9 ms·

You may be significantly underestimating your ability to get phenomenal gains from algorithmic improvements. I think DJB would argue that the compiler optimizat

• DannyBee 9y ago
  " I think DJB would argue that the compiler optimizations are insignificant in comparison to those improvements for a very large portion of hot code." Except he has no data to show this, and every piece of data i have (and my colleagues have), says that he is wrong. Additionally, the compilers can and do replace algorithms if users let them. Most users don't want it, even if it makes code faster. Remember that the vast majority of people want software that works and is reliable, or has fast development cycles, or a million other things, and happily choose that over software that is fast. Compilers changing algorithms on you isn't highly compatible with this, especially when, for example, most programming languages don't even give you a good way to express the invariants you are trying to maintain (except good ol' Ada). However, besides that, certainly you can see that in the end, the people lose. It's like arguing that computer would never win at go. If i really really really cared about some piece of hot code, i wouldn't pay an expert to optimize it, i'd throw the cycles of 100k spare cpus at optimizing it. That is also the other strange part of his argument to me, he assumes the capabilities of compilers based on compilers he sees that are meant to operate pretty much single threaded on single machines, and complains "they will never beat people". It's like looking at the machine learning algorithm that takes months or years or roughly forever to train on a single cpu and say "it'll never do a good job". The world is not that small anymore. It's blindingly obvious the optimizing compilers win if we want them to. In any case, this particular debate will restart again as gpus and accelerators follow the same old compiler cycle.
  • oconnor0 9y ago
    > Additionally, the compilers can and do replace algorithms if users let them. Most users don't want it, even if it makes code faster. What compilers do this? With what algorithms?
    • jerven 9y ago
      If you look at modern JIT compilers such as Truffle-Ruby, then yes it will change algorithms. e.g. changing sort algorithms depending on data size. if the data is small enough it will do a bubblesort on registers, when its big it will do TimSort on heap data etc...[1] [1]: http://chrisseaton.com/rubytruffle/small-data-structures/ http://chrisseaton.com/rubytruffle/small-data-structures/
      • coldtea 9y ago
        This example is about the compiler changing between (existing) internal versions of sorting algorithms for language data structures. Not about changing what algorithms the user coded.
        • danielbarla 9y ago
          Exactly, that's two completely different things. It's akin to a compiler using the famous "switch statement becomes if-else chain if less than 7 items, otherwise lookup table". The parent posts are referring to macro-level optimisations, where large changes to how a system works are implemented, in order to get massive gains in performance. Picking out the essence of the larger implementation would be a difficult task for a compiler, but more importantly it wouldn't have enough information about the context of the application to make optimisation decisions. Things like: do we need to read from that file each time we need this piece information, or is it a once-off that can be cached? Unless this contextual information is expressed somehow (and it basically never is), only the programmer will be able to make the change.
          • jerven 9y ago
            def clamp(num, min, max) [min, num, max].sort[1] end being turned into def clamp(num, min, max) if (num < min) { if (num < max) return num; else return max; } if (min < max) return min; else return max; end Seems to me to be algorithm rewriting. As is superword optimization and partial evaluation. All things that Truffle-Ruby + Graal do today.
  • flamedoge 9y ago
    > the compilers can and do replace algorithms if users let them I think idiom recognizers are worthy of mention. That said, it being a manual pattern recognition is probably going to eventually lose edge to things like LLVM souper or other superoptimizers (exhaustive or predictive)
  • coldtea 9y ago
    >Except he has no data to show this, and every piece of data i have (and my colleagues have), says that he is wrong. That's because those data only represent quantitative statistics on code, not whether said code would get great benefits if the programmer actually bothered to make algorithmic adjustments.
    • DannyBee 9y ago
      Why do you assume we haven't measured what optimal looks like?
  • flukus 9y ago
    > Remember that the vast majority of people want software that works and is reliable, or has fast development cycles, or a million other things, and happily choose that over software that is fast. People want software that is fast, slow (and bloated) software frustrates normal people, they just can't articulate it or identify the correct source of their frustrations. As I said in another post just yesterday, you'll often hear people say everything from "god I hate Microsoft" to "i think I need a new machine" to "I might have a virus". Once you dig into the problem it's often just a particular app (or bloated website) that is slow, or something else on the machine eating up way to much RAM and causing everything to swap. I'd love to run an experiment where group A has native apps written in a low level language and group B has entirely electron based ones. I'd bet my house that group B has a hell of a lot more general complaints.
  • mpweiher 9y ago
    >> compiler optimizations are insignificant in comparison to those improvements > he has no data to show this Yes he does. And so do countless of other people. For example, I just gave a talk where I describe an afternoon optimization session giving me a 1000x and and 8000x improvement for a data-conversion and a query task, respectively. https://www.youtube.com/watch?v=kHG_zw75SjE https://www.youtube.com/watch?v=kHG_zw75SjE Was the compiler helpful? Probably, I didn't measure with and without. But the vast majority of the improvement was not the compiler. And I don't see how the compiler could have come up with the changes (moving from a layered architecture of Model Objects -> ORM -> SQLite to a hexagonal architecture with an optimized data representation and completely different access procedure for the one data type that mattered). And this is what I find almost everywhere I go. For example the BBC system I made somewhere between a hundred and a thousand times faster: https://www.slideshare.net/MarcelWeiher/in-processrest https://www.slideshare.net/MarcelWeiher/in-processrest , https://link.springer.com/chapter/10.1007/978-1-4614-9299-3_11 https://link.springer.com/chapter/10.1007/978-1-4614-9299-3_... And so on and so on. It is also interesting that in none of these cases did I revert to assembler, so I think that part of DJB's thesis is not entirely correct, but only in being too specific: hand-coding assembler is just one part of hand-optimizing, and very often it is not the or even a crucial part (though in security algorithms it might very well be). But the idea that optimizing compilers are getting squeezed at both ends is, IMHO, 100% correct. On the one hand, so much code is so incredibly cold, see Ruby on Rails web sites or most UI programming. RoR can do maybe 10-100 requests/s. It's so slow it's comical. However, most sites would be happy to receive 100 hits per day, so "comically slow" is fast enough. Or when I set the end-parameters for an animation that is going to take half a second, it wouldn't matter if the code doing that were run on a LISP interpreter badly written in Ruby. Or why hasn't PyPy completely replaced Python? Why was TCL adequate for steering supercomputer workloads? On the other hand, the performance critical parts have gotten more critical. That means that the "make everything a bit faster, but without any guarantees" that optimizing compilers are great at providing is simply not that useful any more for a lot of workloads. And yes, there are workloads where it is useful, and it seems that you have some of these workloads. That doesn't change the broader picture.
    • regularfry 9y ago
  • big_spammer 9y ago
    > most programming languages don't even give you a good way to express the invariants you are trying to maintain (except good ol' Ada). What does Ada do that helps with this in ways C and other languages don't?
    • nostrademons 9y ago
      Ada has a rich set of language features to express preconditions, postconditions, and invariants directly in the code, and then have those logical statements available to the compiler so that it can assume certain facts about the program's execution when optimizing: https://en.wikibooks.org/wiki/Ada_Programming/Contract_Based_Programming https://en.wikibooks.org/wiki/Ada_Programming/Contract_Based... Also, the type system lets you specify legal values with more precision than the word-based types that most languages inherited from C. For example, you can have range types that may hold integers of only a certain range, and the compiler will enforce that when assigning to variables of the type, or throw an exception if an arithmetic operation results in a value outside the range (you also have modulus types that explicitly give the wraparound behavior familiar in C). Containers allow generic constraints, and you can explicitly specify the dimensions of array types. https://en.wikibooks.org/wiki/Ada_Programming/Type_System https://en.wikibooks.org/wiki/Ada_Programming/Type_System
• oconnor0 9y ago
  I think one of the issues here is what compiler optimizations encompass. In a language like C or C++, you may be able to write optimal algorithms but end up with poor performance because of memory access. An unoptimized compile puts almost everything on the stack - a = b + c involves two loads, one add, and one store in an unoptimized compile. All of that extra memory access is going to kill performance; even if you have optimal data structures and algorithms. I see "optimizations are bad/unnecessary/problematic/whatever" in my job writing a compiler. The underlying issue is almost never "I don't want optimizations" - even when someone thinks that's what they want; it's "I don't want optimizations that produce unexpected behavior".
• dagw 9y ago
  You may be significantly underestimating your ability to get phenomenal gains from algorithmic improvements. My HPC lecturer used to say that the speed difference between bad Matlab code and good Matlab code will often be greater than the speed difference between good Matlab code and good Fortran code. Tweaking the hell out of your code at the assembly level is a waste of time if you're using a naive O(2^N) algorithm.
  • gpderetta 9y ago
    ... unless N is small.
  • sllabres 9y ago
    You mean O(n^2). 2^n looks a more than 'naive' for most cases.
• dom0 9y ago
  > You may be significantly underestimating your ability to get phenomenal gains from algorithmic improvements. Implementation differences in the exact same, say, simple linear algorithm, can still add up to a tenfold difference. Compiler optimisations can take it further, though some are not generally applicable. E.g. I have made good experiences with PGO (careful analysis and/or knowing your workload is needed), but it's obviously not an option for the typical open source software distribution model. Obviously, combining a better algorithm with a better implementation that is optimized will be fastest.
• Sean1708 9y ago
  > You may be significantly underestimating your ability to get phenomenal gains from algorithmic improvements. Absolutely but often it's very hard to make those changes or those changes lead to a loss of accuracy, whereas compiling with -O3 is very easy and leads to no loss in accuracy (in the common case).