Thinking in an array language

4 years ago (github.com)

Any performance benchmarks between a simple C implementation of matrix multiplication like explained in the article and K one?

Story time: 2 years ago, right before 1st lockdown, I landed a client. A data scientist who had already implemented an algorithm which dealt with matrices, implemented by a previous programmer he hired. Said implementation was in Python, no more than 10 line altogether, which performed well when matrix size where small, like 10x10. But the problem was, his real work need it to have matrices of size 10^6 x 10^6. Not only the Python implementation had to be ran on a beast of server with 4TB of memory, it also took 3 days to finish. And while the algorithm was small in size in Python, its explaining paper was 4 pages in total, which took me 1 week to understand. And then an entire month to implement in C. But in the end, when all was said and done, the run time when used with real data was only 20 minutes and consumed only 8 GB of memory, though it did required at least 16 virtual processors.

Hence my question, in the end performance is what it matters, not number of lines.

  • K timing, squaring a 100x100 matrix in ngn/k:

       a:100 0N#?10000    / 100x100 double matrix
       \t:1000 a(+/*)\:a  / Total of 1k runs in ms
      1514
    

    Following C program outputs 272 when compiled and run with -O3 -march=native.

      #include <stdio.h>
      #include <time.h>
      size_t monoclock(void) {
        struct timespec ts;
        clock_gettime(CLOCK_MONOTONIC, &ts);
        return 1000000000*ts.tv_sec + ts.tv_nsec;
      }
      
      int main() {
        const size_t n = 100;
        double a[n][n], r[n][n];
        size_t t = monoclock();
        for (size_t iter=0; iter<1000; iter++) {
          for (size_t i=0; i<n; i++) {
            for (size_t j=0; j<n; j++) {
              double sum = 0;
              for (size_t k=0; k<n; k++) sum += a[i][k]*a[k][j];
              r[i][j] = sum;
            }
          }
        }
        printf("%lld\n", (monoclock() - t)/1000000);
      }
    

    So K is slower, by a factor of 5 rather than the 200 you saw with Python. K appears to scale to larger matrices better than C: 8664 in K vs 2588 in C for 200x200, but I can't be bothered to sort out heap allocation for the C code to do larger sizes. I would certainly not say that your story supports the idea that performance is what matters. In the month you took to implement the C version, the Python solution could have run ten times over, and there are many computations that only ever need to be performed once. Besides, I bet you were glad to have working Python code to use as a reference!

    • I prefer "middle of the road" languages that are high-level AND readable AND have decent performance optimisation for bulk operations. Python with C libraries suffices for a lot of people, Julia similarly is getting popular.

      Even the older Mathematica language blows both K and C out of the water for readability and performance:

          m = Table[RandomReal[], 100, 100];
          t = RepeatedTiming[MatrixPower[m, 2]];
          First[t]*1000*1000
      
          18.0245 
      

      You would have to know literally zero about Mathematica's Wolfram Language to be able to read that clearly! From a standing start you could understand what another person has created. For K, you'd have to have memorized the K-specific syntax. For C, if you hadn't seen the standard patterns for matrix multiplication you'd have to read the code carefully. A lot of it is just noise, like the verbose for loop syntax.

      Oh, and Mathematica's matrix power function is:

      - Parallel! If I use a 10K x 10K matrix as an input, it uses about 75% CPU on my 8-core laptop. It can complete a single multiplication in 5.3 seconds. For laughs, try that with either K or C and see what you get...

      - Extends to negative or fraction powers.

      - Has optimisations for applying the matrix power directly to a vector.

      - Is extensively documented, unlike the terse K snippet or the hand-rolled C code: https://reference.wolfram.com/language/ref/MatrixPower.html....

      Essentially what I'm trying to say is that terseness is an anti-pattern, and doesn't even begin to approach the utility of a well designed high-level language intended for teams of collaborating humans.

    • When you compile the C example with the compiler option "-Wall", you'll get a warning that the variable 'r' is set but not used, so the compiler would be free to simply skip the for loops. In fact, if you compile with clang instead of gcc, the compiler will do just that and you'll get almost zero computation time.

      It would be better to do something with the computed result so the compiler does not remove the computation, e.g. print a randomly selected value.

      I also benchmarked the fixed C code against Python and "Python" (which uses OpenBLAS under the hood in my case) was 10 times faster:

          import time
          import numpy as np
      
          a = np.random.rand(100, 100)
          t = time.perf_counter()
      
          for _ in range(1000): a @ a
      
          print(time.perf_counter() - t, "seconds")
      

      Implementation matters a lot for matrix multiplication.

      1 reply →

    • Naive implementations of stock matrix math can't get anywhere close to numpy or julia, which both use BLAS and automatically parallelize across cores.

        % python matrix.py
        Timing 10 squares of a random 10000 x 10000 matrix
        97.3976636590669 seconds
        python matrix.py  364.41s user 8.10s system 379% cpu 1:38.25 total
      

      julia has more overhead, and the first multiply triggers code compilation so there's an additional warm-up square outside of the timing loop, but its "warm" performance is equivalent to numpy. Turning on extra optimizations (-O3) can even make it a couple seconds faster than numpy once warmed up.

        % julia matrix.jl
        Timing 10 squares of a random 10000 x 10000 matrix
         97.787679 seconds (31 allocations: 7.451 GiB, 0.33% gc time)
        julia matrix.jl  405.34s user 8.13s system 375% cpu 1:50.09 total
      

      If you're going to wait for that C implementation, or the other comment's K implementation, to finish that loop, you'll want a book.

    • Nice benchmarking, thank you for your effort.

      Also the scientist was leading a team, so my program would've been used by at least 20 people, 10 times per day. That was why Python one was a no go for them from beginning.

      4 replies →

I was just playing with Nils M Holm's Klong this morning: https://t3x.org/klong/index.html (Klong rather than the others mostly because the C implementation looks like C so I have a ghost of a chance of actually grokking it.)

These folks are really onto something, but I think they get sidetracked in the (admittedly very very fun) minutia of the languages and lose sight of the crucial insight in re: mathematical notation, to wit: it's a means of human communication.

For APL or K to get out of their niches would require, I am convinced, something like a tome of documentation of a ratio of about 1.5 paragraphs per line of code. That would give us mere mortals a fighting chance at grokking these tools.

A similar problem plagues the higher-order stuff they're pursuing over in Haskell land. I know e.g. "Functional programming with bananas, lenses, envelopes and barbed wire" and "Compiling to Categories" are really important and useful, but I can't actually use them unless some brave Prometheus scales Olympus and returns with the fire.

Stuff dribbles out eventually. Type inference and checking have finally made it into the mainstream after how many decades?

Array languages might make good maths notation. It's terse and easy to write, and there's a logical naming scheme (for instance, matrix multiplication is just (+/*)\: ). I suppose the trick is to think of (+/*)\: as one unit.

"What if we could make entire programs as unreadable as regexes?" -K

  • And yet regular expressions are a part of most programming languages as a library or built-in syntax.

    Why is that?

    • Probably due to they're a great notation for the problem area which for regex is concisely describing text patterns.

      For example 'a*b' is any number of 'a's followed by 'b'.

      How else would you concisely state that?

      6 replies →

    • They're very quick to write and they (in appropriate cases) would be quite difficult to implement otherwise (tedious state machine stuff). They are still massively overused though. Using a regex at all is a huge red flag. Sometimes they are appropriate, but not in 90% of cases in my experience.

      Anyway I'm not sure the same is true for K. At least for the given example the for loop was not exactly difficult to write.

Array notation is great, but only for single array operations and for dense array operations.

The moment you need to run complex group bys and store non-contiguous data efficiently, it gets awkward pretty quick.

On the other hand, operations on dense data is pretty cumbersome in SQL. You can only do so much with limited support of proper scan algorithms, merge joins or bolted on window functions.

Please somebody combine APL with SQL and you win the programming language wars.

  • > Please somebody combine APL with SQL and you win the programming language wars.

    kdb+ and q fit this description. Docs here: https://code.kx.com/q/basics/qsql/

    Here's an example.

    In SQL:

        SELECT stock, SUM(amount) AS total FROM trade GROUP BY stock
    

    In q:

        q)select total:sum amt by stock from trade

    • Was gonna say the same plus kdb+ is a columnar store so you can get vectorizarion in your sql execution as well.

  • I would say APL-family languages today have largely addressed these concerns with operators such as Key[0][1] and nested arrays. J also has built-in support for sparse arrays. Some more complicated things like storing a list of lists in a compact representation (perhaps lengths and data) aren't supported natively, but I'd consider that a niche concern as a list of lists will have similar performance, just with more memory use.

    There's a lot of database software built on array languages, with kdb and Jd being the most prominent as far as I know.

    [0] https://aplwiki.com/wiki/Key

    [1] https://code.jsoftware.com/wiki/Vocabulary/slashdot#dyadic

Sort of related: thinking in relational algebra, or SQL. It appears to be "natural" to think about computing one atomic value at a time, in loops, or slightly less intuitively, recursive functions. (That latter choice may follow from whether your first language was pure Lisp.)

I was fortunate to have a teacher whose database course drilled relational algebra into us. This was in the 70s, shortly after Codd's paper, and well before SQL was invented, much less established. Now I think about much computation algebraically (and often functionally). But I do see that this is "unnatural" for many, including students, having taught databases for several years.

SQL reflects this. I often see students writing nested subqueries, because that is more procedural, where joins would be a cleaner choice. A colleague of mine wrote a paper many years ago, pointing out that thinking procedurally is more "natural" for many: https://dl.acm.org/doi/10.1145/319628.319656. But thinking set-at-a-time instead of one-at-a-time is a valuable skill, not that far off from thinking functionally.

  • It's easier not to mess up table based filters using explicit semi-join operators (eg. in, not in, exists) instead of using regular joins because joins can introduce duplicates.

    Give me 'any join' operation - ie. just select the first value instead of all, and I'll happily use joins more. They are actually more intuitive.

    It's not that relational algebra is untintuitive. It's because standard SQL sucks.

    • Indeed, I've taught myself to only use JOIN when I actually need some data from the table I join. For everything else I use EXISTS and friends.

      I was thinking SQL could do with a keyword for that, like maybe FILTER, that looks like a JOIN but works like EXISTS.

      1 reply →

    • My problem with semijoins is that the semantics of "what exactly does a SELECT evaluate to inside an expression" are sometimes murky and might vary across databases.

      3 replies →

  • Having taught it, do you have any recommendations for folks who are looking to improve their thinking in relations/sets/SQL skills?

    • I teach, in order, more or less (there is some interleaving):

      - Data modeling - Relational algebra - SQL - DBMS architecture (buffering, btrees, query processing, index selection, ...) - Query optimization - Transactions

      A few assignments have students write an in-memory relational algebra implementation, and then use it to write queries. A typical query is a deeply nested set of function calls (as a "one-liner"). And only then to we get to SQL. The hope is that RA is so ingrained, that the connections to SQL are easier to see. And this background is also really useful in understanding query optimization.

      All of this material, including assignments, is available online (this is my own webserver): http://geophile.com/115.

  • I think this is related to "wholemeal" programming as some haskellers/FP-ists do, thinking in sets/tree/graphs operations.

  • > SQL reflects this. I often see students writing nested subqueries, because that is more procedural, where joins would be a cleaner choice.

    In my experience, in the non-ad-hoc use-case, views can often be substituted for the procedural approach, forming the equivalent of a POSIX pipe.

    *> A colleague of mine wrote a paper many years ago, pointing out that thinking procedurally is more "natural" for many: https://dl.acm.org/doi/10.1145/319628.319656. But thinking set-at-a-time instead of one-at-a-time is a valuable skill, not that far off from thinking functionally.

    Hmm. Given the proliferation of tabular data tools (especially spreadsheets) over the intervening 40 years, I wonder if those results would remain the same today (and whether there would be any difference among Excel power users that use pivot tables, etc.)