A GC latency SLO should include operating-system memory pressure. Otherwise, a page-fault problem will look like a collector problem and lead to the wrong fix.
This is not entirely correct. If you care about latency, then it doesn't matter on which major fault your application gets paused. Disabling swap protects you from data pages being evicted, but code pages can still be paged out.
If you care about latency, mlock() your memory, do not disable swap. Swap is good and gives the kernel an equal opportunity to evict data and code pages.
For GC enabled languages swap is universally bad. Some gc-pauses are indistinguishable from a system crash. It's a side effect on not having tightly specified memory limits.
I'd rather have applications be oom_killed than having them swap out, the former is rather obvious and demands action.
If you want you application to stay in memory, then make it explicitly with mlock()/mlockall().
Disabling swap will just moves pressere elsewhere: to code pages. And evicted code page is no better: full stall while kernel loads that page from disk.
Under memory pressure the system will free the ram used for code pages because it can always load them back from the executable on disk. It’s the same virtual memory mechanism as swap but without needing dedicated swap space. The op is saying disabling swap doesn’t prevent long pauses during memory pressure, because the system just swaps code out instead of dynamically allocated memory.
There are two types of pages: anonymous and mapped from files. The code pages are mapped from a binary; they are not very special.
Under memory pressure, the kernel evicts less popular pages from memory. If a page has been mapped from a file, it is dropped (if dirty, then it is written out first). If it is needed later, the kernel can read it back from the file. If a page is anonymous (read: heap page), then there is no backing file and the kernel copies it to swap before dropping it. This is swapping.
So, what happens if you disable swap and the kernel is low on memory? What can it evict? Anonymous pages cannot be evicted: there is no swap to put a copy in. The only choice the kernel has is to evict pages that are mapped from files. Those include pages mapped from the executable. You don't eliminate stalls by disabling swap, you just move them elsewhere: the kernel will page out code and your app gets paused whenever the execution flow hits such a page.
>So, what happens if you disable swap and the kernel is low on memory? What can it evict?
Your userspace early OOM killer triggers and lets you know you're trying to run more than will fit in memory so you don't do it again. (In my experience, the kernel OOM killer can't be trusted to kill processes soon enough.)
I haven't had to analyze the performance of no-swap processes before. My assumption is that code is hot enough to avoid eviction and that evicted code pages are rather the exception. To strong-man the argument, I can imagine long running complex (bloated) services could have parts that are not touched unless a specific request comes in.
While this is technically true, and useful to know, it's not useful to think of this problem in these terms.
In a scenario like the one describe, one is not running out of memory because of file-based pages. That's <1% of memory usage in typical cases. One is almost always running out of memory because of allocated anonymous pages.
You are missing that the file LRU is not only the code pages or memory mapped files but all file data. The active code pages are specifically excluded from being discarded during the reclaim process.
I wish this is true but evidence is against you. Just try testing memory pressure on a no swap system. You will see that the system grinds to a half for a few minutes before you even get to an OOM event.
Could you clarify what exactly is "not true"? Also, could you explain what exactly your evidence proves? What do you think happens when the system becomes unresponsive under memory pressure, and why?
I'll try again. Why do you think that unresponsiveness is caused by code pages being discarded? Do you have evidence for this? Have you considered that there are other mechanisms that cause programs to become unresponsive under memory pressure?
It's the explanation that I've read when scouring forums for reasons why it happens. I'd love to hear alternative explanations and how to measure/verify them. But the measurement tools also become unresponsive so I can't check anything.
But frankly, as a user I don't really care what the real technical reason is. I just don't think the system should become unresponsive for minutes on end during memory pressure. Just OOM-kill earlier, when it's supposed to.
The assumption that a system gets unresponsive due to code pages being reclaimed doesn't make any sense. It is a common myth, cargo-cult knowledge. The kernel doesn't discard any active referenced pages at all, and the active code pages are the last that are getting reclaimed.
The proper explanation is relatively straightforward. Normally, any memory request for a process is served by getting a page from the free memory pool. But when the system free memory goes below the low watermark (calculated from vm.min_free_kbytes and vm.watermark_scale_factor), the kernel stops giving the pages from the free memory pool and instead goes into direct reclaim, trying to find unused pages. While the kernel does reclaim, the allocating process remains stalled. The kernel can't return control to the process - there is no page mapped that the process tried to access. This applies to any types of pages - anon, file-backed, code pages. So, under the condition of free memory below the low watermark, any process trying to allocate memory gets stalled. This is measured as memory pressure stalls (PSI) in /proc/pressure/memory.
The problem with the system being stuck in an unresponsive state is that the design choice for the kernel OOM handling is to be optimistic and keep the workload running as long as the kernel is making any progress reclaiming pages. The direct reclaim process needs to fail to make any progress 16 times before the OOM killer is invoked.
The solution is to use cgroup memory limits to avoid getting into the situation of memory exhaustion and use user-space oomd to recover from this situation earlier.
So you're saying processes are stuck for minutes on end at trying to get memory from the kernel? I find this implausible because relatively simple processes such as ssh and bash shouldn't be getting memory from the kernel at a continuous basis. Unless they're suddenly processing a lot of new data (like me trying to cat /dev/urandom), they should be spending most of the time working with the memory they already have. Especially when all I'm doing is pressing keystrokes in ssh+bash, not even hitting enter. Note that malloc doesn't ask memory from the kernel at every call. Glibc malloc is (in)famously stubborn with caching memory it has gotten from the kernel, hesitant to release it back to the kernel.
Any repeatable way to prove your thesis or disprove the trashing-code-pages thesis?
You seem to confuse user-space memory allocations (malloc) with the kernel memory page allocation and mapping. The processes don't need to allocate any new memory with malloc. They get stuck when accessing memory that is not mapped to a physical memory page or when they enter a syscall that needs in-kernel memory allocation.
Ok fine. But again: any repeatable way to prove your thesis or disprove the trashing-code-pages thesis? And how to do this when the system is freezing up?
I have already pointed you to /proc/pressure/memory. Prometheus node_exporter tracks it as node_pressure_memory_waiting_seconds_total and node_pressure_memory_stalled_seconds_total. You'll see them spiking when you experience stalling.
You can also track variables from /proc/vmstat exported as node_vmstat_allocstall_(normal|dma32|movable), node_vmstat_pgscan_direct, node_vmstat_pgsteal_direct - you'll see them spike when the system enters direct reclaim.
Install and configure systemd-oomd and track its log - it will activate on memory pressure.
To track code page eviction/re-faults specifically, you'll need to write some eBPF code, there are no direct measurements for this available.
Perhaps I wasn't clear, but I did say that all file pages are treated equally.
I would assume that page with PC is excluded from reclamation process, but other code pages are valid targets.
All recently referenced code pages in the active LRU are excluded from reclaim.
Code pages are paged in on demand: for any sufficiently large binary just some code pages will be mapped right after start. Given enough memory pressure, the active LRU will be shrunk often enough to push code pages out of the active LRU and then they are eventually evicted.
I'm not aware how it works in modern MGLRU, but I would assume similar ovservable behevior holds.
Active code pages with the "Accessed" bit set are specifically excluded from being moved to inactive LRU when active LRU is being shrunk. Instead, they are moved to the head of the active LRU. All other pages are moved from active LRU to the inactive LRU unconditionally, even if they have the "Accessed" bit set. So, no, the active code pages are not going to be eventually evicted.
I think we are saying the same thing. Active code pages are not evicted; non-active code pages (that is, pages that had not been accessed since last time of active LRU shrinking) may be evicted.
There are good sibling replies answering your question, but to give a specific example: if your application writes something but then doesn't need it anymore. Since we're talking about go, perhaps a data structure you need also holds a reference to data you don't need.
When the kernel needs memory, it goes hunting for a page it can discard. But since that's transparent, the kernel can only discard a page if it knows it can get it back (after all, it's still got valid data on it).
If there's swap, a page full of stale/unneeded data can be written out to swap. But if there's no swap, your page of "dangling data that you'll never use, but is still valid & referenced" can't be discarded; the kernel doesn't know you won't want it later, and it can't recreate the page if it throws it away.
So like sibling said, at that point it has to find other pages it can evict from memory, ones that _do_ have somewhere persistent they can be written out to. Pages loaded from binaries on disk satisfy that, so those will get dropped instead.
For JIT languages most of the code is dynamically generated, not available to be loaded back from the disk. Morealso, the allocations do happen at large chucks where the oom_killer also happens, so it's significantly less of an issue.
That doesn't necessarily improve matters. Now any code that gets JIT'd is also going to be permanently pinned in memory too, no matter whether it'll get used again, if you don't have swap.
I think discussing it in a vacuum is pointless, it should depend on how much memory you have. You can have a swap on, but with a low `vm.swappiness` number.
Surely as long as it is all in memory these pauses are not going to be that significant, details like that should be left for the language devs to optimize.
You can tell an average Golang user to use mutexes, you shouldn't be telling them to lock memory manually.
Maybe, but it still there and probably has a different value between distros as well. Default value is 60, what is that good for exactly?
If you're in a position where you still want swap to be there, but only used in extreme situations, then you should set it yourself. Otherwise disable swap or switch a language.
What do you think this value of 60 means? Also, as I said, in extreme situations, when the system is right before invoking the OOM killer, swap is used regardless of `vm.swappiness` setting.
Value 60 is not an arbitrary number. It makes the kernel reclaim 140 file pages for each 60 swapped-out anon pages. This proportion is supposed to reflect the relative performance cost between sequential file access and random swap I/O on HDD. For SSDs, the value 100 is more appropriate.
> I don't see a reason why the default shouldn't be 0 then on modern systems.
The zero value makes reclaim very unbalanced and is not appropriate for generic systems. This will cause the kernel to discard all the page cache, slowing down the system with uncached I/O, while keeping all unused anonymous pages in memory.
I don't get the point that you're even arguing. Are you saying that 60 is fine as a default? Or are you saying that people should even think about vm.swappiness because it doesn't matter? The alternative here is to disable swap completely, that's what started this whole thread.
While I've seen systems grinding to a halt due to evicting code pages, the idea that disabling swap distributes load between code pages and data pages makes no sense to me. Data pages are also disk backed, so they could also be evicted at any time. Why would data pages be written to swap if they came fron the disk in the first place?
So that only leaves anonymous data pages, i.e., regular data in memory, that could be swapped. Ok but you're not easing the load on code pages, those still get evicted during memory pressure whether you have swap or not.
The assumption that data pages can always be easily discarded is incorrect. Dirty data pages can't be discarded until they are written back. Recently accessed data pages are also not discarded until all inactive data and code pages are discarded.
FYI: This `delamon` account seems to be meat proxying us. Just relaying garbage AI answers without understanding his own words. Doubt he's ever used mlock himself for anything.
I see that you've rewritten your AI generated reply there. So when have you used mlock in production for Go services, the way you're advising others to do?
Yes, I was and still running go and zig binaries with mlockall. Apps have large in-application caches and also perform large streaming reads of datasets bigger than available RAM. If it happens that several streaming reads running at same time, they push memory pressure hard enough, so kernel swaps out cold parts of in-app caches. That later leads to elevated latencies => mlockall to avoid that.
Run your process in a cgroup that's set to disallow swap, then. The rest of your system benefits from having swap. This was an easy and safe thing that required minimal convincing to roll out to production at my company.
Okay, to be more explicit: disable swap and ensure you have sufficient memory for your workload. Your 14 MB Go binary is very unlikely to be the source of your memory pressure, it's almost always going to be the allocated anonymous pages.
The advice of using mlock is just wrong in the case of a Go server allocating memory for its heap. Don't try to do that. Have sufficient memory, monitor it, and let the Go GC and kernel manage memory normally.
Can you elaborate, why mlock is bad? And how disabling swap is better?
You are correct, that 90MB Go binary is not major memory consumer. But it doesn't matter what is the source of memory pressure. What matters is what kernel do under pressure. If you disable swap, then under memory pressure kernel can only pageout memory mapped files, no matter how small they are. And your code will be paged out.
We're trying to solve a problem: we don't want huge latency spikes caused by swapping memory to disk. The only good solution to that is to have sufficient RAM. It's that simple.
Introducing mlock is operationally annoying and trying to solve the wrong problem. It requires elevated container permissions, it's not something your SRE team is going to expect, almost no one does it, and it has all kinds of technical implications and interactions with kernel OOM system, forking, etc.
There are very specific scenarios where mlock is exactly the right solution but the typical Go HTTP service is not one of them. The KISS principle applies.
That would work reasonably well. With one important caveat: memory controller tracks all memory, including page cache. Setting memory.min will prevent reclaim of page cache charged to that cgroup, as long as total memory usage stays below the limit.
Another important caveat: people keep assuming that page cache can be easily discarded and it doesn't count towards the working memory set. Then they get surprised by unexpected OOMs. There is a decade-old issue on Kubernetes https://github.com/kubernetes/kubernetes/issues/43916
It was the default for every production binary. I read the c++ implementation, not really that crazy. It was for latency, and in some cases (with other measures) to keep the binary running if the disk failed.
The alternative (which I’ve also done), is to put the binary in tmpfs, after making sure it’s not bloated.
Your Go binary was being swapped out due to memory pressure, causing latency, and using mlock was your fix for this problem?
Or did someone else do this work and you assumed it was effective? Or did you personally implement this change and then validate that the before/after metrics improved as a result?
As I mentioned, it had long been the standard, the core library gave you a simple flag to use. I would occasionally find cases where it was not set correctly and you could see it causing short spikes (the simplest way to measure is something like a pacemaker task on an event loop, or task scheduling delay since it strips out other variables).
Alright, so you didn't implement it. And you're describing a "pacemaker task on an event loop"? That doesn't sound like a Go program at all, where we don't have a single event loop to measure and we have built-in runtime metrics.
Just seems kinda weird to give advice on forums when you're not speaking from actual relevant experience. I'm getting meat proxy answers and confident replies from people who've never done what they're suggesting others do.
As I said, I was more familiar with the C++ implementation, but we had it for every language in use.
I ultimately change the runtime environment to no longer allow any local durable storage for job (with a handful of exceptions), including swap and binary staging. The fact that almost everything was already using mlock made the transition simpler.
I am struggling to understand which specific problems that pertain to the use of mlock(2) you are alluding to.
Java is an example of the most widely deployed garbage-collected language in production around the world, deployment number and scale-wise. Most of the world's missing critical systems today run on Java. Java can run as a process or in a container. Containerised Java deployments do not require unique operational SRE arrangements, nor do they require any unique/dedicated permissions to run in a container.
Upon start-up, the JVM checks whether the system has sufficient memory to satisfy the max heap size constraint, and it will duly fail to start up if not. Afterwards, the JVM will allocate the required amount of memory, and it will mlock it. Multi-terabyte[0] RAM JVM deployments[1] do exist and exhibit exactly the same behaviour. Java has had such behaviour since its inception.
The prevailing number of Java deployments run on systems with swap enabled, and, from the operational standpoint, it is pure insanity to run mission critical (banking, payments, government etc) apps without a swap. Swapless JVM deployments are virtually unknown or are operational negligence.
The principal exceptions for running a JVM on a system without a swap are hard real-time scenarios and embedded systems. That is it.
What makes Go different or unique that it can't use mlock and that it would require special permissions or unique/distinct SRE/operational arrangements? It is eluding me.
[0] The Azul JVM supports up to 20 Tb of RAM which it will mlock.
[1] See the published Azul/Forrester evidence for reference. It is an interesting read regardless of the mlock debate.
I'm pretty sure almost all of our (very large) java systems were running without swap (it took effort to enable swap for your jobs, and eventually was almost impossible).
Java is generally simpler to run without swap since you know in advance how large the heap can be, so it's just a small amount of variability in the jvm internals to consider.
You don't want swap of heap/stack, or paged out code pages. The latency hit is pretty bad even for non-"real"-time systems. And really, you should know how much memory you are going to use (to within some fair bounds).
Adding only a small amount of swap for very cold pages does not save very much ram. And using it for anything else is very bad for interactive systems. In the same vein paging out code pages you won't ever run just does not save very much.
It requires some discipline in monitoring, load testing, etc, but you should be doing that anyway.
The nasty bit is that swap doesn't just make the allocation slower; if GC metadata gets paged out, you have turned memory pressure into a stop-the-world latency spike.
Immediate reclamation, yes. There do exist a number of reference-allocation algorithms that performs deferred reclamation similar to how tracing GC does.
ref counting is expensive in multi-threaded applications. Overall it would have worse performance. When it comes to predictability: deallocating a linked list (for instance) would have to deallocate all of the elements. Dealing with reference cycles is also not simple, either.
Linked-list freeing need not all happen in one go. There can be a queue of free-but-not-zeroed, push RC=0 stuff there instead of recurring on referents, anyone doing memory allocation work can pop a few items off that queue and decrement the counts of their referents, etc.
That loses timeliness, adds problem of queue management.
A weird trick I stumbled across years ago is to not strictly pop from the head of the stack/queue -- instead choose randomly from the last N elements of push end of the stack. This seemed to blunt the sort of growth that you get from "visiting wrong". I never did figure out the theory of why this worked.
Because the industry has plenty of experience with referece counting as the very first GC algorithm, already in the early 1960's, in early Lisp implementations, BASIC, Cedar, and several other languages.
The predictable runtime performance is also a myth, because they never take into account the use of NUMA memory, lock contention, possible stack overflow and stop the world in the case of cascaded deletions in naive implementations.
Places that were doing C++, but found out that the right GC, and JIT compiler, can achieve good enough performance for their business case.
There is nothing "special GC" about it, the failure is to assume there is only one way to implement GC algorithms, as if there is only one way to implement hash tables, tree re-balancing algorithms, .... and then place all languages into the same bucket.
Then we have the modern times with AI driven code generation, where no one cares how their agents are actually doing the work, with what kinds of resource management approaches.
Can you elaborate on how cascaded deletions "stop the world" with reference counting? I understand how GC could lead to arbitrarily large latencies for whatever task triggers that kind of cascaded deletions, but as I understand it, "stop the world" usually means that no threads are allowed to execute application code.
Imagine a graph or tree data structure where the deletion of a node causes a cascade deletion of all child nodes, which also causes a deletion of their children and so on.
This is proportional to the data structure being deleted.
Unless you use techniques to move the deletion into background threads, e.g. C++/WinRT with COM AddRef/Release, the thread will be "blocked" doing busy work cleaning all those nodes, running the cleanup code (destructors, deinit, whatever), node after node.
Yes, that's exactly what I said I understood. When Go (and Java) people say "stop the world" for GC, they mean all goroutines/threads stop, not just one. I don't think you get that with even naive reference counting.
Reference counting has high runtime overhead, especially atomic reference counting with multiple threads. If you overwrite a pointer, you'd need to also look up and update two counts and do all that in a manner that is effectively atomic -- and that is complex on current hardware. I've heard that Swift programs could have as much as 40% runtime overhead from ARC.
I still think that reference counting is promising though. First because it meshes well with static analysis memory-management techniques such as inference of uniqueness and borrowing -- that can optimise away RC altogether. (and Swift's compiler already does some of that). I have not seen any work that could optimise away tracing GC in a similar way.
Second, because I believe that it would be possible to design hardware with object-memory addressing that would performs atomic reference counting with no additional runtime cost.
Either the size of the data structures created is limited and known, or it isn't. Whether the work is done in a refcount-decrementing recursive descent or a tracing garbage collector doesn't change that part.
I don't understand why they felt necessary to rewrite a core component in another language. If you have a garbage collection problem my first intuition would be to produce less garbage!
For example better using the stack, or pulling out the big gun of manual memory management.
I'm sure they had reasons to choose Go when they first designed this project but they don't go into them at all.
Feels like they just wanted to play with a new toy.
One of big issue with GC is it need to scan every reachable objects to mark it is reachable. That mean even if your code produce no garbage but have a large amount of long lived object the GC still need to traverse all of those objects every cycle.
Problem with languages like Go is, because they have GC, their API is less tuned to the ability to prevent new allocations.
In C you see lots of functions whose first parameter is a buffer that will be re-used.
And even in Rust you don't see this often, because reusing a buffer means having to do &mut, which means having to re-structure your code. And it's really easy to do Vec::new().
Unfortunately that's somewhat expected - regardless of implementation, the sole fact that the GC needs to somehow walk the entire tree to trace still referenced sections makes swap highly impractical -- even if you had not had 40ms stop-the-world pauses the LRU cache used by swap would get thrown away at every GC.
I believe that's actually the same reason why Apple stopped using GC in their frameworks in favour of automatic reference counting.
Useful to them usually, or it goes to show how little our profession cares about learning the proper terms, or its historical background for that matter.
Anyway it doesn't matter any longer, AI is going to replace most of us, and it comes with automatic everything.
Whether reference counting is a GC algorithm depends on how you define what GC is.
I prefer to consider GC only the methods of memory management where reclaiming the no longer used memory is done either asynchronously with the main program or as late as possible, i.e. when new allocation requests cannot be satisfied.
In the normal implementation of reference counting, memory is freed as soon as possible, i.e. exactly like stack memory, when blocks are exited, so I do not consider reference counting as GC.
The problem with GC in the strict sense is that you cannot predict when it will happen. With both stack memory and reference counted heap memory you know that whenever you exit a block, some time will be spent with running destructors and for freeing memory, but such interruptions will not happen in other points of the program.
I define the academic view of GC algorithms in Computer Science, not what random developers decide to call GC.
Which in an industry where some folks call themselves Software Engineers after a bootcamp, without any kind of accreditation, I rather stay with the definition from those that do language design and compiler algorithms research.
I prefer to use the terms as they were originally defined by the authors who introduced them, and not with the modified meanings that become fashionable after the authors of some paper written some decades later decide randomly to change the definitions of the old terms.
So as fellow digital archaeologist I would find interesting to find Lisp, BASIC, CLU, SIMULA, or Cedar papers, where the authors mention reference counting not being a form of garbage collection algorithms.
Cedar was already combining reference counting with a cycle collector, as one of the very first systems programming languages with automatic resource management.
In the normal implementation of reference counts, counters can be decremented only at block exits and not at any other program point.
At a block exit some of the local variables that are freed may contain references, so freeing them will decrement some reference counts. Then some counters will reach zero, triggering other deallocations and the decrementing of other counters. This will repeat until no other counters reach zero.
All the memory deallocation happens predictably, only at block exits.
If a variable is not freed immediately when a counter reaches zero, but the deallocation is deferred for a later time, which is not predictable, that is no longer classic memory management with reference counts, but it is a garbage collector, which happens to also use reference counts, probably in combination with some tracing algorithm.
When reference counts are implemented, manual memory deallocation, like with C free() or C++ delete, should be forbidden, but even if it were used that would just introduce other program points besides the block exits, where it is known that memory deallocation will happen.
> Unfortunately that's somewhat expected - regardless of implementation
Yes, that's expected and no the GC implementation may be improved.
See this:
>>
I remember a research paper about swap and GC, where the GC cooperated with the OS to avoid this kind of issue. AFAIK it went nowhere, too bad.
perma-lienembedenregistrermodifierdésactiver les réponses de la messagerieeffacerrépondre
[–]andreiross[S] 15 points il y a 21 jours
Are you talking about this one? https://cse.buffalo.edu/\~mhertz/bc-pldi-2005.pdf. If so, yes. Too bad. I don't know the repercusions this paper had in the past, though, in the sense of pros and cons of the bookmark collector. Don't know if anyone tried to actually implement it or design it at some point.
Modern OSes need a facility to signal to a thread that it hit a paged out chunk of memory. It's not like it's not possible to create a swap-aware GC (or any other kind of code that touches lots of memory).
I would go further - I think the whole swap lifecycle needs to be communicated.
Before the OS swaps out a page, if it could invoke the GC which would clean up that piece of memory so that we dont end up writing garbage to swap.
The OS should also allow marking pages as piority to stop them from being swapped out.
> The OS should also allow marking pages as piority to stop them from being swapped out.
This is IIUC possible using the mlock(2) family of syscalls: https://man7.org/linux/man-pages/man2/mlock.2.html. (On Linux, though I'm guessing other UNIXen and operating systems have it or something equivalent.)
Apple should just stop using a silly mark & sweep collector. All the good collectors are copying, if you have enough memory. Only on tiny devices you need mark & sweep.
I started wondering if there could be swap-aware GC, like first make the required page swapped in (not that there's any obvious API for that...) and only then pause the world?
This. Memory bloat bugs are relatively easy to fix in Go, but sometimes it’s an adventure to remove swap bloat. And I think it’s worth removing all swap bloat!
When my system starts using swap space, I've already given up all hopes of a decently running system, so IMHO this particular "bug" doesn't matter much.
> I'm being downvoted, but honestly, is anyone always testing their software while swap is active?
Yes, not only is my system using swap 100% of the time, but I am working on and using an application (Rust) which often is used during a high demand of swap and expected to run reliably.
Choosing a language that forgives bad memory management practices and promotes sloppy programming practices is your decision. Don't be surprised that it has to do its job. You could have chosen modern C++ or Rust and use RAII and proper automatic deallocation.
160 comments
[ 0.24 ms ] story [ 122 ms ] thread"Stop doing that"
If you care about latency, disable swap. System wide or for the specific the cgroup.
If you care about latency, mlock() your memory, do not disable swap. Swap is good and gives the kernel an equal opportunity to evict data and code pages.
I'd rather have applications be oom_killed than having them swap out, the former is rather obvious and demands action.
Disabling swap will just moves pressere elsewhere: to code pages. And evicted code page is no better: full stall while kernel loads that page from disk.
Under memory pressure, the kernel evicts less popular pages from memory. If a page has been mapped from a file, it is dropped (if dirty, then it is written out first). If it is needed later, the kernel can read it back from the file. If a page is anonymous (read: heap page), then there is no backing file and the kernel copies it to swap before dropping it. This is swapping.
So, what happens if you disable swap and the kernel is low on memory? What can it evict? Anonymous pages cannot be evicted: there is no swap to put a copy in. The only choice the kernel has is to evict pages that are mapped from files. Those include pages mapped from the executable. You don't eliminate stalls by disabling swap, you just move them elsewhere: the kernel will page out code and your app gets paused whenever the execution flow hits such a page.
Your userspace early OOM killer triggers and lets you know you're trying to run more than will fit in memory so you don't do it again. (In my experience, the kernel OOM killer can't be trusted to kill processes soon enough.)
I haven't had to analyze the performance of no-swap processes before. My assumption is that code is hot enough to avoid eviction and that evicted code pages are rather the exception. To strong-man the argument, I can imagine long running complex (bloated) services could have parts that are not touched unless a specific request comes in.
In a scenario like the one describe, one is not running out of memory because of file-based pages. That's <1% of memory usage in typical cases. One is almost always running out of memory because of allocated anonymous pages.
This is not true, or the system wouldn't grind to a halt due to having to reload code pages from disk over and over.
But frankly, as a user I don't really care what the real technical reason is. I just don't think the system should become unresponsive for minutes on end during memory pressure. Just OOM-kill earlier, when it's supposed to.
The proper explanation is relatively straightforward. Normally, any memory request for a process is served by getting a page from the free memory pool. But when the system free memory goes below the low watermark (calculated from vm.min_free_kbytes and vm.watermark_scale_factor), the kernel stops giving the pages from the free memory pool and instead goes into direct reclaim, trying to find unused pages. While the kernel does reclaim, the allocating process remains stalled. The kernel can't return control to the process - there is no page mapped that the process tried to access. This applies to any types of pages - anon, file-backed, code pages. So, under the condition of free memory below the low watermark, any process trying to allocate memory gets stalled. This is measured as memory pressure stalls (PSI) in /proc/pressure/memory.
The problem with the system being stuck in an unresponsive state is that the design choice for the kernel OOM handling is to be optimistic and keep the workload running as long as the kernel is making any progress reclaiming pages. The direct reclaim process needs to fail to make any progress 16 times before the OOM killer is invoked.
The solution is to use cgroup memory limits to avoid getting into the situation of memory exhaustion and use user-space oomd to recover from this situation earlier.
Any repeatable way to prove your thesis or disprove the trashing-code-pages thesis?
You can also track variables from /proc/vmstat exported as node_vmstat_allocstall_(normal|dma32|movable), node_vmstat_pgscan_direct, node_vmstat_pgsteal_direct - you'll see them spike when the system enters direct reclaim.
Install and configure systemd-oomd and track its log - it will activate on memory pressure.
To track code page eviction/re-faults specifically, you'll need to write some eBPF code, there are no direct measurements for this available.
Code pages are paged in on demand: for any sufficiently large binary just some code pages will be mapped right after start. Given enough memory pressure, the active LRU will be shrunk often enough to push code pages out of the active LRU and then they are eventually evicted.
I'm not aware how it works in modern MGLRU, but I would assume similar ovservable behevior holds.
When the kernel needs memory, it goes hunting for a page it can discard. But since that's transparent, the kernel can only discard a page if it knows it can get it back (after all, it's still got valid data on it).
If there's swap, a page full of stale/unneeded data can be written out to swap. But if there's no swap, your page of "dangling data that you'll never use, but is still valid & referenced" can't be discarded; the kernel doesn't know you won't want it later, and it can't recreate the page if it throws it away.
So like sibling said, at that point it has to find other pages it can evict from memory, ones that _do_ have somewhere persistent they can be written out to. Pages loaded from binaries on disk satisfy that, so those will get dropped instead.
The only code in danger is the one from libraries (e.g. zlib)
You can tell an average Golang user to use mutexes, you shouldn't be telling them to lock memory manually.
If you're in a position where you still want swap to be there, but only used in extreme situations, then you should set it yourself. Otherwise disable swap or switch a language.
Some arbitrary number.
https://docs.kernel.org/admin-guide/sysctl/vm.html#swappines...
> At 0, the kernel will not initiate swap until the amount of free and file-backed pages is less than the high watermark in a zone.
I don't see a reason why the default shouldn't be 0 then on modern systems.
> I don't see a reason why the default shouldn't be 0 then on modern systems.
The zero value makes reclaim very unbalanced and is not appropriate for generic systems. This will cause the kernel to discard all the page cache, slowing down the system with uncached I/O, while keeping all unused anonymous pages in memory.
> If you care about latency, disable swap.
So that only leaves anonymous data pages, i.e., regular data in memory, that could be swapped. Ok but you're not easing the load on code pages, those still get evicted during memory pressure whether you have swap or not.
His reply below was presumably killed by mods for doing this: https://news.ycombinator.com/item?id=49970841
Yes, I was and still running go and zig binaries with mlockall. Apps have large in-application caches and also perform large streaming reads of datasets bigger than available RAM. If it happens that several streaming reads running at same time, they push memory pressure hard enough, so kernel swaps out cold parts of in-app caches. That later leads to elevated latencies => mlockall to avoid that.
This is not exclusive to GC, any program that reads a rarely accessed piece of memory is vulnerable to this.
The advice of using mlock is just wrong in the case of a Go server allocating memory for its heap. Don't try to do that. Have sufficient memory, monitor it, and let the Go GC and kernel manage memory normally.
You are correct, that 90MB Go binary is not major memory consumer. But it doesn't matter what is the source of memory pressure. What matters is what kernel do under pressure. If you disable swap, then under memory pressure kernel can only pageout memory mapped files, no matter how small they are. And your code will be paged out.
Introducing mlock is operationally annoying and trying to solve the wrong problem. It requires elevated container permissions, it's not something your SRE team is going to expect, almost no one does it, and it has all kinds of technical implications and interactions with kernel OOM system, forking, etc.
There are very specific scenarios where mlock is exactly the right solution but the typical Go HTTP service is not one of them. The KISS principle applies.
Have you actually done this yourself for Go programs that you've deployed into production? If so, what were the circumstances?
The alternative (which I’ve also done), is to put the binary in tmpfs, after making sure it’s not bloated.
Or did someone else do this work and you assumed it was effective? Or did you personally implement this change and then validate that the before/after metrics improved as a result?
The impact is pretty obvious.
Just seems kinda weird to give advice on forums when you're not speaking from actual relevant experience. I'm getting meat proxy answers and confident replies from people who've never done what they're suggesting others do.
I ultimately change the runtime environment to no longer allow any local durable storage for job (with a handful of exceptions), including swap and binary staging. The fact that almost everything was already using mlock made the transition simpler.
Java is an example of the most widely deployed garbage-collected language in production around the world, deployment number and scale-wise. Most of the world's missing critical systems today run on Java. Java can run as a process or in a container. Containerised Java deployments do not require unique operational SRE arrangements, nor do they require any unique/dedicated permissions to run in a container.
Upon start-up, the JVM checks whether the system has sufficient memory to satisfy the max heap size constraint, and it will duly fail to start up if not. Afterwards, the JVM will allocate the required amount of memory, and it will mlock it. Multi-terabyte[0] RAM JVM deployments[1] do exist and exhibit exactly the same behaviour. Java has had such behaviour since its inception.
The prevailing number of Java deployments run on systems with swap enabled, and, from the operational standpoint, it is pure insanity to run mission critical (banking, payments, government etc) apps without a swap. Swapless JVM deployments are virtually unknown or are operational negligence.
The principal exceptions for running a JVM on a system without a swap are hard real-time scenarios and embedded systems. That is it.
What makes Go different or unique that it can't use mlock and that it would require special permissions or unique/distinct SRE/operational arrangements? It is eluding me.
[0] The Azul JVM supports up to 20 Tb of RAM which it will mlock.
[1] See the published Azul/Forrester evidence for reference. It is an interesting read regardless of the mlock debate.
Java is generally simpler to run without swap since you know in advance how large the heap can be, so it's just a small amount of variability in the jvm internals to consider.
You don't want swap of heap/stack, or paged out code pages. The latency hit is pretty bad even for non-"real"-time systems. And really, you should know how much memory you are going to use (to within some fair bounds).
Adding only a small amount of swap for very cold pages does not save very much ram. And using it for anything else is very bad for interactive systems. In the same vein paging out code pages you won't ever run just does not save very much.
It requires some discipline in monitoring, load testing, etc, but you should be doing that anyway.
(though of course a swap is a swap - but you can "trigger" it depending on your memory or file access pattern)
Not really, reference counting can cause a single object deallocation to trigger an arbitrarily long chain of deallocations.
Variable time yes but you know when you're going to pay it
(I mean yes you can put your gc.run() there as well, but it might not give you the results you want)
That loses timeliness, adds problem of queue management.
A weird trick I stumbled across years ago is to not strictly pop from the head of the stack/queue -- instead choose randomly from the last N elements of push end of the stack. This seemed to blunt the sort of growth that you get from "visiting wrong". I never did figure out the theory of why this worked.
The predictable runtime performance is also a myth, because they never take into account the use of NUMA memory, lock contention, possible stack overflow and stop the world in the case of cascaded deletions in naive implementations.
I'm not sure most GC implementations worry about the rest neither (as by the several complaints we see going around)
(makes me wonder who's buying those - things like Azul, etc)
There is nothing "special GC" about it, the failure is to assume there is only one way to implement GC algorithms, as if there is only one way to implement hash tables, tree re-balancing algorithms, .... and then place all languages into the same bucket.
Then we have the modern times with AI driven code generation, where no one cares how their agents are actually doing the work, with what kinds of resource management approaches.
This is proportional to the data structure being deleted.
Unless you use techniques to move the deletion into background threads, e.g. C++/WinRT with COM AddRef/Release, the thread will be "blocked" doing busy work cleaning all those nodes, running the cleanup code (destructors, deinit, whatever), node after node.
I still think that reference counting is promising though. First because it meshes well with static analysis memory-management techniques such as inference of uniqueness and borrowing -- that can optimise away RC altogether. (and Swift's compiler already does some of that). I have not seen any work that could optimise away tracing GC in a similar way. Second, because I believe that it would be possible to design hardware with object-memory addressing that would performs atomic reference counting with no additional runtime cost.
Such hardware has been designed in the past, Lisp machines, Ada machines, the famous iAPX 432 Intel's failure.
if the arch is not the last: atomic check
> You lose nothing when going out of scope with GC.
actually, you don't know
> actually, you don't know
I know because I’ve verified it.
* subset of all allocations until it runs
I use Rust where i need low latency.
For example better using the stack, or pulling out the big gun of manual memory management.
I'm sure they had reasons to choose Go when they first designed this project but they don't go into them at all.
Feels like they just wanted to play with a new toy.
In C you see lots of functions whose first parameter is a buffer that will be re-used.
And even in Rust you don't see this often, because reusing a buffer means having to do &mut, which means having to re-structure your code. And it's really easy to do Vec::new().
I believe that's actually the same reason why Apple stopped using GC in their frameworks in favour of automatic reference counting.
Many make the mistake to think there is only one way to do a GC.
That is a tracing GC by the way.
There are also tracing GC implementations with deterministic resource management APIs, .NET and D have them for example.
Anyway it doesn't matter any longer, AI is going to replace most of us, and it comes with automatic everything.
Garbage collection of the type: usually there's no garbage, but when there is, it's not collected.
I prefer to consider GC only the methods of memory management where reclaiming the no longer used memory is done either asynchronously with the main program or as late as possible, i.e. when new allocation requests cannot be satisfied.
In the normal implementation of reference counting, memory is freed as soon as possible, i.e. exactly like stack memory, when blocks are exited, so I do not consider reference counting as GC.
The problem with GC in the strict sense is that you cannot predict when it will happen. With both stack memory and reference counted heap memory you know that whenever you exit a block, some time will be spent with running destructors and for freeing memory, but such interruptions will not happen in other points of the program.
Which in an industry where some folks call themselves Software Engineers after a bootcamp, without any kind of accreditation, I rather stay with the definition from those that do language design and compiler algorithms research.
Cedar was already combining reference counting with a cycle collector, as one of the very first systems programming languages with automatic resource management.
It’s the same with ARC. You also don’t know when the counter will reach zero.
In the normal implementation of reference counts, counters can be decremented only at block exits and not at any other program point.
At a block exit some of the local variables that are freed may contain references, so freeing them will decrement some reference counts. Then some counters will reach zero, triggering other deallocations and the decrementing of other counters. This will repeat until no other counters reach zero.
All the memory deallocation happens predictably, only at block exits.
If a variable is not freed immediately when a counter reaches zero, but the deallocation is deferred for a later time, which is not predictable, that is no longer classic memory management with reference counts, but it is a garbage collector, which happens to also use reference counts, probably in combination with some tracing algorithm.
When reference counts are implemented, manual memory deallocation, like with C free() or C++ delete, should be forbidden, but even if it were used that would just introduce other program points besides the block exits, where it is known that memory deallocation will happen.
Yes, that's expected and no the GC implementation may be improved. See this: >> I remember a research paper about swap and GC, where the GC cooperated with the OS to avoid this kind of issue. AFAIK it went nowhere, too bad.
perma-lienembedenregistrermodifierdésactiver les réponses de la messagerieeffacerrépondre
[–]andreiross[S] 15 points il y a 21 jours
Are you talking about this one? https://cse.buffalo.edu/\~mhertz/bc-pldi-2005.pdf. If so, yes. Too bad. I don't know the repercusions this paper had in the past, though, in the sense of pros and cons of the bookmark collector. Don't know if anyone tried to actually implement it or design it at some point.
perma-lienembedenregistrerparentsignalerrépondre
[–]renozyx 9 points il y a 21 jours
Yes, congratulations for finding it. And I don't know either,. Except that they did implement it on Linux (of course) https://plasma.cs.umass.edu/emery/cooperative-memory-managem...
<<
I would go further - I think the whole swap lifecycle needs to be communicated. Before the OS swaps out a page, if it could invoke the GC which would clean up that piece of memory so that we dont end up writing garbage to swap.
The OS should also allow marking pages as piority to stop them from being swapped out.
This is IIUC possible using the mlock(2) family of syscalls: https://man7.org/linux/man-pages/man2/mlock.2.html. (On Linux, though I'm guessing other UNIXen and operating systems have it or something equivalent.)
I'd guess it ain't gonna perform great either, though.
Any program that has a rarely accessed, but vital chunk of memory is vulnerable to this sort of issue.
All the better systems use that for decades.
Yes, not only is my system using swap 100% of the time, but I am working on and using an application (Rust) which often is used during a high demand of swap and expected to run reliably.