This is my all-time favorite paper. It's so easy to read, and there's so much to think about, so much that still applies to everyday programming and language dedign.
Also there's Knuth admitting he avoids GO TO because he is afraid of being scolded by Edsger Dijkstra.
"It is clearly better to write programs in a language that reveals the control structure, even if we are intimately conscious of the hardware at each step; and therefore I will be discussing a structured assembly language called PL/MIX in the fifth volume of The art of computer programming"
This statement in the introduction applies to so many things in CS:
I have the
uncomfortable feeling that others are making
a religion out of it, as if the conceptual
problems of programming could be solved by
a single trick, by a simple form of coding
discipline!
Yeah, it's one of my favourites as well. And it weirds me out that the "premature optimization" bit is the most quoted piece of that paper — arguably, that's the least interesting part of it, compared to the musings on the language design and (semi)automatic transformation of algorithms!
I personally find that "break/continue" with (optionally labelled) loops cover about 95% of GOTO's use cases today (including "one-and-a-half" loops), although e.g. Rust and Zig, with their expression-oriented design (and support for sum types) offer an option to experiment with Zahn's event mechanism... but usually factoring an inner loop body into a separate function is a more clear-to-read approach (as for efficiency? Hopefully you have a decent inliner idk).
The only thing I truly miss, occasionally, is a way to concisely write an one-and-a-half range/FOR loop. Something like this:
for key, value in enumerate(obj):
emit(f'{key}={value}')
between:
emit(',')
When you have a generic "while True:" loop, conditional "break" works fine enough, but here? Ugh.
Love this paper and read it several times, most recently around 10 years ago when thinking about whether there were looping constructs missing from popular programming languages.
I have made the same point several times online and in person that the famous quote is misunderstood and often suggest people take the time to go back to the source and read it since it’s a wonderful read.
I will say I have worked on projects where sales and upper management were both saying the same thing: customers don't like that our product is slow(er than a competitors), and the devs just shrug and say we've already done everything we can. In one case the most senior devs even came up charts to prove this was all they were going to get.
Somehow I found 40%. Some of it was clever, a lot of it was paying closer attention to the numbers, but most of it was grunt work.
Besides the mechanical sympathy, the other two main tools were 1) stubbornness, and 2) figuring out how to group changes along functional testing boundaries, so that you can justify making someone test a change that only improves perf by 1.2% because they're testing a raft of related changes that add up to more than 10%.
Most code has orphaned performance improvements in 100 little places that all account for half of the runtime because nobody can ever justify going in and fixing them. And those can also make parallelism seem unpalatable due to Amdahl.
I think the problem with the quote is that everyone forgets the line that comes after it.
We should forget about small efficiencies, say about 97% of the time: premature optimization is the root of all evil.
vvvvvvvvvv
Yet we should not pass up our opportunities in that critical 3%. A good programmer will not be lulled into complacency by such reasoning, he will be wise to look carefully at the critical code; but only after that code has been identified.
^^^^^^^^^^
This makes it clear, in context, that Knuth defines "Premature Optimization" as "optimizing before you profile your code"
@OP, I think you should lead with this. I think it gets lost by the time you actually reference it. If I can suggest, place the second paragraph after
> People always use this quote wrong, and to get a feeling for that we just have to look at the original paper, and the context in which it was written.
The optimization part gets lost in the middle and this, I think, could help provide a better hook to those who aren't going to read the whole thing. Which I think how you wrote works good for that but the point (IMO) will be missed by more inattentive readers. The post is good also, so this is just a minor critique because I want to see it do better.
Amdahl’s Law is the single best thing I learned in 4 years of university. It sounds obvious when spelled out but it blew my mind.
No amount of parallelization will make your program faster than the slowest non-parallelizable path. You can be as clever as you want and it won’t matter squat unless you fix the bottleneck.
This extends to all types of optimization and even teamwork. Just make the slowest part faster. Really.
It is exactly this "lulled into complacency" that I rail against when most people cite that line. Far too many people are trying to shut down down dialog on improving code (not just performance) and they're not above Appeal to Authority in order to deflect.
"Curiosity killed the cat, but satisfaction brought it back." Is practically on the same level.
If you're careful to exclude creeping featurism and architectural astronautics from the definition of 'optimization', then very few people I've seen be warned off of digging into that sort of work actually needed to be reined in. YAGNI covers a lot of those situations, and generally with fewer false positives. Still false positives though. In large part because people disagree on what "The last responsible moment" in part because our estimates are always off by 2x, so by the time we agree to work on things we've waited about twice as long as we should have and now it's all half assed. Irresponsible.
A lot of people are still thinking of Knuth's comment as being about just finding the slow path or function and making it faster. What Knuth has talked about, however, and why any senior engineer who cares about performance has either been taught or discovered, is that most real optimization is in the semantics of the system and - frankly - not optimizing things but finding ways not to do them at all.
Tuning the JSON parser is not nearly as effective as replacing it with something less stupid, which is, in turn, not nearly as effective as finding ways to not do the RPC at all.
Most really performant systems of a certain age are also full of layer violations for this reason. If you care about performance (as in you are paid for it), you learn these realities pretty quickly. You also focus on retaining optionality for future optimizations by not designing the semantics of your system in a way that locks you permanently into low performance, which is unfortunately very common.
Another example of this is using the right algorithm for the job. I see a lot of code like let's say you have two lists and you need to go through the first one and find the corresponding element by id in the second. The naive implementation uses linear search of the second list, making the algorithm O(n^2). This will stall and take hours, days or longer to run when the lists get large, say into the tens to hundreds of thousands of elements.
If both lists come from your own database then you should have used a foreign key constraint and had the database join them for you. If you can't do that, let's say one or both lists come from a third-party api, you can create a dictionary(hashmap) of the second list to allow O(1) lookup rather than O(n) which brings your full algorithm's complexity to O(n). Now the you can have millions of elements and it'll still run in milliseconds or seconds. And if this is something you need to do very often then consider storing the second list in your database so you can use a join instead. Avoid doing the work, as you say.
These are the most common mistakes I see people make. That and situations where you make one request, wait for a response, then use that to make another request, wait for response and so on.
You don't need a profiler to avoid this type of problem, you just need to understand what you're doing at a fundamental level. Just use a dictionary by default, it doesn't matter if it might be slightly slower for very small inputs - it'll be fast enough and it'll scale.
We had one of those, except it was two loops and a python "in" on a list, so actually O(n^3). Changing the "in" to be on a set brought it from around two hours to around five minutes. And the best part is related to the topic at hand: This was a small postprocessing function, and other more experienced developers had tried to speed up the overall process but assumed the slowness was earlier in the process instead of in this function, and after failing they assumed it couldn't be sped up further. They never bothered to check which part of the process was actually slow.
> This makes it clear, in context, that Knuth defines "Premature Optimization" as "optimizing before you profile your code"
As with that Google testing talk ("We said all your tests are bad, but that's not entirely true because most of you don't have any tests"), the reality is that most people don't have any profiling.
If you don't have any profiling then in fact every optimisation is premature.
The operative word was always "premature". As with everything, whether or not a particular optimization is worthwhile depends on its ROI, which in the case of software depends on how often your code is going to be run. If you are building a prototype that's only going to run once, then the only optimizations that are worthwhile are the ones that speed things up by times comparable to the time it take to write them. On the other hand, if you're writing the inner loop of an LLM training algorithm, a 1% speedup can be worth millions of dollars, so it's probably worth it even if it takes months of effort.
You can almost simplify it to simply observing that PREMATURE optimization is not the same as OPTIMIZATION.
Most people I see who get offended are reacting to the claim that optimization is never useful. But it's pretty easy to knock that claim over.
I don't deny that plenty of people use adjectives very sloppily, and that much writing is improved by just ignoring them, but Knuth is not one of them.
Yes but the second half of the second half is "only after that code has been identified." So the advice is still 'don't waste your time until you profile.'
If you have benchmarked something then optimizations are not premature. People often used to 'optimize' code that was rarely run, often making the code harder to read for no gain.
beware too of premature pessimization. Don't write bubble sort just because you haven't benchmarked your code to show it is a bottleneck - which is what some will do and then incorrectly cite premature optimization when you tell them they should do better. Note that any compitenet languare has sort in the standard library that is better than bubble sort.
Most of the time, the arguments aren’t even that cogent, they’re more along the lines of “I’m not going to specify the columns I need from this SELECT query, because that’s premature optimization.”
The word “typesetting” does not refer to all aspects of visual appearance, but merely to the crucial decision of where each character goes: what glyph (from the fonts available) is placed at what position on the page. This paper was first published in the ACM Computing Surveys journal, and the typesetting is fine (as you'll find if you pick up a physical copy of the journal) — I think what you're complaining about is that all PDF versions that have been put up online so far (including the official one available via https://doi.org/10.1145/356635.356640) are from the same poor scan of the journal: it's not the typesetting that is atrocious, but the scanning quality (converting the printed page into digital images).
> Usually people say “premature optimization is the root of all evil” to say “small optimizations are not worth it” but […] Instead what matters is whether you benchmarked your code and whether you determined that this optimization actually makes a difference to the runtime of the program.
In my experience the latter is actually often expressed. What else would “premature” mean, other than you don’t know yet whether the optimization is worth it?
The disagreement is usually more about small inefficiencies that may compound in the large but whose combined effects are difficult to assess, compiler/platform/environment-dependent optimizations that may be pessimizations elsewhere, reasoning about asymptotic runtime (which shouldn’t require benchmarking — but with cache locality effects sometimes it does), the validity of microbenchmarks, and so on.
The way I often hear it expressed has nothing to do with small efficiency changes or benchmarking and it’s more of a yagni/anticipating hyper scale issue. For example, adding some complexity to your code so it can efficiently handle a million users when you’re just fine writing the simple to read and write version for hat isn’t optimal but will work just fine for the twenty users you actually have.
I understood it to mean that optimising for speed at the expense of size is a bad idea unless there are extremely obvious performance improvements in doing so. By default you should always optimise for size.
I like this article. It’s easy to forget what these classic CS papers were about, and I think that leads to poorly applying them today. Premature optimisation of the kind of code discussed by the paper (counting instructions for some small loop) does indeed seem like a bad place to put optimisation efforts without a good reason, but I often see this premature optimisation quote used to:
- argue against thinking about any kind of design choice for performance reasons, eg the data structure decisions suggested in this article
- argue for a ‘fix it later’ approach to systems design. I think for lots of systems you have some ideas for how you would like them to perform, and you could, if you thought about it, often tell that some designs would never meet them, but instead you go ahead with some simple idea that handles the semantics without the performance only to discover that it is very hard to ‘optimise’ later.
Oh man, I hate how often this is used. Everyone knows there's nothing more permanent than a temporary fix lol.
But what I think people don't realize is that this is exactly what tech debt is. You're moving fast but doing so makes you slow once we are no longer working in a very short timeline. That's because these issues compound. Not only do we repeat that same mistake, but we're building on top of shaky ground. So to go back and fix things ends up requiring far more effort than it would have taken to fix it early. Which by fixing early your efforts similarly compound, but this time benefiting you.
I think a good example of this is when you see people rewrite a codebase. You'll see headlines like "by switching to rust we got a 500% improvement!" Most of that isn't rust, most of that is better algorithms and design.
Of course, you can't always write your best code. There's practical constraints and no code can be perfect. But I think Knuth's advice still fits today, despite a very different audience. He was talking to people who were too obsessed with optimization while today were overly obsessed with quickly getting to some checkpoint. But the advice is the same "use a fucking profiler". That's how you find the balance and know what actually can be put off till later. It's the only way you can do this in an informed way. Yet, when was the last time you saw someone pull out a profiler? I'm betting the vast majority of HN users can't remember and I'd wager a good number never have
I completely agree with most of what you've said, but personally I rarely use a profiler. I don't need it, I just think about what I'm doing and design things to be fast. I consider the time complexity of the code I'm writing. I consider the amount of data I'm working with. I try to set up the database in a way that allows me to send efficient queries. I try to avoid fetching more data than I need. I try to avoid excessive processing.
I realize this is a very personal preference and it obviously can't be applied to everyone. Someone with less understanding might find a profiler very useful and I think those people will learn the same things I'm talking about - as you find the slow code and learn how to make it fast you'll stop making the same mistakes.
A profiler might be useful if I was specifically working to optimize some code, especially code I hadn't written myself. But for my daily work it's almost always good enough to keep performance in mind and design the system to be fast enough from the bottom up.
Most code doesn't have to be anywhere near optimal, it just has to be reasonably fast so that users don't have to sit and stare at loading spinners for seconds at a time. Some times that's unavoidable, some times you're crunching huge amounts of data or something like that. But most of the time, slow systems are slow because the people who designed and implemented them didn't understand what they were doing.
> It’s easy to forget what these classic CS papers were about, and I think that leads to poorly applying them today.
Notably, pretty much the entire body of discourse around structured programming is totally lost on modern programmers failing to even imagine the contrasts.
It is interesting that there's so much discourse about the effort people have had to put into data structure and algorithm stuff for interviews, but then people refuse to take advantage of the knowledge studying that gives you towards trivial effort optimizations (aka your code can look pretty similar, just using a different data structure under the hood for example).
That's because people don't get it. You see people saying things like "It's pointless to memorize these algorithms I'll never use" - you're not supposed to memorize the specific algorithms. You're supposed to study them and learn from them, understand what makes them faster than the others and then be able to apply that understanding to your own custom algorithms that you're writing every day.
I 100% agree. I could have written the same comment.
The biggest way I see this is picking an architecture or programming language (cough Python) that is inherently slow. "We'll worry about performance later" they say, or frequently "it's not performance critical".
Cut to two years later, you have 200k lines of Python that spends 20 minutes loading a 10GB JSON file.
I failed to put an important point on the above comment, which is spending too much time designing the perfect thing can lead to a system that is either never finished, or one that meets the goals but doesn’t do the work it was originally intended to do, and then it’s too late to pivot to doing the right thing.
If you develop in an environment where you have high velocity (eg python) you can much sooner learn that you are building the wrong thing and iterate.
Most systems do not have very high performance requirements. A typical python web application is not going to need to service many requests and the slow stuff it does is likely going to be waiting on responses from a database or other system. The thing to make such a program faster is to send better db queries rather than optimising the O(1 web page) work that is done in python.
I know this was an example, but if you get to that point and haven’t swapped out Python’s stdlib json library for orjson or some other performance-oriented library, that’s on you.
You also need to remember that back when those classic CS papers were written, the CPU/RAM speed ratios were entirely different from what they are today. Take, for instance, a Honeywell 316 from 1969 [0]: "Memory cycle time is 1.6 microseconds; an integer register-to-register "add" instruction takes 3.2 microseconds". Yep, back in those days, memory fetches could be twice as fast as the most simple arithmetic instruction. Nowadays, even the L1 fetch is 4 times as slow as addition (which takes a single cycle).
No wonder the classical complexity analysis of algorithms generally took memory access to be instantaneous: because it, essentially, was instantaneous.
It is generally better just to focus on algorithmic complexity - O(xN^k). The first version of the code should bring code to the lowest possible k (unless N is very small then who cares). Worry about x later. Don't even think about parallelizing until k is minimized. Vectorize before parallelizing.
For parallel code, you basically have to know in advance that it is needed. You can't normally just take a big stateful / mutable codebase and throw some cores at it.
Not sure I agree. Yes algorithmic complexity is important and should always be considered. But you shouldn't ignore X. Picking Rust instead of Python only changes X but it can easily reduce X by a factor of 100 or more, and it's not something you can realistically decide later.
Hmmm. I wasn't thinking about making performance decisions that early in a project - more like an optimization approach for a given feature / work.
In terms of deciding on what programming language to use for a project, that is a complicated one. Usually it is some combination of the team and the problem I suppose.
Choose an algorithm that's reasonably efficient and easy to implement. Then parallelize if it wasn't fast enough. Then try to find a better algorithm. And only then vectorize. Pick the low-hanging fruit first, and only resort to options that require more work and make the code less maintainable if that wasn't enough.
Software engineers and computer scientists are often reluctant to parallelize their code, because they have learned that it's difficult and full of hidden pitfalls. But then they lose the performance gains from multithreaded loops and other simple approaches, which are often essentially free. Multithreaded code is not particularly difficult or dangerous, as long as you are not trying to be clever.
- Knuth puts sentinels at the end of an array to avoid having to bounds check in the search.
- Knuth uses the register keyword.
- Knuth custom writes each data structure for each application.
When I was young someone pointed out to me how each hype cycle we cherry-pick a couple interesting bits out of the AI space, name them something else respectable, and then dump the rest of AI on the side of the road.
When I got a bit older I realized people were doing this with performance as well. We just call this part architecture, and that part Best Practices.
When Knuth wrote his paper compilers/optimizer were not nearly as good as today, and CPUs were much more deterministic (only disk access shows the issues we now see with cache misses). Most of his optimizations are things that the compiler will do better than you can (most is not all!) and so not even worth thinking about. Meanwhile the compiler will almost never turn a O(n^2) into O(n*ln(n)) despite that resulting in a much faster speedup than anything the compiler can do.
The famous "premature optimization" quote isn't from a dedicated paper on optimization, but from Knuth's 1974 "Structured Programming with go to Statements" paper where he was discussing broader programming methodology.
I happen to have also reread the paper last week. The last time I read it was 30 years ago, closer to when it was written than to the present, and I had forgotten a lot of it. One of the few bits that stuck with me was the Shanley Design Criterion. Another was "n and a half loops".
The bit about sequential search and optimization, the topic of this whole blog post, is kind of a minor detail in the paper, despite being so eloquently phrased that it's the part everyone quotes—sometimes without even knowing what “optimization” is. There's a table of contents on its second page, which is 33 lines long, of which "A Searching Example" and "Efficiency" are lines 4 and 5. They are on pages 266–269, 3 pages of a 41-page paper. (But efficiency is an issue he considers throughout.)
Mostly the paper is about control structures, and in particular how we can structure our (imperative!) programs to permit formal proofs of correctness. C either didn't exist yet or was only known to less than five people, so its break/continue control structures were not yet the default. Knuth talks about a number of other options that didn't end up being popular.
It was a really interesting reminder of how different things looked 51 years ago. Profilers had just been invented (by Dan Ingalls, apparently?) and were still not widely available. Compilers usually didn't do register allocation. Dynamically typed languages and functional programming existed, in the form of Lisp and APL, but were far outside the mainstream because they were so inefficient. You could reliably estimate a program's speed by counting machine instructions. People were sincerely advocating using loops that didn't allow "break", and subroutines without early "return", in the interest of building up their control flow algebraically. Knuth considered a recursive search solution to the N-queens problem to be interesting enough to mention it in CACM; similarly, he explains the tail-recursion optimization as if it's not novel but at least a bit recondite, requiring careful explanation.
He mentions COBOL, BCPL, BLISS, Algol W, Algol 60, Algol 68, other Algols, PL/I (in fact including some example code), Fortran, macro assemblers, "structured assemblers", "Wirth's Pascal language [97]", Lisp, a PDP-10 Algol compiler called SAIL (?), META-II, MIXAL, PL360, and something called XPL, but not Smalltalk, CLU, APL, FORTH, or BASIC.
He points out that it would be great for languages to bundle together the set of subroutines for operating on a particular data type, as Smalltalk and CLU did, but he doesn't mention CLU; it had only been introduced in the previous year. But he's clearly thinking along those lines (p.295):
> (...) it turns out that a given level of abstraction often involves several related routines and data definitions; for example, when we decide to represent a table in a certain way, we also want to specify the routines for storing and fetching data from that table. The next generation of languages will probably take into account such related routines.
Often when I read old papers, or old software documentation like the TENEX EXEC manual, it's obvious why the paths not taken were not taken; what we ended up doing is just obviously better. This paper is not like that. Most of the alternatives mentioned seem like they might have turned out just as well as what we ended up with.
> Yet we should not pass up our opportunities in that critical 3 %
The funny thing is, we forgot how to write programs that spend most of their time in 3% of the code.
Profile a modern application and you see 50 level deep stacks and tiny slices of 0.3% of CPU time spent here and there. And yet these slices amount to 100%.
I think the C# dev team had an interesting way to describe that kind of profile [0]:
> In every .NET release, there are a multitude of welcome PRs that make small improvements. These changes on their own typically don’t “move the needle,” don’t on their own make very measurable end-to-end changes. However, an allocation removed here, an unnecessary bounds check removed there, it all adds up. Constantly working to remove this “peanut butter,” as we often refer to it (a thin smearing of overhead across everything), helps improve the performance of the platform in the aggregate.
This is a bit different because C# is designed and implemented by industry experts - far above the average code monkey in terms of skill, understanding and experience, with added help from community experts, and the software in question is used by millions of systems all over the world.
Practically every part of C# is a hot path because it runs on probably billions of devices every minute of every day. This code is worth micro-optimizing. A 1% improvement in a popular C# library probably saves megawatts of electricity (not to mention the time saves) on a global scale.
Most code is not like that at all. Most code is not worth putting this kind of effort into. Most code is fine being 10% or even 500% slower than it technically could be because computers are lightning fast and can crunch numbers at an incredible speed. Even if the computer is doing 5 times as much work as it needs to it'll still be instant for most use cases. The problem for these kinds of applications arises when it's not doing 5 times as much work as it needs to but 5000 times or 5 million times. Or when the data is just really large and even the optimal solution would take a noticeable amount of time.
Most programs inherently don't have a small bit of code they spend most of their time in. There are exceptions like scientific computing, AI, etc. But most programs have their runtime distributed through lots of code so the idea that you can ignore performance except to profile and optimise some hotspots is nonsense.
Most programs - after dealing with bugs & low hanging fruit - don't have hotspots.
I'm not sure about the "most programs" bit. I think you're forgetting whole swathes of programs like system utilities, databases, tools etc.
But, in any case, this is why we build large programs using modular architectures. You don't profile the Linux kernel hoping to find a hotspot in some obscure driver somewhere, you profile the driver directly. Same for filesystems, networking etc. Similarly we build libraries for things like arithmetic and SAT solvers etc. that will be used everywhere and optimise them directly.
Your comment reads like you're disagreeing with your parent, but in fact you are reinforcing the point. We've forgotten how to build programs like this and end up with big balls of mud that can't be profiled properly.
I completely disagree. If the program doesn't have "hotspots" then it's probably pretty fast and it's following Knuth's advice. But in my experience most applications are completely littered with "hotspots" - endpoints taking 5+ seconds to respond, reports taking minutes or hours to generate, stuff like that. I've never worked on an application where I didn't quickly find multiple easily avoidable hotspots. The average developer, at least in my area of the world, just really doesn't understand how to write performant code at all. They write algorithms with terrible time and/or space complexity, they write inefficient database queries, design their database schemas in ways that don't even allow efficient queries etc.
Then they add caching to compensate for their crappy system design and stuff like that, bloating the system with lots of unnecessary code. They store the data naively then spend significant processing time preparing the data for delivery rather than just storing the data in an ideal format in the first place. Lots of stuff like that.
Pick your top 3 and optimize if you really have to. But if the whole chain is 0.3% here and there, theres very little room for large optimization wins without some redesign.
Wait... people understood "premature optimisation" to mean "small optimisations are not worth it"? I've always understood it to mean exactly what it's supposed to mean, namely don't optimise something until you've shown that it's actually needed. I honestly don't know how it could be interpreted any other way.
It is sad how often people cite premature optimization in the real world even when it isn't premature. Using a good algorithm is never premature optimization, particularly when the good algorithm is likely built into your language as well. Likewise if you have run the profiler it is not premature optimization, and running the profiler to find those places is not premature optimization.
I mean basically what he's saying is check the impact of your optimization. Every time there's an optimization, theres a complexity and brittleness cost. However sometimes unoptimized code is actually more difficult to read than the optimized version. I mean its quite logical tbf.
I understand the Knuth's _Premature Optimization_ saying as: (1) find the most hot code (best way to have full heat chart), (2) optimize there. That's all of it, and it really works. E.g. If you have some code that prepares data for a loop that is n^2, and you work hard on this data preparation, profile it etc., this will not give you any visible improvement, because this is one-time code. What's inside the n^2 loop is more important and optimize there, even work together and prepare the data better to allow the loop work more efficient.
In practice repetition of this quote more often than not leads to a lazy attitude of "don't think about performance, it's too complicated, measure first and then let's talk".
In my experience all great programmers think about performance from the moment they start thinking about a problem.
Like, you're writing an O(n^2) nested loop over an array that's not going to be tiny? Get out of here! That just shouldn't happen and talking about "premature optimization" points in exactly the wrong direction.
I think the best way to go about optimization is to maintain a constant baseline level of mechanical sympathy with the computer's architecture.
If you burn into your soul the table of NUMA latencies and really accept the inevitable conclusions regarding practical computation, a lot of the optimization discussion simply boils down to keeping the closest cache as hot as possible as often as possible. Put differently, if you can get the working set to ~fit in L1, then you have nothing to worry about. The CPU is faster than most people can reason about now. You can talk to L1 1000x before you can talk to the GPU 1x. This stuff makes a ridiculous difference when applied correctly.
Context switching and communication between threads is where most developers begin to lose the rabbit. Very often a problem can be solved on one physical core much faster than if you force it to be solved across many physical cores.
(meta) I am probably wasting my time commenting on linked article here. Nobody does that /s.
I don't think the measurements support conclusion that well.
What I want to have when I see those measurements:
I want language abstract machine and compiler to not get in the way I want code on certain platforms to perform. This is currently not what I get at all. The language is actively working against me. For example, I can not read cache line from an address because my object may not span long enough. The compiler has its own direction at best. This means there is no way to specify things, and all I can do is test compiler output after each minor update. In a multi-year project such updates can happen dozens of times! The ergonomics of trying to specify things actually getting worse. The example with assembly is similar to my other experiences: the compiler ignores even intrinsics now. If it wants to optimize, it does.
I can't run to someone else for magic library solutions every time I need to write code. I need to be able to get things done in a reasonable amount of time. It is my organization development process that should decide if the solution I used should be part of some library or not. It usually means that efforts that cover only some platforms and only some libraries are not that universally applicable to my platforms and my libraries as folks at language conferences think /s.
It's also good practice to chose an appropriate language for the problem. It's not premature optimisation to use a compiled language instead of Python when you already know the code will run thousands of times in a loop because of the application and use a lot of electrical energy in the process.
72 comments
[ 3.3 ms ] story [ 73.1 ms ] threadAlso there's Knuth admitting he avoids GO TO because he is afraid of being scolded by Edsger Dijkstra.
https://pic.plover.com/knuth-GOTO.pdf
From the paper:
"It is clearly better to write programs in a language that reveals the control structure, even if we are intimately conscious of the hardware at each step; and therefore I will be discussing a structured assembly language called PL/MIX in the fifth volume of The art of computer programming"
Looking forward to that!
This statement in the introduction applies to so many things in CS:
Like Brooks said: No Silver BulletsI personally find that "break/continue" with (optionally labelled) loops cover about 95% of GOTO's use cases today (including "one-and-a-half" loops), although e.g. Rust and Zig, with their expression-oriented design (and support for sum types) offer an option to experiment with Zahn's event mechanism... but usually factoring an inner loop body into a separate function is a more clear-to-read approach (as for efficiency? Hopefully you have a decent inliner idk).
The only thing I truly miss, occasionally, is a way to concisely write an one-and-a-half range/FOR loop. Something like this:
When you have a generic "while True:" loop, conditional "break" works fine enough, but here? Ugh.I have made the same point several times online and in person that the famous quote is misunderstood and often suggest people take the time to go back to the source and read it since it’s a wonderful read.
Things have changed.
Remember this: "speed is a feature"?
If you need fast softwqare to make it appealing then make it fast.
Somehow I found 40%. Some of it was clever, a lot of it was paying closer attention to the numbers, but most of it was grunt work.
Besides the mechanical sympathy, the other two main tools were 1) stubbornness, and 2) figuring out how to group changes along functional testing boundaries, so that you can justify making someone test a change that only improves perf by 1.2% because they're testing a raft of related changes that add up to more than 10%.
Most code has orphaned performance improvements in 100 little places that all account for half of the runtime because nobody can ever justify going in and fixing them. And those can also make parallelism seem unpalatable due to Amdahl.
@OP, I think you should lead with this. I think it gets lost by the time you actually reference it. If I can suggest, place the second paragraph after
The optimization part gets lost in the middle and this, I think, could help provide a better hook to those who aren't going to read the whole thing. Which I think how you wrote works good for that but the point (IMO) will be missed by more inattentive readers. The post is good also, so this is just a minor critique because I want to see it do better.https://dl.acm.org/doi/10.1145/356635.356640 (alt) https://sci-hub.se/10.1145/356635.356640
No amount of parallelization will make your program faster than the slowest non-parallelizable path. You can be as clever as you want and it won’t matter squat unless you fix the bottleneck.
This extends to all types of optimization and even teamwork. Just make the slowest part faster. Really.
https://en.wikipedia.org/wiki/Amdahl%27s_law
"Curiosity killed the cat, but satisfaction brought it back." Is practically on the same level.
If you're careful to exclude creeping featurism and architectural astronautics from the definition of 'optimization', then very few people I've seen be warned off of digging into that sort of work actually needed to be reined in. YAGNI covers a lot of those situations, and generally with fewer false positives. Still false positives though. In large part because people disagree on what "The last responsible moment" in part because our estimates are always off by 2x, so by the time we agree to work on things we've waited about twice as long as we should have and now it's all half assed. Irresponsible.
Tuning the JSON parser is not nearly as effective as replacing it with something less stupid, which is, in turn, not nearly as effective as finding ways to not do the RPC at all.
Most really performant systems of a certain age are also full of layer violations for this reason. If you care about performance (as in you are paid for it), you learn these realities pretty quickly. You also focus on retaining optionality for future optimizations by not designing the semantics of your system in a way that locks you permanently into low performance, which is unfortunately very common.
Another example of this is using the right algorithm for the job. I see a lot of code like let's say you have two lists and you need to go through the first one and find the corresponding element by id in the second. The naive implementation uses linear search of the second list, making the algorithm O(n^2). This will stall and take hours, days or longer to run when the lists get large, say into the tens to hundreds of thousands of elements.
If both lists come from your own database then you should have used a foreign key constraint and had the database join them for you. If you can't do that, let's say one or both lists come from a third-party api, you can create a dictionary(hashmap) of the second list to allow O(1) lookup rather than O(n) which brings your full algorithm's complexity to O(n). Now the you can have millions of elements and it'll still run in milliseconds or seconds. And if this is something you need to do very often then consider storing the second list in your database so you can use a join instead. Avoid doing the work, as you say.
These are the most common mistakes I see people make. That and situations where you make one request, wait for a response, then use that to make another request, wait for response and so on.
You don't need a profiler to avoid this type of problem, you just need to understand what you're doing at a fundamental level. Just use a dictionary by default, it doesn't matter if it might be slightly slower for very small inputs - it'll be fast enough and it'll scale.
As with that Google testing talk ("We said all your tests are bad, but that's not entirely true because most of you don't have any tests"), the reality is that most people don't have any profiling.
If you don't have any profiling then in fact every optimisation is premature.
Most people I see who get offended are reacting to the claim that optimization is never useful. But it's pretty easy to knock that claim over.
I don't deny that plenty of people use adjectives very sloppily, and that much writing is improved by just ignoring them, but Knuth is not one of them.
BTW, my junior developers don't know what a debugger is.
They are all worth reading.
beware too of premature pessimization. Don't write bubble sort just because you haven't benchmarked your code to show it is a bottleneck - which is what some will do and then incorrectly cite premature optimization when you tell them they should do better. Note that any compitenet languare has sort in the standard library that is better than bubble sort.
You may also want to look at the version published in Knuth's collection called Literate Programming (with corresponding errata: https://cs.stanford.edu/~knuth/lp.html#:~:text=Errata,correc... ), and I've just uploaded a scan here: https://shreevatsa.net/tmp/2025-06/DEK-P67-Structured.progra...
In my experience the latter is actually often expressed. What else would “premature” mean, other than you don’t know yet whether the optimization is worth it?
The disagreement is usually more about small inefficiencies that may compound in the large but whose combined effects are difficult to assess, compiler/platform/environment-dependent optimizations that may be pessimizations elsewhere, reasoning about asymptotic runtime (which shouldn’t require benchmarking — but with cache locality effects sometimes it does), the validity of microbenchmarks, and so on.
- argue against thinking about any kind of design choice for performance reasons, eg the data structure decisions suggested in this article
- argue for a ‘fix it later’ approach to systems design. I think for lots of systems you have some ideas for how you would like them to perform, and you could, if you thought about it, often tell that some designs would never meet them, but instead you go ahead with some simple idea that handles the semantics without the performance only to discover that it is very hard to ‘optimise’ later.
But what I think people don't realize is that this is exactly what tech debt is. You're moving fast but doing so makes you slow once we are no longer working in a very short timeline. That's because these issues compound. Not only do we repeat that same mistake, but we're building on top of shaky ground. So to go back and fix things ends up requiring far more effort than it would have taken to fix it early. Which by fixing early your efforts similarly compound, but this time benefiting you.
I think a good example of this is when you see people rewrite a codebase. You'll see headlines like "by switching to rust we got a 500% improvement!" Most of that isn't rust, most of that is better algorithms and design.
Of course, you can't always write your best code. There's practical constraints and no code can be perfect. But I think Knuth's advice still fits today, despite a very different audience. He was talking to people who were too obsessed with optimization while today were overly obsessed with quickly getting to some checkpoint. But the advice is the same "use a fucking profiler". That's how you find the balance and know what actually can be put off till later. It's the only way you can do this in an informed way. Yet, when was the last time you saw someone pull out a profiler? I'm betting the vast majority of HN users can't remember and I'd wager a good number never have
I realize this is a very personal preference and it obviously can't be applied to everyone. Someone with less understanding might find a profiler very useful and I think those people will learn the same things I'm talking about - as you find the slow code and learn how to make it fast you'll stop making the same mistakes.
A profiler might be useful if I was specifically working to optimize some code, especially code I hadn't written myself. But for my daily work it's almost always good enough to keep performance in mind and design the system to be fast enough from the bottom up.
Most code doesn't have to be anywhere near optimal, it just has to be reasonably fast so that users don't have to sit and stare at loading spinners for seconds at a time. Some times that's unavoidable, some times you're crunching huge amounts of data or something like that. But most of the time, slow systems are slow because the people who designed and implemented them didn't understand what they were doing.
Notably, pretty much the entire body of discourse around structured programming is totally lost on modern programmers failing to even imagine the contrasts.
The biggest way I see this is picking an architecture or programming language (cough Python) that is inherently slow. "We'll worry about performance later" they say, or frequently "it's not performance critical".
Cut to two years later, you have 200k lines of Python that spends 20 minutes loading a 10GB JSON file.
If you develop in an environment where you have high velocity (eg python) you can much sooner learn that you are building the wrong thing and iterate.
Most systems do not have very high performance requirements. A typical python web application is not going to need to service many requests and the slow stuff it does is likely going to be waiting on responses from a database or other system. The thing to make such a program faster is to send better db queries rather than optimising the O(1 web page) work that is done in python.
No wonder the classical complexity analysis of algorithms generally took memory access to be instantaneous: because it, essentially, was instantaneous.
[0] https://en.wikipedia.org/wiki/Honeywell_316#Hardware_descrip...
For parallel code, you basically have to know in advance that it is needed. You can't normally just take a big stateful / mutable codebase and throw some cores at it.
In terms of deciding on what programming language to use for a project, that is a complicated one. Usually it is some combination of the team and the problem I suppose.
Software engineers and computer scientists are often reluctant to parallelize their code, because they have learned that it's difficult and full of hidden pitfalls. But then they lose the performance gains from multithreaded loops and other simple approaches, which are often essentially free. Multithreaded code is not particularly difficult or dangerous, as long as you are not trying to be clever.
When I got a bit older I realized people were doing this with performance as well. We just call this part architecture, and that part Best Practices.
The bit about sequential search and optimization, the topic of this whole blog post, is kind of a minor detail in the paper, despite being so eloquently phrased that it's the part everyone quotes—sometimes without even knowing what “optimization” is. There's a table of contents on its second page, which is 33 lines long, of which "A Searching Example" and "Efficiency" are lines 4 and 5. They are on pages 266–269, 3 pages of a 41-page paper. (But efficiency is an issue he considers throughout.)
Mostly the paper is about control structures, and in particular how we can structure our (imperative!) programs to permit formal proofs of correctness. C either didn't exist yet or was only known to less than five people, so its break/continue control structures were not yet the default. Knuth talks about a number of other options that didn't end up being popular.
It was a really interesting reminder of how different things looked 51 years ago. Profilers had just been invented (by Dan Ingalls, apparently?) and were still not widely available. Compilers usually didn't do register allocation. Dynamically typed languages and functional programming existed, in the form of Lisp and APL, but were far outside the mainstream because they were so inefficient. You could reliably estimate a program's speed by counting machine instructions. People were sincerely advocating using loops that didn't allow "break", and subroutines without early "return", in the interest of building up their control flow algebraically. Knuth considered a recursive search solution to the N-queens problem to be interesting enough to mention it in CACM; similarly, he explains the tail-recursion optimization as if it's not novel but at least a bit recondite, requiring careful explanation.
He mentions COBOL, BCPL, BLISS, Algol W, Algol 60, Algol 68, other Algols, PL/I (in fact including some example code), Fortran, macro assemblers, "structured assemblers", "Wirth's Pascal language [97]", Lisp, a PDP-10 Algol compiler called SAIL (?), META-II, MIXAL, PL360, and something called XPL, but not Smalltalk, CLU, APL, FORTH, or BASIC.
He points out that it would be great for languages to bundle together the set of subroutines for operating on a particular data type, as Smalltalk and CLU did, but he doesn't mention CLU; it had only been introduced in the previous year. But he's clearly thinking along those lines (p.295):
> (...) it turns out that a given level of abstraction often involves several related routines and data definitions; for example, when we decide to represent a table in a certain way, we also want to specify the routines for storing and fetching data from that table. The next generation of languages will probably take into account such related routines.
Often when I read old papers, or old software documentation like the TENEX EXEC manual, it's obvious why the paths not taken were not taken; what we ended up doing is just obviously better. This paper is not like that. Most of the alternatives mentioned seem like they might have turned out just as well as what we ended up with.
The funny thing is, we forgot how to write programs that spend most of their time in 3% of the code.
Profile a modern application and you see 50 level deep stacks and tiny slices of 0.3% of CPU time spent here and there. And yet these slices amount to 100%.
> In every .NET release, there are a multitude of welcome PRs that make small improvements. These changes on their own typically don’t “move the needle,” don’t on their own make very measurable end-to-end changes. However, an allocation removed here, an unnecessary bounds check removed there, it all adds up. Constantly working to remove this “peanut butter,” as we often refer to it (a thin smearing of overhead across everything), helps improve the performance of the platform in the aggregate.
[0]: https://devblogs.microsoft.com/dotnet/performance-improvemen...
Practically every part of C# is a hot path because it runs on probably billions of devices every minute of every day. This code is worth micro-optimizing. A 1% improvement in a popular C# library probably saves megawatts of electricity (not to mention the time saves) on a global scale.
Most code is not like that at all. Most code is not worth putting this kind of effort into. Most code is fine being 10% or even 500% slower than it technically could be because computers are lightning fast and can crunch numbers at an incredible speed. Even if the computer is doing 5 times as much work as it needs to it'll still be instant for most use cases. The problem for these kinds of applications arises when it's not doing 5 times as much work as it needs to but 5000 times or 5 million times. Or when the data is just really large and even the optimal solution would take a noticeable amount of time.
Most programs - after dealing with bugs & low hanging fruit - don't have hotspots.
But, in any case, this is why we build large programs using modular architectures. You don't profile the Linux kernel hoping to find a hotspot in some obscure driver somewhere, you profile the driver directly. Same for filesystems, networking etc. Similarly we build libraries for things like arithmetic and SAT solvers etc. that will be used everywhere and optimise them directly.
Your comment reads like you're disagreeing with your parent, but in fact you are reinforcing the point. We've forgotten how to build programs like this and end up with big balls of mud that can't be profiled properly.
Then they add caching to compensate for their crappy system design and stuff like that, bloating the system with lots of unnecessary code. They store the data naively then spend significant processing time preparing the data for delivery rather than just storing the data in an ideal format in the first place. Lots of stuff like that.
On a long enough timeline, the probability that someone will call your code in a for-loop approaches 1.
In my experience all great programmers think about performance from the moment they start thinking about a problem.
Like, you're writing an O(n^2) nested loop over an array that's not going to be tiny? Get out of here! That just shouldn't happen and talking about "premature optimization" points in exactly the wrong direction.
If you burn into your soul the table of NUMA latencies and really accept the inevitable conclusions regarding practical computation, a lot of the optimization discussion simply boils down to keeping the closest cache as hot as possible as often as possible. Put differently, if you can get the working set to ~fit in L1, then you have nothing to worry about. The CPU is faster than most people can reason about now. You can talk to L1 1000x before you can talk to the GPU 1x. This stuff makes a ridiculous difference when applied correctly.
Context switching and communication between threads is where most developers begin to lose the rabbit. Very often a problem can be solved on one physical core much faster than if you force it to be solved across many physical cores.
I don't think the measurements support conclusion that well.
What I want to have when I see those measurements:
I want language abstract machine and compiler to not get in the way I want code on certain platforms to perform. This is currently not what I get at all. The language is actively working against me. For example, I can not read cache line from an address because my object may not span long enough. The compiler has its own direction at best. This means there is no way to specify things, and all I can do is test compiler output after each minor update. In a multi-year project such updates can happen dozens of times! The ergonomics of trying to specify things actually getting worse. The example with assembly is similar to my other experiences: the compiler ignores even intrinsics now. If it wants to optimize, it does.
I can't run to someone else for magic library solutions every time I need to write code. I need to be able to get things done in a reasonable amount of time. It is my organization development process that should decide if the solution I used should be part of some library or not. It usually means that efforts that cover only some platforms and only some libraries are not that universally applicable to my platforms and my libraries as folks at language conferences think /s.
Disclaimer. I work in gamedev.