The return of the frame pointers
brendangregg.com
brendangregg.com
During my first two years in college, if one of our programs did something funny, I would attach gdb and see what was happening at assembly level. I got used to "walking the stack" manually, though the debugger often helped a lot. Happy times, until all of the sudden, "-fomit-frame-pointer" was all the rage, and stack traces stopped making sense. Just like that, debugging that segfault or illegal instruction became exponentially harder. A short time later, I started using Python for almost everything to avoid broken debugging sessions. So, I lost an order of magnitude or two with "-fomit-frame-pointer". But learning Python served me well for other adventures.
There's a persistent myth that frame pointers have a huge overhead, because there was a single Python case that had a +10% slow down (now fixed). The actual measured overhead is under 1%, which is far outweighed by the benefits we've been able to make in certain applications.
Initially we just turned off frame pointers for the Python 3.9 interpreter in Fedora. They are back on in Python 3.12 where it seems the upstream bug has been fixed, although I can't find the actual fix right now.
Fedora tracking bug: https://bugzilla.redhat.com/2158729
Fedora change in Python 3.9 to disable frame pointers: https://src.fedoraproject.org/rpms/python3.9/c/9b71f8369141c...
FWIW your results don't quite match the measurements from Linux kernel folks who claim that the overhead is anywhere between 5-10%. Source: https://lore.kernel.org/lkml/20170602104048.jkkzssljsompjdwy...
I didn't preserve the data involved but in a variety of workloads including netperf, page allocator microbenchmark, pgbench and sqlite, enabling framepointer introduced overhead of around the 5-10% mark.
While it might call itself a benchmark, it behaves very microbenchmark-y.
The other numbers I and others have shared have been from actual production workloads. Not a simple program that tests same sequence of commands over and over.
The pgbench docs make it sound microbenchmark-y by describing making the same call over and over. If people find that this simulates actual production workloads, then yes, it can be considered a macro-benchmark.
From https://www.tpc.org/tpcb/, a TPC-B workload that pgbench runs by default:
In August 1990, the TPC approved its second benchmark, TPC-B. In contrast to TPC-A, TPC-B is not an OLTP benchmark. Rather, TPC-B can be looked at as a database stress test, characterized by:
Significant disk input/output
Moderate system and application execution time
Transaction integrity
TPC-B measures throughput in terms of how many transactions per second a system can perform. Because there are substantial differences between the two benchmarks (OLTP vs. database stress test), TPC-B results cannot be compared to TPC-A.
...
Transactions are submitted by programs all executing concurrently.You also missed the word "Obsolete" splattered all over the website you sent me, and the text that TPC-B was "Obsolete as of 6/6/95".
> which was about whether databases could even have micro-benchmarks.
No, this was an argument of yours that you pulled out out of nowhere. The topic very specifically was about the pgbench and not whether or not databases can have micro-benchmarks. Obvious answer is, yes, they can as any other software out there.
I think that you kinda tried to imply that pgbench was one of such micro-benchmarks in disguise and which is why I c/p the description which proves that it is not.
> You also missed the word "Obsolete"
I did not since that was not the topic being discussed at all. And in a technical sense, it doesn't matter at all. pgbench still runs so it is very much "not obsolete".
I'm guessing that this discussion would be more productive if you would please say who you are and the company you work for. I'm Brendan Gregg, I work for Intel, and I'm well known in the performance space. Who are you?
Wow, just wow. How ridiculous this is?
> Your position is also completely untenable: this benchmark was obsoleted by its creators 29 years ago, who very clearly say it is obsolete, and you're arguing that it isn't because it "still runs."
My only position in the whole thread was "1% of overhead cannot be universally claimed" and I still stand by it 100%. pgbench experiment from Linux kernel folks was just one of the counter-examples that can be found in the wild that goes against your claim. And which you first disputed by saying that it is a micro-benchmark (meaning that you have no idea what it is) and now you're disputing it by saying it's obsolete (yes, but still very relevant in database development, used in the Linux kernel and not the slightest technical reasoning after all).
Personally, I couldn't care less about this but if names is what you're after, you're not disagreeing with me but with the methodology and results presented by Mel Gorman, Linux kernel developer.
You are backing away from your other positions, for example:
> I fail to understand the reasoning of it "being simple" or "microbenchmarkey". It's far from the truth I think.
Do you now agree that TPC-B is too simple and microbenchmarky? And if not, please tell me (as I'm working on the problem of industry benchmarking in general) what would it take to convince someone like you to stop elevating obsoleted benchmarks like TPC-B? Is there anything?
Your replies demonstrate lack of technical depth in certain areas and which makes me personally doubt in your experiment results. And you know what, that is totally fine.
But your attitude? A total disbelief.
> Do you now agree that TPC-B is too simple and microbenchmarky?
No, why would I, you did not present any evidence that would support that claim of yours?
And contrary to you, and to your own embarrassment, I do not use personal attacks when I go out of technical depth.
> what would it take to convince someone like you to stop elevating obsoleted benchmarks like TPC-B? Is there anything?
You don't have to convince me anything. Remember that this is just a random internet page where people are exchanging their opinions.
Wish you a pleasant day.
I am just a minor PostgreSQL contributor and consultant of no import, but do you seriously think you are superior to PostgreSQL core devs and know more than them about PostgreSQL just because you are a Linux kernel dev? I really do not like your attitude here and your appeal to your own authority when you are not even an authority here.
Pgbench is used heavily both in the development of PostgreSQL and by people who tune PostgreSQL. So it is not obsolete. Maybe it is a bad benchmark and that the PostgreSQL community should stop using it but then you need a stronger argument than just some claim about obsoleteness from 1995. If a PostgreSQL core developer says in 2024 that it is still relevant I think that weighs a bit higher than a claim from 1995.
Some programs, like Python, have reported worse, 2-7% [2], but there is traction on tackling that [1a] (see both rwmj's and brendangregg's replies to sibling comments, they've both done a lot of upstreamed work wrt. frame pointers, performance, and profiling).
As has been frequently pointed out, the benefits from improved profiling cannot be understated, even a 10% cost to having frame pointers can be well worth it when you leverage that information to target the actual bottlenecks that are eating up your cycles. Plus, you can always disable it in specific hotspots later when needed, which is much easier than the reverse.
Something, something, premature optimisation -- though in seriousness, this information benefits actual optimisation, exactly because we don't have the information and understanding that would allow truly universal claims, precisely because things like this haven't been available, and so haven't been widely used. We know frame pointers, from additional register pressure and extended function prologue/epilogue, can be a detriment in certain hotspots; that's why we have granular control. But without them, we often don't know which hotspots are actually affected, so I'm sure even the databases would benefit... though the "my database is the fastest database" problem has always been the result of endless micro-benchmarking, rather than actual end-to-end program performance and latency, so even a claimed "10%" drop there probably doesn't impact actual real-world usage, but that's a reason why some of the most interesting profiling work lately has been from ideas like causal profilers and continuous profilers, which answer exactly that.
[0]: https://rwmj.wordpress.com/2023/02/14/frame-pointers-vs-dwar... [1a]: https://pagure.io/fesco/issue/2817#comment-826636 [1b]: https://pagure.io/fesco/issue/2817#comment-826805 [2]: https://discuss.python.org/t/the-performance-of-python-with-...
The choice made by Intel when they have designed 8086 to use 2 separate registers for the stack pointer and for the frame pointer was a big mistake.
It is very easy to use a single register as both the stack pointer and the frame pointer, as it is standard for instance in IBM POWER.
Unfortunately in the Intel/AMD CPUs using a single register is difficult, because the simplest implementation is unreliable since interrupts may occur between 2 instructions that must form an atomic sequence (and they may clobber the stack before new space is allocated after writing the old frame pointer value in the stack).
It would have been very easy to correct this in new CPUs by detecting that instruction sequence and blocking the interrupts between them.
Intel had already done this once early in the history of the x86 CPUs, when they have discovered a mistake in the design of the ISA, that interrupts could occur between updating the stack segment and the stack pointer. Then they had corrected this by detecting such an instruction sequence and blocking the interrupts at the boundary between those instructions.
The same could have been done now, to enable the use of the stack pointer as also the frame pointer. (This would be done by always saving the stack pointer in the top of the stack whenever stack space is allocated, so that the stack pointer always points to the previous frame pointer, i.e. to the start of the linked list containing all stack frames.)
Few can leverage that information because the open source software you are talking about lacks telemetry in the self hosted case.
The profiling issue really comes down to the cultural opposition in these communities to collecting telemetry and opening it for anyone to see and use. The average user struggles to ally with a trustworthy actor who will share the information like profiling freely and anonymize it at a per-user level, the level that is actually useful. Such things exist, like the Linux hardware site, but only because they have not attracted the attention of agitators.
Basically users are okay with profiling, so long as it is quietly done by Amazon or Microsoft or Google, and not by the guy actually writing the code and giving it out for everyone to use for free. It’s one of the most moronic cultural trends, and blame can be put squarely on product growth grifters who equivocate telemetry with privacy violations; open source maintainers, who have enough responsibilities as is, besides educating their users; and Apple, who have made their essentially vaporous claims about privacy a central part of their brand.
Of course people know the answer to your question. Why doesn’t Google publish every profile of every piece of open source software? What exactly is sensitive about their workloads? Meta publishes a whole library about every single one of its customers, for anyone to freely read. I don’t buy into the holiness of the backend developer’s “cleverness” or whatever is deemed sensitive, and it’s so hypocritical.
No; the groups are approximately "cares whether software respects the user, including privacy", or "doesn't know or doesn't care". I seriously doubt that any meaningful number of people are okay with companies invading their privacy but not smaller projects.
Continuous profiling of your system that gets relayed to someone else by telemetry is very different from continuous profiling of your own system, handled only by yourself (or, generalising, your community/group/company). You seem to be imagining we're operating more in the former, whereas I am imagining more in the latter.
When it's our own system, better instrumented for our own uses, and we're the only ones getting the information, then there's nothing to worry about, and we can get much more meaningful and informative profiling done when more information about the system is available. I don't even need telemetry. When it's "someone else's" system, in other words, when we have no say in telemetry (or have to exercise a right to opt-out, rather than a more self-executing contract around opt-in), then we start to have exactly the kinds of issues you're envisaging.
When it's not completely out of our hands, then we need to recognise different users, different demands, different contexts. Catering to the user matters, and when it comes to sensitive information, well, people have different priorities and threat models.
If I'm opening a calendar on my phone, I don't expect it to be heavily instrumented and relaying all of that, I just want to see my calendar. When I open a calendar on my phone, and it is unreasonably slow, then I might want to submit relevant telemetry back in some capacity. Meanwhile, if I'm running the calendar server, I'm absolutely wanting to have all my instrumentation available and recording every morsel I reasonably can about that server, otherwise improving it or fixing it becomes much harder.
From the other side, if I'm running the server, I may want telemetry from users, but if it's not essential, then I can "make do" with only the occasional opt-in telemetry. I also have other means of profiling real usage, not just scooping it all up from unknowing users (or begrudging users). Those often have some other "cost", but in turn, they don't have the "cost" of demanding it from users. For people to freely choose requires acknowledging the asymmetries present, and that means we can't just take the path of least resistance, as we may have to pay for it later.
In short, it's a consent issue. Many violate that, knowingly, because they care not for the consequences. Many others don't even seem to think about it, and just go ahead regardless. And it's so much easier behind closed doors. Open source in comparison, even if not everything is public, must contend with the fact that the actions and consequences are (PRs, telemetry traffic, etc), so it inhabits a space in which violating consent is much more easily held accountable (though no guarantee).
Of course, this does not mean it's always done properly in open source. It's often an uphill battle to get telemetry that's off-by-default, where users explicitly consent via opt-in, as people see how that could easily be undermined, or later invalidated. Many opt-in mechanisms (e.g. a toggle in the settings menu) often do not have expiration built in, so fail to check at a later point that someone still consents. Not to say that's the way you must do it, just giving an example of a way that people seem to be more in favour of, as with the generally favourable response to such features making their way into "permissions" on mobile.
We can see how the suspicion creeps in, informed by experience... but that's also known by another word: vigilance.
So, users are not "okay" with it. There's a power imbalance where these companies are afforded the impunity because many are left to conclude they have no choice but to let them get away with it. That hasn't formed in a vacuum, and it's not so simple that we just pull back the curtain and reveal the wizard for what he is. Most seem to already know.
It's proven extremely difficult to push alternatives. One reason is that information is frequently not ready-to-hand for more typical users, but another is that said alternatives may not actually fulfil the needs of some users: notably, accessibility remains hugely inconsistent in open source, and is usually not funded on par with, say, projects that affect "backend" performance.
The result? Many people just give their grandma an iPhone. That's what's telling about the state of open source, and of the actual cultural trends that made it this way. The threat model is fraudsters and scammers, not nation-state actors or corporate malfeasance. This app has tons of profiling and privacy issues? So what? At least grandma can use it, and we can stay in contact, dealing with the very real cultural trends towards isolation. On a certain level, it's just pragmatic. They'd choose differently if they could, but they don't feel like they can, and they've got bigger worries.
Unless we do different, the status quo will remain. If there's any agitation to be had, it's in getting more people to care about improving things and then actually doing them, even if it's just taking small steps. There won't be a perfect solution that appears out of nowhere tomorrow, but we only have a low bar to clear. Besides, we've all thought "I could do better than that", so why not? Why not just aim for better?
Who knows, we might actually achieve it.
For the record I'm NOT OK with being profiled, measured or otherwise studied in any way without my explicit consent. That even extends to the unethical human experiments that corporations run on people and which they euphemistically call A/B tests. I don't care if it's Google or a hobbyist developer, I will block it if I can and I will not lose a second of sleep over it.
True. But such people are like cockroaches. They know what they are doing will be unpopular with their targets, so they keep it hidden. This is easy enough to do in closed designs, car manufacturers selling your driving habits to insurance companies and health monitoring app selling menstrual cycle data to retailers selling to women.
Compare that do say Debian and RedHat. They too collect performance data. But the code is open source, Debian has repeatable builds so you can be 100% sure that is the code in use, and every so often someone takes a look it. Guess what, the data they send back is so unidentifiable it satisfies even the most paranoid of their 1000's of members.
All it takes is a little bit of sunlight to keep the cockroaches at bay, and then we can safely let the devs collect the data they need to improve code. And everyone benefits.
You're correct, results might be stale, although I wouldn't hold my breath for it since there has been no fundamental change in a way how frame pointers are handled as far as my understanding goes. Perhaps smaller improvements in compiler technology but CPUs did not undergo any significant change w.r.t. that context.
That said, nowhere in this thread have we seen a dispute of those Linux kernel results other than categorically rejecting them as being "microbenchmarks", which they are not.
> though the "my database is the fastest database" problem has always been the result of endless micro-benchmarking, rather than actual end-to-end program performance and latency
Quite the opposite. All database benchmarks are end-to-end program performance and latency analysis. "Cheating" in database benchmarks is done elsewhere.
* use framepointers [1]
* use LBR (but LBR has a limited depth, and may not work on on all CPUs, I'm assuming due to bugs in perf)
* implement some deep changes in how perf works to handle the 2 stacks in OCaml (I don't even know if this would be possible), or write/adapt some eBPF code to do it
OCaml 5 has a separate stack for OCaml code and C code, and although GDB can link them based on DWARF info, perf DWARF call-graphs cannot (https://github.com/ocaml/ocaml/issues/12563#issuecomment-193...)
If you need more evidence to keep it enabled in future releases, you can use OCaml 5 as an example (unfortunately there aren't many OCaml applications, so that may not carry too much weight on its own).
[1]: I haven't actually realised that Fedora39 has already enabled FP by default, nice! (I still do most of my day-to-day profiling on an ~CentOS 7 system with 'perf record --call-graph dwarf -F 47 -a', I was aware that there was a discussion to enable FP by default, but haven't noticed it has actually been done already)
Stuff like this is making me lean more towards getting an Intel for my next computer, at least on desktop where their worse power efficiency is less of an issue. But then they keep gating AVX-512 due to their E-cores not supporting it... You really can't win these days.
The reason we removed them was not a myth but comes from the pre-64 bit days. Not that long ago actually.
Even today if you want to repurpose older 64 bit systems with a new life then this of optimization still makes sense.
Ideally it should be the default also for security critical systems because not everything needs to be optimized for "observability"
Isn't that just 32-bit x86, which isn't used in IoT? The other 32-bit ISAs aren't register-starved like x86.
That said, I don’t know how valuable getting rid of FP on ARM is - I once benchmarked ffmpeg on 32-bit x86 before and after enabling FP and PIC (basically removing 2 GPRs) and the difference was huge (>10%) but that’s an extreme example.
> The frame pointer register (x29) must always address a valid frame record. Some functions — such as leaf functions or tail calls — may opt not to create an entry in this list. As a result, stack traces are always meaningful, even without debug information.
https://developer.apple.com/documentation/xcode/writing-arm6...
Even if $BIG_COMPANY makes a decision to compile everything with frame pointers, the rest of the community is not. So we'll be stuck fighting an unwinnable argument with a much larger community. Turns out that it was a ~20 year argument.
I ended up writing some patches to make libunwind work for gperftools and maintained libunwind for some number of years as a consequence of that work.
Having moved on to other areas of computing, I'm now a passive observer. But it's fascinating to read history from the other perspective.
In what way would you be stuck? What functional problems does adding frame pointers introduce?
Aee also this benchmark from Phoronix [0]:
Of the 100 tests carried out for this article, when taking the geometric mean of all these benchmarks it equated to about a 14% performance penalty of the software with -O2 compared to when adding -fno-omit-frame-pointer.
The actual «wastage» is in having to generate a prologue and an epilogue for each function – 2x instructions to preserve the old frame pointer and set a new one up, and 2x instruction at the point of return – to restore the previous frame pointer.
Generally, it is not a big deal with an exception of a pathological case of a very large number of very small functions calling each other frequently where the extra 4x instructions per each such a function will be filling up the L1 instruction cache «unnessarily».
There is also a significant number of projects that go to great lengths to force-override CFLAGS/CXXFLAGS (usually with «-O2 -g» or even with «-O») or make it extraordinary difficult to change the project's default CFLAGS, for no apparent reasons which eliminates a number of advanced optimisations in builds with default build settings.
All of the companies involved believed that they were special and decided to build their own (poorly managed) distribution called "third party code" and having to deal with it was not my best experience working at these companies.
What existed was a x86 to x86 cross compilation environment and the libraries involved were manually imported by developers who needed that particular library.
My argument was about the cost of ensuring that those libraries were compiled with frame pointers when much of the open source community was defaulting to omit-fp.
In 2024, you'd probably create a "build container" that all developers are required to use to build binaries or pay a linux distributor to build that container.
But cross compilation was the preferred approach back then. So all binaries had a rpath (run time search path to look for shared library) that ignored the distributor supplied libraries.
Having come from a open source background, I found this system hard to digest. But there was a lot of social pressure to work as a bee in a system that thousands of other very competent engineers are using (quite successfully).
I remember briefly talking to a chrome OS related group who were using the "build your own custom distro" approach, before deciding to move to another faang.
What does this mean?
Since Google would rather have the best brains in the industry build the next search indexing algorithm or the browser, they didn't have the time to invest human capital into building a better open source friendly dev environment.
A natural alternative is to contract out the work. Linux distributors were good candidates for such contract work.
But the vibe back then was Google could build better alternatives to some of these libraries and therefore bridging the gap between dev experience as an open source dev vs in house software engineer wasn't important.
You could see the same argument play out in git vs monorepo etc, where people take up strong positions on a particular narrow tech topic, whereas the larger issue gets ignored as a result of these belief systems.
https://bugzilla.redhat.com/show_bug.cgi?id=2267598 https://github.com/tukaani-project/xz/commit/82ecc538193b380...
Why do we normally store the return addresses near to the local variables in the first place, again? There are so many downsides.
In exchange, it becomes very difficult for the stack smashing to corrupt the return address.
[1] https://en.wikipedia.org/wiki/Forth_(programming_language)
STOIC, a variant of Forth, includes a file stack when loading words
The dictionary pointer can also be manipulated in some dialects. That can be used directly as the stack variant of the arena allocator idea. It is particularly useful for text concatenation.
[1] https://forth-standard.org/standard/search [2] https://forth-standard.org/standard/core/MARKER
The advantage of storing them elsewhere is not quite clear (unless you have hardware support for things like shadow stacks).
You'd have to argue that the cost of moving things to this other page and managing two pointers (where one is less powerful in the ISA) is meaningfully cheaper than the other equally effective mitigation of stack cookies/protectors which are already able to provide protection only where needed. There is no real security benefit to doing this over what we currently have with stack protectors since an arbitrary read/write will still lead to a CFI bypass.
The classic buffer overflow issue should spring immediately to mind. By having a separate return address stack it's far less vulnerable to corruption through overflowing your data structures. This stops a bunch of attacks which purposely put crafted return addresses into position that will jump the program to malicious code.
It's not a panacea, but generally keeping code pointers away from data structures is a good idea.
The main driver is in (https://github.com/titzer/virgil/blob/master/rt/native/Nativ... the rest of the code in the directory implements the decoding of metadata.
I think frame pointers only make sense if frames are dynamically-sized (i.e. have stack allocation of data). Otherwise it seems weird to me that a dynamic mechanism is used when a static mechanism would suffice; mostly because no one agreed on an ABI for the metadata encoding, or an unwind routine.
I believe the 1-2% measurement number. That's in the same ballpark as pervasive checks for array bounds checks. It's weird that the odd debugging and profiling task gets special pleading for a 1% cost but adding a layer of security gets the finger. Very bizarre priorities.
And for those who value performance and don't want to pay the cost of "a lot more than 1-2%" ;p
Btw, in my experience, std::vector is fast. Insanely fast. "I don't understand how it can be so fast", "barely distinguishable from raw arrays in C" fast. Not doing bounds checking is probably part of that, though far from the whole story.
https://www.youtube.com/watch?v=RK9bfrsMdAM https://llvm.org/devmtg/2023-05/slides/TechnicalTalks-May11/...
It's a myth that bounds checking has extraordinary performance costs and cannot be enabled without slowing everything to a halt. Maybe this was the case 10 years ago or 20 years ago or something, but not today.
> Profiling has been broken for 20 years and we've only now just fixed it.
It was a shame when they went away. Lots of people, certainly on other systems and probably Linux too, have found the absence of frame pointers painful this whole time and tried to keep them available in as many environments as possible. It’s validating (if also kind of frustrating) to see mainstream Linux bring them back.
Since it can't unwind the user space stacks, the kernel has to copy the entire stack (8192 bytes by default) into the output file (perf.data) and then the perf user space program will unwind it later. It does this for each sample, which is usually hundreds of times per second, per CPU. Though it depends how you configured the collection.
That does have a significant overhead: first, runtime overhead: copying 8k bytes, hundreds of times per second, and writing it to disk, all don't come for free. You spend quite a bit of CPU time doing the memcpy operation which consumes memory bandwidth too. You also frequently need to increase the size of the perf memory buffer to accommodate all this data while it waits for user space to write it to disk. Second, disk space overhead, since the 8k stack bytes per sample are far larger than the stack trace would be. And third, it does require that you install debuginfo packages to get the DWARF info, which is usually a pain to do on production machines, and they consume a lot of disk space on their own.
Many of these overheads aren't too bad in simple cases (lower sample rates, fewer CPUs, or targeting a single task). But with larger machines with hundreds of CPUs, full system collections, and higher frequencies, the overhead can increase exponentially.
I'm not certain I know what you mean by the "slow unwinding path for perf", as there is no faster path for user space when frame pointers are disabled (except Intel LBR as outlined in the blog).
The difference between -g (--call-graph fp) and --call-graph dwarf is large even with perf linked directly against binutils, at least in my experience (and it was much, much worse until I got some patches into binutils to make it faster). This is both on record and report.
There are also weird bugs with --call-graph dwarf that perf upstream isn't doing anything about, around inlining. It's not perfect by any means.
* Many frame pointer unwinders don't account for a problem they have that DWARF unwind info doesn't have: the fact that the frame set-up is not atomic, it's done in two instructions, `push $rbp` and `mov $rsp $rbp`, and if when a snapshot is taken we are in the `push`, we'll miss the parent frame. I think this might be able to be fired by inspecting the code, but I think this might only be as good as a heuristic as there could be other `push %rbp` unrelated to the stack frame. I would love to hear if there's a better approach!
* I developed the solution Brendan mentions which allows faster, in-kernel unwinding without frame pointers using BPF [0]. This doesn't use DWARF CFI (the unwind info) as-is but converts it into a random-access format that we can use in BPF. He mentions not supporting JVM languages, and while it's true that right now it only supports JIT sections that have frame pointers, I planned to implement a full JVM interpreter unwinder. I have left Polar Signals since and shifted priorities but it's feasible to get a JVM unwinder to work in lockstep with the native unwinder.
* In an ideal world, enabling frame pointers should be done on a case-by-case. Benchmarking is key, and the tradeoffs that you make might change a lot depending on the industry you are in, and what your software is doing. In the past I have seen large projects enabling/disabling frame pointers not doing an in-depth assessment of losses/gains of performance, observability, and how they connect to business metrics. The Fedora folks have done a superb and rigorous job here.
* Related to the previous point, having a build system that enables you to change this system-wide, including libraries your software depends on can be awesome to not only test these changes but also put them in production.
* Lastly, I am quite excited about SFrame that Indu is working on. It's going to solve a lot of the problems we are facing right now while letting users decide whether they use frame pointers. I can't wait for it, but I am afraid it might take several years until all the infrastructure is in place and everybody upgrades to it.
- [0]: https://web.archive.org/web/20231222054207/https://www.polar...
PUSH EBP
MOV ESP,ESP
SUB SP,N
As for the reason why ENTER isn't used, it was deemed too slow. LEAVE (MOV SP,BP; POP BP) is used as it's just as fast as, if not faster, than the sequence it replaces. If ENTER were just the PUSH/MOV/SUB sequence, it probably would be used, but it's that other operand (which is 0 above in my example) that kills it performance wise (it's for nested functions to gain access to upper stack frames and is every expensive to use).
It would be really nice to have fine-grained control over frame pointer inclusion: provided fine-grained profiling, we could determine whether we needed the frame pointers for a given function or compilation unit. I wouldn't be surprised if we see that only a handful of operations are dramatically slowed by frame pointer inclusion while the rest don't really care.
No it's not, particularly when it can help you identify hotspots via profiling that can net you improvements of 10% or more.
If you're netflix then "enable frame pointers" is a no-brainer. But if you're a distro who's building code for millions of users, many of whom will likely never need to fire up a profiler, I think the question is at least a little trickier. The overall best tradeoff might end up being still to enable frame pointers, but I can see the other side too.
In the end a 2% of performance of every application it's a big deal. On a single computer may not be that significant, think about all the computers, servers, clusters, that run a Linux distro. And yes, I would ask a Google engineer that if scaled on the I don't know how many servers and computers that Google has a 2% increase in CPU usage is not a big deal: we are probably talking about hundreds of kW more of energy consumption!
We talk a lot of energy efficiency these days, to me wasting a 2% only to make the performance analysis of some software easier (that is that you can analyze directly the package shipped by the distro and you don't have to recompile it) it's something stupid.
Not having them enabled on all dependencies makes them significantly less useful if your application interacts with the system in a meaningful way. "Just recompile" (glibc + xorg + qt + whatever else you use) is a very hefty thing.
Having them enabled by default also enables devs to ask their users for traces. I can reasonably tell someone to install sysprof or something, click a few buttons and send me a trace. I cannot reasonably tell them to compile the project (+ dependencies). And tbh, even devs cba if they have to jump through too many hoops.
> We talk a lot of energy efficiency these days, to me wasting a 2% only to make the performance analysis of some software easier (that is that you can analyze directly the package shipped by the distro and you don't have to recompile it) it's something stupid.
We're trading all kinds of efficiencies for all kinds of niche benefits all the time. This enables analysis of what is shipped in real use across the whole stack with the alternative being much less useful or a lot more (human) work. To me it's worth 2% and I'm shocked (well... not really) it's not for more people.
Code eking out every possible cycle of performance can enable a no-frame-pointer optimization and see if it helps. But it's a bad default for libc, and for the kernel.
Don't completely trust the benchmarks on this; they are a bit synthetic and real-world applications tend to produce very different results.
Plus, profiling is important. I was able to speed up various segments of my code by up to 20 per cent by profiling them carefully.
And, at the end of the day, if your application is so sensitive about any loss of performance, you can simply profile your code in your lab using frame pointers, then omit them in the version released to your customers.
That is what should be done but TFA is about distros shipping code with frame pointers to end uses because some developers are too lazy to recompile libc when profiling. Somehow shipping different copies of libc, one indended for end users on low-powered devices and one indended for developers is not even considered.
__attribute__((optimize("no-omit-frame-pointer")))
__attribute__((optimize("omit-frame-pointer")))I think Intel's DNN things dump their info out to some common file that perf can read instead, but because the *kernels* themselves reuse rbp throughout oneDNN, it's totally useless.
Finally, can any JVM folks explain this claim about DWARF info from the article:
> Doesn't exist for JIT'd runtimes like the Java JVM
that just sounds surprising to me. Is it off by default or literally not available? (Google searches have mostly pointed to people wanting to include the JNI/C side of a JVM stack, like https://github.com/async-profiler/async-profiler/issues/215).
The fact that people complain about the performance of the mechanism that enables the system to be profiled, and so performance problems be identified, is beyond ironic. Surely the epitome of premature optimisation.
the funny thing is i am not sure if the world would actually work properly if we didn't have both of these kinds of people.
I don't see the people writing hyperfast low-level code asking for this.
My experience with profiling is anyway that you don't need that fine-grained profiling. It's main use is finding stuff like "we spend 90% of our time reallocating a string over and over to add characters one at a time". After a few of those it's just "it's a little bit slow everywhere".
And after that, you need fine-grained profiling to find multiple 1% wins and apply them repeatedly.
(I do this for a living, basically)
That is, on i686 you have 7 GPRs without frame pointers, while on x86_64 you have 14 GPRs even with frame pointers.
Copying a comment of mine from an older related discussion (https://news.ycombinator.com/item?id=38632848):
"To emphasize this point: on 64-bit x86 with frame pointers, you have twice as many registers as on 32-bit x86 without frame pointers, and these registers are twice as wide. A 64-bit value (more common than you'd expect even when pointers are 32 bits) takes two registers on 32-bit x86, but only a single register on 64-bit x86."
So this is a solution for a problem, and it arrives just at the moment that people have solved the problem more generically ;)
(Prodfiler coauthor here, we had solved all of this by the time we launched in Summer 2021)
https://gitlab.com/freedesktop-sdk/freedesktop-sdk/-/issues/...
So not a perf issue there, but they don't think the workflow is suitable for whole-system profiling. Perf issues were in the context of `perf` using DWARF:
https://gitlab.com/freedesktop-sdk/freedesktop-sdk/-/issues/...
So it's both similarly fragile, but one is almost never disabled.
The broader point is: For HLL runtimes you need to be able to switch between native and interpreted unwinds anyhow, so you'll always do some amount of lifting in eBPF land.
And yes, having frame pointers removes a lot of complexity, so it's net a very good thing. It's just that the situation wasnt nearly as dire as described, because people that care about profiling had built solutions.
Things get even more complicated because context switches can mean CPU migrations, making many of your data useless.
> Things get even more complicated because context switches can mean CPU migrations, making many of your data useless.
No it doesn't. If a user space thread is blocked on doing kernel work, its stack isn't going to change, not even if that thread ends up resuming on a different thread.
But I think you're only thinking about CPU profiling at <= 100 Hz / core. However, Brendan's article is also talking about Off-CPU profiling, and as far as I can tell, all known techniques (scheduler tracing, wall clock sampling) require stack unwinding to occur 1-3 orders of magnitude more often than for CPU profiling.
For those use cases, I don't think .eh_frame unwinding will be good enough, at least not for continuous profiling. E.g. see [1][2] for an example of how frame pointer unwinding allowed the Go runtime to lower execution tracing overhead from 10-20% to 1-2%, even so it was already using a relatively fast lookup table approach.
[1] https://go.dev/blog/execution-traces-2024
[2] https://blog.felixge.de/reducing-gos-execution-tracer-overhe...
Does this mean Microsoft decided they weren't going to support breaking profilers and debuggers OR is there some magic in the pdata section that makes it work even if you omit the frame-pointer?
“Recovering a broken stack on x64 machines on Windows is trickier because the x64 uses unwind codes for stack walking rather than a frame pointer chain.”
More details are here: https://learn.microsoft.com/en-us/cpp/build/exception-handli...
The default is omission. If you have a Windows machine, in all likelihood almost no 64 bit code running on it has frame pointers.
> OR is there some magic in the pdata section that makes it work even if you omit the frame-pointer
You haven't ever needed frame pointers to unwind using ... unwind information. The same thing exists for linux as `.eh_frame` section.
And so, (as you again have noted), your best bet is to just use plain old logging and tracing (like what datadog provides) to find out where the waiting is happening.
Some people may believe that 100ms latency is acceptable for a CLI tool, but what if it could be 3ms? On some aesthetic level, it also feels good to be able to eliminate excess. Finally, you should learn it because you won't necessarily have that job forever.
> The overhead to walk DWARF is also too high, as it was designed for non-realtime use.
Source: I wrote multiple profilers.
Granted, this is using optimized tooling which uses a bunch of tricks to side-step the problem of DWARF being slow, I only profile native code (and some VMs which do ahead-of-time codegen) and I've never worked with JVM, but in principle I don't see why it wouldn't be practical on JVM too, although it certainly would be harder and might require better tooling (which might not exist currently). If you have the luxury of enabling frame pointers then that certainly would be easier and simpler.
(Somewhat related, but I really wish we would standardize on something better than DWARF for unwinding tables and basic debug info. Having done a lot of work with DWARF and its complexity I wouldn't wish it upon my worst enemy.)
DWARF-based unwinding can be a bottleneck for time-sensitive program analysis tools. For instance the perf profiler is forced to copy the whole stack on taking each sample and to build the backtraces offline: this solution has a memory and time overhead but also serious confidentiality and security flaws.
However, paper also mentions
Similarly, the Linux kernel by default relies on a frame pointer to provide reliable backtraces. This incurs in a space and time overhead; for instance it has been reported (https://lwn.net/Articles/727553/) that the kernel’s .text size increases by about 3.2%, resulting in a broad kernel-wide slowdown.
Measurements have shown a slowdown of 5-10% for some workloads (https://lore.kernel.org/lkml/20170602104048.jkkzssljsompjdwy@suse.de/T/#u).> Javier Honduvilla Coto (Polar Signals) did some interesting work using an eBPF walker to reduce the overhead, but...Java.
I'm waiting for this to happen: https://github.com/open-telemetry/community/issues/1918
Samply unwinds significantly faster than `perf` because it caches unwind information.
That being said, this approach still has some limitations, such as that very deep stacks won't be unwound, as the size of the process stack the kernel sends is quite limited.
Is it because I avoided Linux and C most of my life? Perhaps it's because I used debug, and Periscope before that... and never gdb?
in the discussion the issue gets conflated with jit-ted languages, but that has nothing to do with the crusade to enable frame pointer for system libraries.
and if you care that much for dwarf overhead... just cache the unwind information in your system-level profiler? no need to rebuild everything.
This is solvable as Brendan calls out, we’ve created an eBPF-based profiler at Polar Signals, that essentially does what you said, it optimized the unwind tables, caches them in bpf maps, and then synchronously unwinds as opposed to copying the whole stack into user-space.
By default on most Linux distros the frame tables are built into all the binaries, and end up in the GNU_EH_FRAME segment, which is always available in any running process. Doesn't sound a broken and disjointed mess to me. Sounds more like a smoothly running solved problem.
- [0] This line information might or might not be present in an executable but luckily there's debuginfod (https://sourceware.org/elfutils/Debuginfod.html)
I wasn’t doing extreme performance stuff when -fomit-frame-pointer became the norm, so maybe it was a big win for enough people to be a sane default, but even that seems dubious: “just works” profiling is how you figure out when you’re in an extreme performance scenario (if you’re an SG14 WG type, you know it and are used to all the defaults being wrong for you).
I’m deeply grateful for all the legends who have worked on libunwind, gperf stuff, perftool, DTrace, eBPF: these are the too-often-unsung heroes of software that is still fast after decades of Moore’s law free-riding.
But they’ve been fighting an uphill battle against a weird alliance of people trying to game compiler benchmarks and the really irresponsible posture that “developer time is more expensive” which is only sometimes true and never true if you care about people on low-spec gear, which is the community of users who that is already the least-resourced part of the global community.
I’m fortunate enough to have a fairly modern desktop, laptop, and phone: for me it’s merely annoying that chat applications and music players and windowing systems offer nothing new except enshittification in terms of features while needing 10-100x the resources they did a decade ago.
But for half of my career and 2/3rds of my time coding, I was on low-spec gear most of the time, and I would have been largely excluded if people didn’t care a lot about old computers back then.
I’m trying to help a couple of aspiring hackers get started right now it’s a real struggle to get their environments set up with limitations like Intel Macs and WSL2 as the Linux option (WSL2 is very cool but it’s not loved up enough by e.g. yarn projects).
If you want new hackers, you need to make things work well on older computers.
Thanks again Brendan et al!
"I could say that times have changed and now the original 2004 reasons for omitting frame pointers are no longer valid in 2024."
The original 2004 reason for omitting frame pointers is still valid in 2024: it's still a big performance win on the register-starved 32-bit x86 architecture. What has changed is that the 32-bit x86 architecture is much less relevant nowadays (other than legacy software, for most people it's only used for a small instant while starting up the firmware), and other common 32-bit architectures (like embedded 32-bit ARM) are not as register-starved as the 32-bit x86.
There definitely have been regressions with frame pointers being enabled, although we've fixed all the ones we've found in current (2024) Fedora.
-fno-schedule-insns2
-fno-omit-frame-pointers
-fno-optimize-sibling-calls
-mno-omit-leaf-frame-pointer
-fpatchable-function-entry=18,16
-fno-inline-functions-called-once
I thought you were asking about the design of the hardware: it's designed that way because compatibility means that the vast majority of people want something backwards compatible.
Very little is fully "compiled from scratch" when you consider for example that libc is in the set of things that must be recompiled.
If one were to design a modern ISA from scratch it would make sense though.
[1] D Language Reference: Function Pointers, Delegates and Closures:
https://dlang.org/spec/function.html#function-pointers-deleg...
[2] Bret Victor The Future of Programming:
But those days has 32bit processors with few registers.
Times change.
I guess the similar situation with msvcrt’s.
Far more interesting I recall there might be an instruction where rbp isn't allowed.