> We plan to deliver improvements to [..] purging mechanisms
During my time at Facebook, I maintained a bunch of kernel patches to improve jemalloc purging mechanisms. It wasn't popular in the kernel or the security community, but it was more efficient on benchmarks for sure.
Many programs run multiple threads, allocate in one and free in the other. Jemalloc's primary mechanism used to be: madvise the page back to the kernel and then have it allocate it in another thread's pool.
One problem: this involves zero'ing memory, which has an impact on cache locality and over all app performance. It's completely unnecessary if the page is being recirculated within the same security domain.
The problem was getting everyone to agree on what that security domain is, even if the mechanism was opt-in.
I'm really surprised to see you still hocking this.
We did extensive benchmarking of HHVM with and without your patches, and they were proven to make no statistically significant difference in high level metrics. So we dropped them out of the kernel, and they never went back in.
I don't doubt for a second you can come up with specific counterexamples and microbenchnarks which show benefit. But you were unable to show an advantage at the system level when challenged on it, and that's what matters.
You probably weren't there when servers were running for many days at a time.
By the time you joined and benchmarked these systems, the continuous rolling deployment had taken over. If you're restarting the server every few hours, of course the memory fragmentation isn't much of an issue.
> But you were unable to show an advantage at the system level when challenged on it, and that's what matters.
You mean 5 years after I stopped working on the kernel and the underlying system had changed?
The patches were written in 2011 and published in 2012. They did what they were supposed to at the time.
For the peanut gallery: this is a manifestation of an internal eng culture at fb that I wasn't particularly fond of. Celebrating that "I killed X" and partying about it.
You didn't reply to the main point: did you benchmark a server that was running several days at a time? Reasonable people can disagree about whether this a good deployment strategy or not. I tend to believe that there are many places which want to deploy servers and run for months if not days.
For the peanut gallery more: I worked with both of these guys at Meta on this.
The "servers are only on for a few hours" thing was like never true so I have no idea where that claim is coming from. The web performance test took more than a few hours to run alone and we had way more aggressive soaks for other workloads.
My recollection was that "write zeroes" just became a cheaper operation between '12 and '14.
A fun fact to distract from the awkwardness: a lot of the kernel work done in the early days was exceedingly scrappy. The port mapping stuff for memcached UDP before SO_REUSEPORT for example. FB binaries couldn't even run on vanilla linux a lot of the time. Over the next several years we put a TON of effort in getting as close to mainline as possible and now Meta is one of the biggest drivers of Linux development.
I wouldn't be surprised if both 'adsharma' and 'jcalvinowens' were right, just at different points in time, perhaps in a bit different context. Things change.
It wasn't any cgroup. If you put two untrusting processes in a memory cgroup, there is a lot that can go wrong.
If you don't like the idea of memory cgroups as a security domain, you could tighten it to be a process. But kernel developers have been opposed to tracking pages on a per address space basis for a long time. On the other hand memory cgroup tracking happens by construction.
Note the complementary language usage here. You seem to have interpreted that as me writing that it didn't matter what cgroup they are in, which is an odd thing to claim that I implied. I meant within the same cgroup obviously.
Yes, you can read memory out of another process through other means.. but you shouldn't map pages, be able to read them and see what happened in another process. That's the wild part. It strikes me as asking for problems.
I was unaware of MAP_UNINITIALIZED, support for which was disabled by default and for good reason. Seems like it was since removed.
I was clarifying that there are CPU cgroups, network cgroups etc and the proposal touched only memory cgroups.
The people deploying it are free to restrict the cgroup to one process before requesting MAP_UNINITIALIZED if there is a concern around security. At that point the memory cgroup becomes a way to get around the page tracking restriction.
But I get why aesthetically this idea sounds icky to a lot of people.
Some more historical context. It wasn't a random optimization idea that I thought about in the shower and implemented the next day. Previous work on company wide profiling, where my contribution was low level perf_events plumbing:
The profiling clearly showed kernel functions doing memzero at the top of the profiles which motivated the change. The performance impact (A/B testing and measuring the throughput) also showed a benefit at the point the change was committed.
The change stopped being impactful sometime after 2013, when a JIT replaced the transpiler. I'm guessing likely before 2016 when continuous deployment came into play. But that was continuously deploying PHP code, not HHVM itself.
By the time the patches were reevaluated I was working on a Graph Database, which sounded a lot more interesting than going back to my old job function and defending a patch that may or may not be relevant.
I'm still working on one. Guilty as charged of carrying ideas in my head for 10+ years and acting on them later. Link in my profile.
This kind of thing always struck me as something that the MMU and the memory controller could team up on. When you give back memory, you could not refresh it for some cycles. Or you could DMA the same page of zeros over all of it, so the CPU isn't involved in menial labor.
I recently started using Microsoft's mimalloc (via an LD_PRELOAD) to better use huge (1 GB) pages in a memory intensive program. The performance gains are significant (around 20%). It feels rather strange using an open source MS library for performance on my Linux system.
There needs to be more competition in the malloc space. Between various huge page sizes and transparent huge pages, there are a lot of gains to be had over what you get from a default GNU libc.
We evaluated a few allocators for some of our Linux apps and found (modern) tcmalloc to consistently win in time and space. Our applications are primarily written in Rust and the allocators were linked in statically (except for glibc). Unfortunately I didn't capture much context on the allocation patterns. I think in general the apps allocate and deallocate at a higher rate than most Rust apps (or more than I'd like at least).
Our results from July 2025:
rows are <allocator>: <RSS>, <time spent for allocator operations>
app1:
glibc: 215,580 KB, 133 ms
mimalloc 2.1.7: 144,092 KB, 91 ms
mimalloc 2.2.4: 173,240 KB, 280 ms
tcmalloc: 138,496 KB, 96 ms
jemalloc: 147,408 KB, 92 ms
app2, bench1
glibc: 1,165,000 KB, 1.4 s
mimalloc 2.1.7: 1,072,000 KB, 5.1 s
mimalloc 2.2.4:
tcmalloc: 1,023,000 KB, 530 ms
app2, bench2
glibc: 1,190,224 KB, 1.5 s
mimalloc 2.1.7: 1,128,328 KB, 5.3 s
mimalloc 2.2.4: 1,657,600 KB, 3.7 s
tcmalloc: 1,045,968 KB, 640 ms
jemalloc: 1,210,000 KB, 1.1 s
app3
glibc: 284,616 KB, 440 ms
mimalloc 2.1.7: 246,216 KB, 250 ms
mimalloc 2.2.4: 325,184 KB, 290 ms
tcmalloc: 178,688 KB, 200 ms
jemalloc: 264,688 KB, 230 ms
tcmalloc was from github.com/google/tcmalloc/tree/24b3f29.
I’m surprised (unless they replaced the core tcmalloc algorithm but kept the name).
tcmalloc (thread caching malloc) assumes memory allocations have good thread locality. This is often a double win (less false sharing of cache lines, and most allocations hit thread-local data structures in the allocator).
Multithreaded async systems destroy that locality, so it constantly has to run through the exception case: A allocated a buffer, went async, the request wakes up on thread B, which frees the buffer, and has to synchronize with A to give it back.
modern tcmalloc uses per CPU caches via rseq [0]. We use async rust with multithreaded tokio executors (sometimes multiple in the same application). so relatively high thread counts.
on the OS scheduler side, i'd imagine there's some stickiness that keeps tasks from jumping wildly between cores. like i'd expect migration to be modelled as a non zero cost. complete speculation though.
tokio scheduler side, the executor is thread per core and work stealing of in progress tasks shouldn't be happening too much.
for all thread pool threads or threads unaffiliated with the executor, see earlier speculation on OS scheduler behavior.
1. tcmalloc is actually the only allocator I tested which was not using thread local caches. even glibc malloc has tcache.
2. async executors typically shouldn’t have tasks jumping willy nilly between threads. i see the issue u describe more often with the use of thread pools (like rayon or tokio’s spawn_blocking). i’d argue that the use of thread pools isn’t necessarily an inherent feature of async executors. certainly tokio relies on its threadpool for fs operations, but io-uring (for example) makes that mostly unnecessary.
This is similar to what I experienced when I tested mimalloc many years ago. If it was faster, it wasn't faster by much, and had pretty bad worst cases.
If you go into Dr Dobbs, The C/C++ User's Journal and BYTE digital archives, there will be ads of companies whose product was basically special cased memory allocator.
Even toolchains like Turbo Pascal for MS-DOS, had an API to customise the memory allocator.
One of the best parts about GC languages is they tend to have much more efficient allocation/freeing because the cost is much more lumped together so it shows up better in a profile.
Agreed, however there is also a reason why the best ones also pack multiple GC algorithms, like in Java and .NET, because one approach doesn't fit all workloads.
Minecraft for somewhat silly reasons was largely stuck using Java8 for ~a decade longer than it should have which meant that it was using some fairly outdated GC algorithms.
Any extra throughput is far overshadowed by trying to control pauses and too much heap allocations happening because too much gets put on the heap. For anything interactive the options are usually fighting the gc or avoiding gc.
When it works. Many programs in GC language end up fighting the GC by allocating a large buffer and managing it by hand anyway because when performance counts you can't have allocation time in there at all. (you see this in C all the time as well)
That's generally a bad idea. Not always, but generally.
It was a better idea when Java had the old mark and sweep collector. However, with the generational collectors (which are all Java collectors now. except for epsilon) it's more problematic. Reusing buffers and objects in those buffers will pretty much guarantees that buffer ends up in oldgen. That means to clear it out, the VM has to do more expensive collections.
The actual allocation time for most of Java's collectors is almost 0, it's a capacity check and a pointer bump in most circumstances. Giving the JVM more memory will generally solve issues with memory pressure and GC times. That's (generally) a better solution to performance problems vs doing the large buffer.
Now, that said, there certainly have been times where allocation pressure is a major problem and removing the allocation is the solution. In particular, I've found boxing to often be a major cause of performance problems.
If your workload is very regular, you can still do better with an arena allocator. Within the arena, it uses the same pointer-bump allocation as Java normally uses, but then you can free the whole area back to the start by resetting the pointer to its initial value. If you use the arena for servicing a single request, for instance, you then reset as soon as you're done with the request, setting you up with a totally free area for the next request. That's more efficient than a GC. But it also requires your algorithm to fall into that pattern where you KNOW that you can and should throw everything from the request away. If you can't guarantee that, then modern collectors are pretty magical and tunable.
For example, some code I had to clean up pretty early on in my career was a dev, for unknown reasons, reinventing the `ArrayList` and then using that invention as a set (doing deduplication by iterating over the elements and checking for duplicates). It was done in the name of performance, but it was never a slow part of the code. I replaced the whole thing with a `HashSet` and saved ~300 loc as a result.
This individual did that sort of stuff all over the code base.
I remember in the early days of web services, using the apache portable runtime, specifically memory pools.
If you got a web request, you could allocate a memory pool for it, then you would do all your memory allocations from that pool. And when your web request ended - either cleanly or with a hundred different kinds of errors, you could just free the entire pool.
it was nice and made an impression on me.
I think the lowly malloc probably has lots of interesting ways of growing and changing.
In many cases you can also do better than using malloc e.g. if you know you need a huge page, map a huge page directly with mmap
Yes, if you want to use huge pages with arbitrary alloc/free, then use a third-party malloc. If your alloc/free patterns are not arbitrary, you can do even better. We treat malloc as a magic black box but it's actually not very good.
I feel like the real thing that needs to change is we need a more expressive allocation interface than just malloc/realloc. I'm sure that memory allocators could do a significantly better job if they had more information about what the program was intending to do.
You can also play tricks with inlining and constant propagation in C (especially on the malloc path, where the ground-truth allocation size is usually statically known).
I think some operating system improvements could get people motivated to use huge pages a lot better. In particular make them less fragile on linux and make them not need admin rights on windows. The biggest factor causing problems there is that neither OS can swap a 2MB page. So someone needs to care enough to fix that.
Just out of curiosity are you getting 1GB huge pages on Xeon or some other platform? I always thought this class of page is the hardest to exploit, considering that the machine only has, if I recall correctly, one TLB slot for those.
Modern x86_64 has supported multiple page sizes for a long time. I'm on commodity Zen 5 hardware (9900X) with 128 GiB of RAM. Linux will still use a base page size of 4kb but also supports both 2 MiB and 1 GiB huge pages. You can pass something like `default_hugepagesz=2M hugepagesz=1G hugepages=16` to your kernel on boot to use 2 MiB pages but reserve 16 1 GiB pages for later use.
The nice thing about mimalloc is that there are a ton of configurable knobs available via env vars. I'm able to hand those 16 1 GiB pages to the program at launch via `MIMALLOC_RESERVE_HUGE_OS_PAGES=16`.
EDIT: after re-reading your comment a few times, I apologize if you already knew this (which it sounds like you did).
Right but on Intel the 1G page size has historically been the odd one. For example Skylake-X has 1536 L2 shared TLB entries for either 4K or 2M pages, but it only has 16 entries that can be used for 1G pages. It wasn't unified until Cascade Lake. But Skylake-like Xeon is still incredibly common in the cloud so it's hard to target the later ones.
So for any process that's using less than 16GB, it's a significant performance boost. And most processes using more RAM, but not splitting accesses across more than 16 zones in rapid succession, will also see a performance boost.
My old Intel CPU only has 4 slots for 1GB pages, and that was enough to get me about a 20% performance boost on Factorio. (I think a couple percent might have been allocator change but the boost from forcing huge pages was very significant)
If there is so much performance difference among generic allocators, it means you need semantic optimized allocators (unless performance is actually not that much important in the end).
Agreed mostly. Going from standard library to something like jemalloc or tcmalloc will give you around 5-10% wins which can be significant, but the difference between those generic allocators seem small. I just made a slab allocator recently for a custom data type and got speedups of 100% over malloc.
I've been using jemalloc for over 10 years and don't really see a need for it to be updated. It always holds up in benchmarks against any new flavor of the month malloc that comes out.
Last time I checked mimalloc which was admittedly a while ago, probably 5 years, it was noticebly worse and I saw a lot of people on their github issues agreeing with me so I just never looked at it again.
One has to wonder if this due to the global memory shortage. ("Oh - changing our memory allocator to be more efficient will yield $XXM dollar savings over the next year").
Facebook had talks already years ago (10+) - nobody was allowed to share real numbers, but several facebook employed where allowed to share that the company has measured savings from optimizations. Reading between the lines, a 0.1% efficiency improvement to some parts of Facebook would save them $100,000 a month (again real numbers were never publicly shared so there is a range - it can't be less than $20,000), and so they had teams of people whose job it was to find those improvements.
Most of the savings seemed to come from HVAC costs, followed by buying less computers and in turn less data centers. I'm sure these days saving memory is also a big deal but it doesn't seem to have been then.
The above was already the case 10 years ago, so LLMs are at most another factor added on.
I don't have many regrets about having spent my career in (relatively) tiny companies by comparison, but it sure does sound fun to be on the other side for this kind of thing - the scale where micro-optimizations have macro impact.
In startups I've put more effort into squeezing blood from a stone for far less change; even if the change was proportionally more significant to the business. Sometimes it would be neat to say "something I did saved $X million dollars or saved Y kWh of energy" or whatever.
I've worked on optimizing systems in that ballpark range, memory is worth saving but it isn't necessarily 1:1 with increasing revenue like CPU is. For CPU we have tables to calculate the infra cost savings (we're not really going to free up the server, more like the system is self balancing so it can run harder with the freed CPU), but for memory as long as we can load in whatever we want to (rec systems or ai models) we're in the clear so the marginal headroom isn't as important. It's more of a side thing that people optimizing CPU also get wins in by chance because the skillsets are similar.
There are some people who think they can beat the system by treating apps like Telegram and Discord as free cloud storage, and they certainly get banned to save storage space.
On top of cost, they probably cannot get as much memory as they order in a timely fashion so offsetting that with greater efficiency matters right now.
Yeah, identifying single-digit millions of savings out of profiles is relatively common practice at Meta. It's ~easy to come up with a big number when the impact is scaled across a very large numbers of servers. There is a culture of measuring and documenting these quantified wins.
Not just shortage, any improvements to LLMs/electricity/servers memory footprint is becoming much more valuable as the time goes. If we can get 10% faster, you can easily get a lead in the LLM race. The incentives to transparently improving performance are tremendous
As an Australian who was just made redundant from a role that involved this type of low level programming - I love working on these these kinds of challenges.
I'm saddened that the job market in Australia is largely React CRUD applications and that it's unlikely I will find a role that lets me leverage my niche skill set (which is also my hobby)
Speaking as an Australian that works on React CRUD applications because there's nothing else in the market, I've been reading through this thread thinking the exact same thing.
Google had some position open working on the kernel for ChromeOS, and Microsoft had some positions working on data center network drivers.
I applied for both and got ghosted, haha.
I also saw a government role as a security researcher. Involves reverse engineering, ghidra and that sort of thing. Super awesome - but the pay is extremely uncompetitive. Such a shame.
Other than that, the most interesting roles are in finance (like HFT) - where you need to juggle memory allocations, threads and use C++ (hoping I can pitch Rust but unlikely).
Sadly they have a reputation of having pretty rough cultures, uncompetitive salaries and it's all in-office
I'm actually looking at HFT companies. Hoping I find one that allows remote working - but looks like there are basically no remote roles going at the moment
I hear you. Actually I read this thread because we’re using jemalloc in an embedded product. The only way I found to work on interesting problems here was to work for myself. (Having said that I think Apple might have some security research in Canberra? Years ago there was LinuxCare there and a lot of smart people. But that was in 2003…)
I remember I was a senior lead softeng of a worldbank funded startup project, and have deployed Ruby with jemalloc in prod. There's a huge noticeable speed and memory efficiency. It did saved us a lot of AWS costs, compare to just using normal Ruby. This was 8 years ago, why haven't projects adopt it yet as de facto.
Usually lack of knowledge that such a thing exists, or just plain ol' momentum. Changing something long in production at established companies, even if there is a tangible benefit, can be a real challenge.
> Facebook's coding AIs to the rescue, maybe? I wonder how good all these "agentic" AIs are at dreaded refactoring jobs like these.
No.
This is something you shouldn't allow coding agents anywhere near, unless you have expert-level understanding required to maintain the project like the previous authors have done without an AI for years.
I've done some work in this sort of area before, though not literally on a malloc. Yes you very much want to be careful, but ultimately it's the tests that give you confidence. Pound the heck out of it in multithreaded contexts and test for consistency.
> ...but ultimately it's the tests that give you confidence. Pound the heck out of it in multithreaded contexts and test for consistency.
I don't think so.
Even on LLM generated code, it is still not enough and you cannot trust it. They can pass the tests and still cause a regression and the code will look seemingly correct, for example in this case study [0].
AI is more than happy to declare the test wrong and “fix it” if you’re not careful. And the cherry on top is that sometimes the test could be wrong or need updating due to changed behavior. So…
If you filter the commits to the past five years, four of the top six committers are Meta employees. The other two might be as well, it just doesn't say that on their Github / personal website.
> With the leverage jemalloc provides however, it can be tempting to realize some short-term benefit. It requires strong self-discipline as an organization to resist that temptation and adhere to the core engineering principles.
This doesn't quite read properly to me. What does it actually mean, does anyone know?
I'm pretty sure it means something like this: "Because jemalloc is used all over the place in our systems that run at tremendous scale, some hack that improves its performance a little bit while degrading the longer-term maintainability of the code can look very appealing -- look, doing this thing will save us $X,000,000 per year! -- and it takes discipline to avoid giving in to that temptation and to insist on doing things properly even if sometimes it means passing up a chance to make the code 0.1% faster and 10% messier."
It would be great if Meta was able to sustain to support more open source projects, especially those they benefit from.
For example they use AsmJit in a lot of projects (both internal and open-source) and it's now unmaintained because of funding issues. Maybe they have now internal forks too.
Surprised not to see any mention of the global memory supply shock. Would love to learn more about how that economic is shifting software priorities toward memory allocation for the first time in my (relatively young) career
While it may seem directly related, it's just not. These things are worked on regardless of how cheap or expensive RAM is, because optimizing memory footprint pretty much always leads to fewer machines leased, which is a worthwhile goal even for smaller shops.
Meta never abandoned jemalloc. https://github.com/facebook/jemalloc remained public the entire time. It's my understanding that Jason Evans, the creator of jemalloc, had ownership over the jemalloc/jemalloc repo which is why that one stopped being updated after he left.
Meta still maintained it and actively pushed commits to it fixing bugs and adding improvements. From this blog post it sounds like they are increasing investment into it along with resurrecting the original repo. When the repo was archived Meta said that development on jemalloc would be focused towards Meta's own goals and needs as opposed to the larger ecosystem.
This looks a lot as if the facebook/jemalloc repo inserted a single commit 70 commits ago and then rebased the changes in the original repo on top. Because the commit SHAs for the changes pulled in change you see this result.
Few months back, some of the services switched to jemalloc for the Java VM. It took months (of memory dumps and tracing sys-calls) to blame the JVM, itself, for getting killed by the oom_killer.
Initially the idea was diagnostics, instead the the problem disappeared on its own.
First impressions: LOL, the blunt commentary in the HN thread title compared to the PR-speak of the fb.com post.
Second thoughts: Actually the fb.com post is more transparent than I'd have predicted. Not bad at all. Of course it helps that they're delivering good news!
Large engineering orgs often underestimate how much CI pipelines amplify performance issues. Even small inefficiencies multiply when builds run hundreds of times a day.
Allocators like that aren't the default for every process because they have higher startup costs. They are targeted to server workloads where startup cost doesn't matter, but it matters a lot if you're doing crud like starting millions of short-lived processes.
I was recently debugging an app double-free segfault on my android 13 samsung galaxy A51 phone, and the internal stack trace pointed to jemalloc function calls (je_free).
scudo has been the default allocator for Android since Android 11, and we are hoping to make it mandatory for the few remaining places that don't use it. Using an allocator without memory protections in 2026 (especially after we have closed nearly all known performance gaps with jemalloc) is really not a great choice.
The only option for cookies is to accept these terms and conditions, I thought implied consent was explicitly not allowed due to GDPR?
"To help personalize content, tailor and measure ads and provide a safer experience, we use cookies. By clicking or navigating the site, you agree to allow our collection of information on and off Facebook through cookies. Learn more, including about available controls:
> Building a software system is a lot like building a skyscraper: The product everyone sees is the top, but the part that keeps it from falling over is the foundation buried in the dirt and the scaffolding hidden from sight.
They should have just called it an ivory tower, as that's what they're building whenever they're not busy destroying democracy with OS Backdoor lobbyism or Cambridge Analytica shenanigans.
Edit: If every thread about any of Elon Musk's companies can contain at least 10 comments talking about Elon's purported crimes against humanity, threads about Zuckerberg's companies can contain at least 1 comment. Without reminders like this, stories like last week's might as well remain non-consequential.
That's a false dichotomy: you optimize both the application and the allocator.
A 0.5% improvement may not be a lot to you, but at hyperscaler scale it's well worth staffing a team to work on it, with the added benefit of having people on hand that can investigate subtle bugs and pathological perf behaviors.
exactly. I can think of at least 5 different projects I have been on where a better allocator would made a world of difference. I can also think of another 5 where it probably would have been a waste of time to even fiddle with.
One project I spent a bunch of time optimizing the write path of I/O. It was just using standard fwrite. But by staging items correctly it was an easy 10x speed win. Those optimizations sometimes stack up and count big. But it also had a few edges on it, so use with care.
During my time at Facebook, I maintained a bunch of kernel patches to improve jemalloc purging mechanisms. It wasn't popular in the kernel or the security community, but it was more efficient on benchmarks for sure.
Many programs run multiple threads, allocate in one and free in the other. Jemalloc's primary mechanism used to be: madvise the page back to the kernel and then have it allocate it in another thread's pool.
One problem: this involves zero'ing memory, which has an impact on cache locality and over all app performance. It's completely unnecessary if the page is being recirculated within the same security domain.
The problem was getting everyone to agree on what that security domain is, even if the mechanism was opt-in.
https://marc.info/?l=linux-kernel&m=132691299630179&w=2