Chronicle Queue) and how this changes the choice of Garbage Collector
Reading this gives me considerable pause - I can’t think of many classes within the codebase I work on that don’t have @Slf4j at the top…
Since there wasn’t a link to the source code in that post, can you help me understand this - for the SLF4J baseline is your logger impl a console appender, a file appender, or a network service like an OTel collector? Does any of that matter for GC context?
All common logging backends create a LogEvent or similar object for each logging call, and logging calls also typically construct new strings, which usually means a new StringBuilder object, its internal array (multiple ones if it grows), the final array it is copied to, and the String object that wraps that array.
These are typically short-lived objects and therefore cheap. Nevertheless, continually creating many such objects increases GC pressure, in particular if the logging happens in code that doesn't otherwise create many objects.
Cheap, not free, and even pretty simple to accidentally fool the GC on the lifetime of these objects.
Consider, for example, if you have a log message like this
if "myOldObject" is large enough or contains references to large things or has just been around for a while, it may be a part of OldGen at this point. And if that's the case, the LogEvent objects will end up automatically promoted to OldGen. Meaning the only time those can be be claimed is in an expensive major collection. The end result is that these things will ultimately fill up old gen and trigger more of the expensive old gen collections.That's why it can be faster in some circumstances to write the more wordy
Nothing saves you, however, if your string being logged is too long. It can be autopromoted to old gen if you are trying to log a 10mb string.> And if that's the case, the LogEvent objects will end up automatically promoted to OldGen.
Why do you think this would happen? There's no mechanism that makes young gen objects that reference old gen objects (or are referenced by old gen objects) get promoted faster. You have to survive a certain number of collections.
Ah shoot, you're right this wouldn't be what I'm thinking.
I was thinking of "GC Nepotism" [1]. That's the case where an object in old gen pointing to an object in new gen will automatically promote that new gen object into old gen. This can be particularly problematic with graph structures.
[1] https://psy-lob-saw.blogspot.com/2016/03/gc-nepotism-and-lin...
> All common logging backends create a LogEvent or similar object for each logging call, and logging calls also typically construct new strings, which usually means a new StringBuilder object, its internal array (multiple ones if it grows), the final array it is copied to, and the String object that wraps that array.
Which then gets discarded because that was a Log.verbose and your minimum log level in production is WARN.
Which is why many libraries have moved towards making your log message returned by a lambda. One constant lambda allocation (so, not a lot, an invokedynamic is absolutely fuck all.) that allows you to straight up skip allocating a full string that most likely is interpolating things and attempting to reach for context present on other threads is strictly better in 99.9% of the cases. The GC pressure is kept minimal and most importantly, constant.
> Which then gets discarded because that was a Log.verbose and your minimum log level in production is WARN.
This isn't true for the LogEvent or equivalent object, which only gets created after the log level is tested to be applicable by the logger implementation.
For call-site object allocation, you can wrap the logging call into an if statement that checks for the corresponding log level. The lambda allocation isn't constant if it captures anything from the surrounding scope, which will generally be the case for logging calls. (Unless by "constant" you mean that it's a single allocation per execution.)
I stopped reading early when it became clear the author didn't understand the terms (or perhaps the entire language) they were using. Several indicia.
Garbage collector?
To quote Bjarne Stroustrup:
> I don't like garbage. I don't like littering. My ideal is to eliminate the need for a garbage collector by not producing any garbage. That is now possible.
[flagged]
I prefer to see C++ as a failed experiment that just keeps going and going rather than garbage. The software industry learned a lot from it, both good and bad. But yea, I haven’t programmed in it since the late 1990s.
I'd phrase it differently, C++ was a set of power-to-performance trade-offs that were optimal in the 1990s.
Time has moved on.
More importantly, a typical 1990s C++ dev was likely someone who learned assembly, then C or C++. Meaning they already knew how to control hardware / memory allocation, and C++ was just a new set of abstraction tools. It was a step forward for them.
To modern devs, C++ is a step backwards. And a tough one at that.
Actually, C++ was rather poor in the 1990s if you ask me (albeit still very usable). Time has moved on - but so has the language. Its implementation tradeoffs were much better IMNSHO after 2011; but it wasn't there yet. And it still isn't! It has a lot of warts that have to stay for backwards compatibility (which is a design goal); and then, it has annoyances I can't believe are not yet addressed (like - where is my 'restrict' keyword, damn it?!)
Anyway, your view of the 1990s devs is incorrect. Almost no programmers who took up C++ learned assembly first (and few ever learned assembly). I believe most of them learned Pascal, or C, or scripting languages like Perl or Tcl or Unix shell scripts. Some may have learned Lisp or some ML variant as their first language, or Fortran 90. They didn't take a step back, they switched to a different set of language design goals and tradeoffs, and found it, well, serviceable.
It is indeed a bit peculiar that the language has had this much staying power. For C, it's much more understandable - because C is such a small and simple language (and one which, as you suggested, often feels like a bunch of syntactic sugar over PDP-7 assembly). But C++ is big, and has its baggage and warts and flaws. I think it's probably because it's been able to adapt and stretch just enough under the influence of trends in programming languages, for people not to ditch it for something new. Maybe Rust will change that; but - C++ might very well "eat its lunch".
> "because C is such a small and simple language (and one which, as you suggested, often feels like a bunch of syntactic sugar over PDP-7 assembly)."
I didn't describe such qualities of C in my previous post. At all.
And to be clear, I didn't mean "devs who learned how to program in the 1990s". I meant "devs who were active in the 1990s", which would have mostly been devs who came up in the 1970s and 1980s (where learning BASIC, then assembly was common).
I also made no claim about first languages in my previous statement. I only said there was a likely progression from assembly to C or C++ (for a C++ dev).
My point was, the talent pool in the 1990s would be able to handle the downsides of C++, and the upsides would have been a bonus. Whereas modern devs aren't used to the downsides, thus C++ is a step back for them.
What I should have added is that for devs in the 1990s, there would have been very few large legacy C++ codebases.
C++ has staying power because there was a surge in its use in the 1990s and early 2000s. Some of those people are now senior managers, who approve C++.
> Some of those people are now senior managers, who never bothered to wonder if better tools exist
Ftfy
You're wrote those words using technology built in C++. Just saying
I'm using words using technology that would be much better built on C or lisp or smalltalk or any number of other languages. C++ is very popular because it was promoted in the right places, but is too abstract for low level tasks and too fiddly and fragile for high level tasks. It's like using a swiss army knife for construction when there's a whole box of better tools for every task.
Are you calling an entire generation of programmers idiots? Do you think they didn't know about C, Smalltalk or Lisp? C++ was the most popular language in the 90s because it was the best tool for a lot of jobs back then, people chose to use it willingly. Arguing that it only got popular because of PR is delusional, it had already gotten popular long before it got any corporation money behind it
I was part of that generation. C++ was the language that was taught in schools, and many many people were trying to get programming jobs that paid well, but they didn't really care about programming. They treated a programming language like a spoken language, and it never occurred to them how easy it was to learn another one. Many many corporate shops were C++ (or java) shops that wrote gang-of-four design pattern programs with no room for any creativity or experimentation, and a lot of that software is just designed to the moon and unnecessarily complicated. contrast that with the open source software of the time (which was more likely to be written in C then C++) and you'll see a clear quality difference.