Book review: Theft of Fire by Devon Eriksen

When I was young, science fiction was the genre of choice for many engineers and scientists. But the genre declined significantly in recent years. Part of the problem is the rise dystopian fiction. In the imagined future, we are no longer conquering space or developing new fantastic technologies, but rather, increasingly, battling the consequences of climate change or of some evil corporation. In some respect, science fiction is always about the present and present has become unimaginative, fearful and generally anxious.

To illustrate, let me quote a recent piece in the Atlantic:

The United States is experiencing an extreme teenage mental-health crisis. From 2009 to 2021, the share of American high-school students who say they feel “persistent feelings of sadness or hopelessness” rose from 26 percent to 44 percent, according to a new CDC study. This is the highest level of teenage sadness ever recorded.

Civilizations are not eternal, they have a life cycle. When young people grow increasingly depressed, when we are more eager to take down than to build up, we are facing a decline. On a per capita basis, most countries in the West are stagnating. Economists like Cowen and Gordon blame in part the relative lack of benefits due to the Internet and technological advancement in computers. But I am sure Roman intellectuals could have blamed the decline of their empire on the fact that late-stage innovations like the onager were not sufficiently impactful.

The covid era is a perfect illustration. Entire societies turned inward, locked up, shut down their industries, their entertainment. We did get an echo of science-fiction, but the dystopian kind. In communist China, we had the so-called big whites, a special police force wearing hazmat suits, enforcing lockdowns. Worldwide, politicians took the habit of wearing black masks. So the Empire from Star Wars was a tangible reality, but without much in the way of rebels. No Luke, no Lea came to be celebrated.

But culture is not a mere reflection of an era. In some sense, culture defines the era, it is a beast that is larger than any one of us. As much as you would like to think that politicians are in charge, they are strictly limited by their culture. Václav Havel made this point in The Power of the Powerless (1985): the ruler of a communist country has no more than a grocer to oppose the prevailing culture, and he may even have less power.

So how do we get back to a future-oriented culture? One where the challenges are met with courage. One where we fight our way through instead of turning inward and regressing?

Culture is all of us. Culture is what we do. It is what we talk about. And it is what we read.

Few books have the power to inspire readers quite like Devon Eriksen’s Theft of Fire. This work of science fiction takes us on a journey through the frozen edge of the solar system, where a hidden treasure lies waiting to be discovered. It is a tale of adventure, intrigue, and the indomitable spirit of humanity.

It is an imperfect, and even somewhat sad future. There are poor people, there are oppressed people. But it is world where you can hope. It is very much the path in our future where Elon Musk’s futuristic and optimistic vision came true. Yes, Elon Musk himself shaped this universe. Eriksen does not tell us that if ‘the man’ (Elon Musk) has his way, we will live forever in harmony. Quite the opposite. If you are intrigued by ChatGPT and Google Gemini, Eriksen has you covered, as well.

Here is a quote to illustrate Eriksen’s style:

In space, everyone can hear you scream. You vacsuit monitors your heartbeat, your blood pressure. It knows when you are injured. And it knows when to cry for help on the 100 MHz emergency band. And everyone is listening. Every suit radio. Every spacecraft. Every robot probe.

Theft of Fire offers a plausible and realistic universe. The characters, Marcus Warnoc and Miranda Foxgrove, are not mere archetypes; they are complex, flawed, and deeply human. Their struggle to trust one another and overcome their own demons is a powerful allegory for the modern human condition. There are class wars, between the very rich and the very poor, but it is not a Marxist story. Marcus might be poor and struggling, but he is never a victim. The rich people can be bad, but so can the poor people.
The world of Theft of Fire is one of contrasts: the cold, unforgiving vacuum of space and the warmth of human connection. It is a testament to the power of storytelling and the enduring appeal of the science fiction genre.

It may maybe telling to look at who published Eriksen’s book: it is a self-published book. As far as I can tell, one of Eriksen’s wives is in charge of marketing. She is the one who reached out to me and suggested this review. Maybe that’s how we change our destiny: from the bottom up. Just like in the novel.

More information: Eriksen’s home page.

Measuring energy usage: regular code vs. SIMD code

Modern processor have fancy instructions that can do many operations at one using wide registers: SIMD instructions. Intel and AMD have 512-bit registers and associated instructions under AVX-512.

You expect these instructions to use more power, more energy. However, they get the job done faster. Do you save energy overall? You should expect so.

Let us consider an example. I can just sum all values in a large array.

float sum(float *data, size_t N) {
  double counter = 0;
  for (size_t i = 0; i < N; i++) {
    counter += data[i];
  return counter;

If I leave it as is, the compiler might be tempted to optimize too much, but I can instruct it to avoid ‘autovectorization’: it will not doing anything fancy.

I can write the equivalent function using AVX-512 intrinsic functions. The details do not matter too much, just trust me that it is expected to be faster for sufficiently long inputs.

float sum(float *data, size_t N) {
  __m512d counter = _mm512_setzero_pd();
  for (size_t i = 0; i < N; i += 16) {
    __m512 v = _mm512_loadu_ps((__m512 *)&data[i]);
    __m512d part1 = _mm512_cvtps_pd(_mm512_extractf32x8_ps(v, 0));
    __m512d part2 = _mm512_cvtps_pd(_mm512_extractf32x8_ps(v, 1));
    counter = _mm512_add_pd(counter, part1);
    counter = _mm512_add_pd(counter, part2);
  double sum = _mm512_reduce_add_pd(counter);
  for (size_t i = N / 16 * 16; i < N; i++) {
    sum += data[i];
  return sum;

Under Linux, we can ask the kernel about power usage. Your software can query the power usage of different components, but I query the overall power usage (per socket). It works well with Intel processors as long as you have privileged access on the system. I wrote a little benchmark that runs both functions.

On a 32-core Ice Lake processors, my results are as follows with a float array that spans about 500 megabytes.

code routine Power (muJ/s) Energy per value (muJ/value)
naive code 0.055 muJ/s 0.11 muJ/value
AVX-512 0.061 muJ/s 0.032 muJ/value

So the AVX-512 uses 3.5 times less energy overall, despite consuming 10% more energy per unit of time.

My benchmark also reports the memory usage of the memory subsystem. In this particular case, the memory usage caused by memory (DRAM) is low and even negligible, according to the kernel.

My benchmark is naive and should only serve as an illustration. The general principle holds, however: if your tasks complete much faster, you are likely to use less power, even if you are using more energy per unit of time.

JSON Parsing: Intel Sapphire Rapids versus AMD Zen 4

Intel has release a new generation of server processors (Sapphire Rapids) while the latest AMD technology (Zen 4) is now broadly available. There are extensive comparisons available. Of particular interest is the open benchmark results which assess various aspects of processor speeds, including JSON parsing performance. In these benchmarks, AMD systems appear to dominate.

I decided to run my own benchmarks using JSON parsing as a reference and commonly available Amazon big nodes. For these tests, I use Amazon Linux 2023 with GCC 11. I use two instances that cost about 5 dollars per hour. Amazon charges me about the same amount for both the AMD and Intel systems.

The AMD instance is of type c7a.24xlarge with an AMD EPYC 9R14 processor (Zen 4 microarchitecture). The Intel instance is of type c7i.metal-24xl with an Intel XeonPlatinum 8488C (Sapphire Rapids microarchitecture). I use systems with multiple cores but my benchmark is entirely single threaded. I could have optimized either system by going with systems that have fewer cores running hotter. In my case, both processors run in practice at a comparable frequency, with a slight advantage for AMD (3.5 GHz vs 3.4 GHz).

The gist of the result is that neither system dominates the other one. In some benchmarks, Intel wins, in others AMD wins. It is very closely matched.

Intel results:

simdjson On-Demand simdjson DOM yyjson rapidjson nlohmann/json Boost JSON
json2msgpack 3.68 GB/s 2.67 GB/s 1.72 GB/s 0.71 GB/s 0.03 GB/s 0.42 GB/s
partial_tweets 6.83 GB/s 4.77 GB/s 2.41 GB/s 0.77 GB/s 0.13 GB/s 0.50 GB/s
distinct_user_id 6.99 GB/s 4.90 GB/s 2.52 GB/s 0.67 GB/s 0.14 GB/s 0.49 GB/s
kostya 2.92 GB/s 2.03 GB/s 0.83 GB/s 0.80 GB/s 0.12 GB/s 0.47 GB/s

AMD results:

simdjson On-Demand simdjson DOM yyjson rapidjson nlohmann/json Boost JSON
json2msgpack 3.09 GB/s 2.45 GB/s 1.93 GB/s 0.68 GB/s 0.03 GB/s 0.38 GB/s
partial_tweets 6.84 GB/s 4.22 GB/s 2.64 GB/s 0.77 GB/s 0.12 GB/s 0.46 GB/s
distinct_user_id 6.94 GB/s 4.26 GB/s 2.58 GB/s 0.77 GB/s 0.13 GB/s 0.47 GB/s
kostya 4.03 GB/s 2.71 GB/s 1.00 GB/s 0.78 GB/s 0.12 GB/s 0.52 GB/s

You can reproduce my results by grabbing simdjson and running bench_ondemand.

I do not pretend that this single data point is sufficient to make purchasing decisions or to assess the Intel and AMD technology. Take it as a data point.

Further reading. On-demand JSON: A better way to parse documents?, Software: Practice and Experience (to appear)

How fast is rolling Karp-Rabin hashing?

A hash function maps values (e.g., strings) into a fixed number of strings, typically smaller than the original. It is useful to compare quickly two long strings, for example. Instead of comparing the strings, you may compare the hash values.

A simple hash function consists in repeatedly multiplying the hash value by some constant B (e.g., you may pick a number such as 31):

uint32_t hash = 0;
for (size_t i = 0; i < len; i++) {
  hash = hash * B + data[i];
return hash;

I credit Karp-Rabin for this type of hash functions. It is an example of recursive hashing: the hash function is computed by taking the hash value and updating it with the next character.

Given a long strings, you may want to hash all sequences of N characters. A naive approach might be as follows:

for(size_t i = 0; i < len-N; i++) {
  uint32_t hash = 0;
  for(size_t j = 0; j < N; j++) {
    hash = hash * B + data[i+j];

You are visiting each character value N times. If N is large, it is inefficient.

You can do better using a rolling hash function: instead of recomputing the hash function anew each time, you just update it. It is possible to only access each character twice:

uint32_t BtoN = 1;
for(size_t i = 0; i < N; i++) { BtoN *= B; }

uint32_t hash = 0;
for(size_t i = 0; i < N; i++) {
  hash = hash * B + data[i];
// ...
for(size_t i = N; i < len; i++) {
  hash = hash * B + data[i] - BtoN * data[i-N];
  // ...

The most expensive component of this routine is the line with two character accesses and two multiplications.

I wrote a simple benchmark in C++ to see how fast a straight-forward implementation could be. I use a set window of size 8 for the purpose of this analysis, but the larger the window, the less competitive the naive implementation becomes.

rolling hashing 0.75 GB/s 13 instructions/byte
naive hashing 0.18 GB/s 61 instructions/byte

Karp-Rabin is not the only type hash functions. Tabulated hashing is another approach that would be much more compute efficient.

However, I suspect we could greatly improve my naive implementation of the Karp-Rabin rolling hash function.

Further reading:

Possibly relevant software:

C23: a slightly better C

One of the established and most popular programming languages is the C programming language. It is relatively easy to learn, and highly practical.

Maybe surprisingly, the C programming language keeps evolving, slowly and carefully. If you have GCC 13 or LLVM (Clang) 16, you already have a compiler with some of the features from the latest standard (C23).

// Only include stdio.h if it exists
#if __has_include (<stdio.h>)
  #include <stdio.h>

#include <stdlib.h>

void f() {}

int g(int x) {
  return x + 1;

int main() {
  f(); // compile-time warning: 'f' is deprecated
  g(1); // compile-time warning
  auto x = 0b1111;
  typeof(x) y = 1'000'000; // type of y is the same as x
  printf("%d\n", x); // prints 15
  printf("%d\n", y); // prints 1000000
  constexpr int N = 10;
  // compile-time asserts using static_assert
  static_assert (N == 10, "N must be 10");
  bool a[N]; // array of N booleans
  for (int i = 0; i < N; ++i) {
    a[i] = true;
  printf("%d\n", a[0]); // prints 1

  1. The first part of the code contains some preprocessor directives, which are instructions for the compiler to process the source code before compiling it. The #if directive checks a condition at compile time and includes the following code only if the condition is true. The __has_include macro is a feature of C++17 adopted by C23 that checks if a header file exists and can be included. In this instance, it is not useful because we know that stdio.h is present, but in other instances, this can prove useful to determine what headers are available.
  2. The next part of the code defines two functions with attributes, which are annotations that provide additional information to the compiler about the behavior or usage of a function, variable, type, etc.
    • The [[deprecated]] attribute is a feature of C++14 adopted by C23 that marks a function as obsolete and discourages its use. The compiler will issue a warning if the function is called or referenced.
    • The [[nodiscard]] attribute is a feature of C++17 adopted by C23 that indicates that the return value of a function should not be ignored or discarded. The compiler will issue a warning if the function is called from a discarded-value expression.

    In this case, the function f is deprecated and does nothing, while the function g returns the input value plus one and should not be ignored. The first two lines of the main function call the functions f and g and trigger the warnings.

  3. The third line of the main function declares a variable x with the auto keyword, which is a feature of C++11 that lets the compiler deduce the type of the variable from its initializer. In this case, the initializer is a binary literal, which is a feature of C++14 and adopted by C23 that allows writing integer constants in binary notation using the prefix 0b. The value of x is 0b1111, which is equivalent to 15 in decimal.
  4. The fourth line declares another variable y with the typeof operator that returns the type of an expression. In this case, the expression is x, so the type of y is the same as the type of x. The initializer of y is a digit separator, which is a feature of C++14 adopted by C23 that allows inserting single quotes between digits in numeric literals to improve readability. The value of y is 1’000’000, which is equivalent to 1000000 in decimal.
  5. The seventh line declares a constant variable N with the constexpr keyword, which is a feature of C++11 adopted by C23 that indicates that the value of the variable can be evaluated at compile time. The value of N is 10. Previously, one would often use a macro to define a compile-time constant (e.g., #define N 10).
  6. The eighth line uses the static_assert keyword, which is a syntax of C++11 adopted by C23 that performs a compile-time assertion check. The keyword takes a boolean expression and an optional string message as arguments. If the expression is false, the compiler will emit an error and stop the compilation, displaying the message. If the expression is true, the compiler will do nothing. In this case, the expression is N == 10, which is true, so the compilation continues.
  7. The ninth line declares an array of N booleans named a. An array is a collection of elements of the same type that are stored in contiguous memory locations. The size of the array must be a constant expression for standard C arrays (otherwise it becomes a variable-length array which may be less efficient), which is why N is declared with constexpr. We also use the keywords true and false which become standard keywords in C23.

There are many more features in C23, but it will take some time for compilers and system librairies to catch up.

My thoughts so far:

  • The introduction of constexpr in C will probably help reduce the dependency on macros, which is a good idea generally. Macros work well in C, but when a bug is introduced, it can be difficult get meaningful error messages. It does not happen too often, but in large code bases, it can be a problem.
  • I personally rarely use auto and typeof in other languages, so I don’t expect to use them very much in C. In some specific cases, it can greatly simply one’s code, however. It is likely going to help reduce the reliance on macros.
  • The idea behind static_assert is great. You run a check that has no impact on the performance of the software, and may even help it. It is cheap and it can catch nasty bugs. It is not new to C, but adopting the C++ syntax is a good idea.
  • The __has_include feature can simplify supporting diverse standard libraries and test for available libraries. For example, it becomes easy to check whether the standard library supports AVX-512. If a header is missing, you can fail the compilation with instructions (e.g., you need to install library X). It is generally a good idea for people who need to write portable code that others can rely upon.
  • I did not include the introduction of `char8_t` in the language. I worked extensively with Unicode in C++ and I have not found good use cases for the `char8_t` type so far: `char` is always sufficient in my experience.

How much memory bandwidth do large Amazon instances offer?

In my previous post, I described how you can write a C++ program to estimate your read memory bandwidth. It is not very difficult: you allocate a large memory region and you read it as fast as you can. To see how much bandwidth you may have if you use multithreaded applications, you can use multiple threads, where each thread reads a section of the large memory region.

The server I used for the blog post, a two-CPU Intel Ice Lake server has a maximal bandwidth of about 130 GB/s. You can double this amount of bandwidth with NUMA-aware code, but it will require further engineering.

But you do not have access to my server. What about a big Amazon server? So I spun out an r6i.metal instance from Amazon. These servers can support 128 physical threads, they have 1 terabyte of RAM (1024 GB) and 6.25 GB/s of network bandwidth.

Running my benchmark program on this Amazon server revealed that they have about 115 GB/s of read memory bandwidth. That is not counting NUMA and other sophisticated tricks. Plotting the bandwidth versus the number of threads used reveals that, once again, you need about 20 threads to maximized memory bandwidth although you get most of it with only 15 threads.

My source code is available.

Estimating your memory bandwidth

One of the limitations of a compute is the memory bandwidth. For the scope of this article, I define “memory bandwidth” as the maximal number of bytes you can bring from memory to the CPU per unit of time. E.g., if your system has 5 GB/s of bandwidth, you can read up to 5 GB from memory in one second.

To measure this memory bandwidth, I propose to read data sequentially. E.g., you may use a function where we sum the byte values in a large array. It is not necessary to sum every byte value, you can skip some because the processor operates in units of cache lines. I do not know of a system that uses cache lines smaller than 64 bytes, so reading one value every 64 bytes ought to be enough.

uint64_t sum(const uint8_t *data,
    size_t start, size_t len, size_t skip) {
  uint64_t sum = 0;
  for (size_t i = start; i < len; i+= skip) {
    sum += data[i];
  return sum;

It may not be good enough to maximize the bandwidth usage: your system has surely several cores. Thus we should use multiple threads. The following C++ code divides the input into consecutive segments, and assigns one thread to each segment, dividing up the task as fairly as possible:

size_t segment_length = data_volume / threads_count;
size_t cache_line = 64;
for (size_t i = 0; i < threads_count; i++) {
  threads.emplace_back(sum, data, segment_length*i,
       segment_length*(i+1), cache_line);
for (std::thread &t : threads) {

I ran this code on a server with two Intel Ice Lake  processors. I get that the more threads I use, the more bandwidth I am able to get up to around 15 threads. I start out at 15 GB/s and I go up to over 130 GB/s. Once I reach about 20 threads, it is no longer possible to get more bandwidth out of the system. The system has a total of 64 cores, over two CPUs. My program does not do any fiddling with locking threads to cores, it is not optimized for NUMA. I have transparent huge pages enabled by default on this Linux system.

My benchmark ought to be make it easy for the processor to maximize bandwidth usage, so I would not expect more complicated software to hit a bandwidth limit with as few as 20 threads.

My source code is available.

This machine has two NUMA nodes. You can double the bandwidth by running the same benchmark using two NUMA nodes. E.g., under Linux you might call:

numactl --cpunodebind=1 --membind=1 ./bandwidth  & numactl --cpunodebind=0 --membind=0 ./bandwidth

Be aware that NUMA has some downsides. For example, the communication between NUMA nodes is relatively expensive.

Further reading: Many tools to measure bandwidth. I also wrote a second blog post on this theme: How much memory bandwidth do large Amazon instances offer?

Implementing the missing sign instruction in AVX-512

Intel and AMD have expanded the x64 instruction sets over time. In particular, the SIMD (Single instruction, multiple data) instructions have become progressively wider and more general: from 64 bits to 128 bits (SSE2), to 256 bits (AVX/AVX2) to 512 bits (AVX-512). Interestingly, many instructions defined on 256 bits registers through AVX/AVX2 are not available on 512 bits registers.

With SSSE3, Intel introduced sign instructions, with the corresponding intrinsic functions (e.g., _mm_sign_epi8). There are 8-bit, 16-bit and 32-bit versions.  It was extended to 256-bit registers in AVX2.

What these instructions do is to apply the sign of one parameter to the other parameter. It is most easily explained as pseucode code:

function sign(a, b): # a and b are integers
   if b == 0 : return 0
   if b < 0 : return -a 
   if b > 0 : return a

The SIMD equivalent does the same operation but with many values at once. Thus, with SSSE3 and psignb, you can generate sixteen signed 8-bit integers at once.

You can view it as a generalization of the absolute function: abs(a) = sign(a,a). The sign instructions are very fast. They are used in numerical analysis and machine learning: e.g., it is used in llama.cpp, the open source LLM project.

When Intel designed AVX-512 they decided to omit the sign instructions. So while we have the intrinsic function  _mm256_sign_epi8, we don’t have _mm512_sign_epi8. The same instructions are missing for 16 bits and 32 bits integers (e.g., no _m512_sign_epi16 is found).

You may implement it for AVX-512 with a several instructions. I found this one approach:

#include <x86intrin.h>

__m512i _mm512_sign_epi8(__m512i a, __m512i b) {
  __m512i zero = _mm512_setzero_si512();
  __mmask64 blt0 = _mm512_movepi8_mask(b);
  __mmask64 ble0 = _mm512_cmple_epi8_mask(b, zero);
  __m512i a_blt0 = _mm512_mask_mov_epi8(zero, blt0, a);
  return _mm512_mask_sub_epi8(a, ble0, zero, a_blt0);;

It is disappointingly expensive. It might compile to four or five instructions:

vpmovb2m k2, zmm1
vpxor xmm2, xmm2, xmm2
vpcmpb k1, zmm1, zmm2, 2
vpblendmb zmm1{k2}, zmm2, zmm0
vpsubb zmm0{k1}, zmm2, zmm1

In practice, you may not need to pay such a high price. The reason the problem is difficult is that we have three cases to handle (three signs b=0, b>0, b<0).  If you do not care about the case ‘b = 0’, then you can do it in two instruction, not counting the zero (one xor):

#include <x86intrin.h>

__m512i _mm512_sign_epi8_cheated(__m512i a, __m512i b) {
   __m512i zero = _mm512_setzero_si512();
  __mmask64 blt0 = _mm512_movepi8_mask(b);
  return _mm512_mask_sub_epi8(a, blt0, zero, a);;

E.g., we implemented…

function sign_cheated(a, b): # a and b are integers
   if b < 0 : return -a 
   if b ≥ 0 : return a

Science and Technology links (December 30th 2023)

  1. Parenting does not appear to be able to determine the personality traits of a child.
  2. When the last ice age ended, 12,000 years ago, the Sahara was green and full of life. It turned into a desert about 5,500 years ago.
  3. Fadnes et al. claim that the UK population could live 10 years older if it changed its eating habits.
  4. By studying 175 different populations, You et al. find that meat intake predicts longevity: people who eat more meat live longer.
  5. According to an editorial in the journal Nature, scientists who work in industry are more satisfied and better paid than are colleagues in academia. Industry scientists report less bullying and discrimination.
  6. The Asch experiment examined the extent to which individuals would conform to the majority view, even when that view was clearly incorrect. The experiment involved a group of participants, one of whom was the actual subject of the experiment, and the rest were people who knew the true purpose of the experiment and acted according to a script. The group was shown a series of images with lines of different lengths and asked to identify which two lines were the same length. The results showed that a significant number of participants conformed to the majority view, even when it was clearly wrong. The Asch experiment is important because it highlights the influence of social factors on individual beliefs. Most people just adopt the prevaling beliefs, even when they are clearly incorrect. In other words, very few people can think for themselves. They just reproduce what they are shown or what they see others doing, like mere monkeys. Unfortunately, the original experiment is robust with respect to replication. We also find that even financial incentive fail to make people more critical.
  7. Weather prediction is one of the first application of powerful computers. To this day, we rely on predictions made by specialized services: we don’t generally compute our own weather predictions. Google Deepmind claims to be able to predict the weather accurately on a normal computer using artificial intelligence.

Measuring the size of the cache line empirically

Our computers do not read or write memory in units of bits or even bytes. Rather memory is accessed in small blocks of memory called “cache lines”. For a given system, the cache line size is usually fixed and small (e.g.,  16 to 256 bytes). All Intel/AMD x64 systems I have used relied on a 64-byte cache line. My current Apple laptop with its ARM-based M2 processor relies on a 128-byte cache line.

How can you measure the size of the cache line? I asked on Twitter/X and a few people (e.g., Peter Cawley, Will Bickford) pointed to a phenomenon called false sharing. I will come back to false sharing in a future blog post. I do not want to discuss or cover false sharing because it depends on parallelism. Furthermore, it may be trickier than it seems. I want a simpler solution.

Many people (Robert Clausecker, Sam Westrick, Tomasz Kowalczewski, Vinoth Deivasigamani, Sergey Slotin and many others) proposed using ‘strided access’ benchmark.

I finally decided to test a strided copy: from a large array, I copy every N bytes to another large array. It is a problem that should be largely “memory access bound” as long as N is not too small. I start N at 16. Importantly, I never read my own writes, so I avoid concerns with 4K aliasing on Intel processors.

If N is larger than twice the cache line, then I can effectively skip one cache line out of two. If N is smaller than the cache line, then every cache line must be accessed. Having a stride value just above the cache line should be not sufficiently to see large gains: but you expect the speed to almost double once you reach twice the size of the cache line if the only thing that matters are cache lines.

Sadly, several other factors come into play on a modern system with such a benchmark. There is more than just the cache-line size as a variable! So we need to verify the model experimentally.

I wrote the benchmark in C, but the actual C compiler is unlikely to be relevant. The original code was in Go and I got the same results, but I switched to C to make sure I avoided Go-specific issues. Interestingly, ChatGPT converted the Go code to C code automatically for  me, with just a couple of small mistakes. Mind you, it is deliberately simple code.

I run each experiment, for each stride size, 10 times and I record the maximum, the minimum and the average. I use an Apple M2 processor running on my laptop and an Intel-based server. I do not require a particular memory alignment when allocating memory. I do not make any other attempt to control the results.

The numbers on the server are quite nice, with hardly any difference between the average, the maximum and the minimum. If your stride is 129, you are 66% faster than when your stride is 64. This suggests that the cache-line size is 64 bytes. The gain is not 2x as I would have expected but the processing might be loading cache lines speculatively. Observe how a stride that is a multiple of 64 (e.g., 128 or 256) is slightly bad: we see the performance dip visibly. It might be due to a caching issue e.g., only half the cache-line addresses are used which makes it more difficult for the processor to make full use of its cache (due to address aliasing).

The result on my laptop are much less clean even though I kept the laptop unused during the benchmarking. In this instance, if your stride is 257, you are more than 2 times faster than when your stride is 128. It suggests that the cache-line size is 128 bytes. Just like the Intel system, a stride of 128 is unfortunate: there is a visible performance dip.

Note that we do not actually copy data at 300 GB/s, that would be impossible on its face on my hardware: but we can copy an array that fast if we just copy one byte per block of 512 bytes.

Thus you can empirically measure the size of the cache line with a strided copy of a large array… As soon you use a stride that is twice the cache line, you should be more than 50% faster.

Fast Buffer-to-String conversion in JavaScript with a Lookup Table

When programming in a JavaScript environment such as Node.js, you might recover raw data from the network and need to convert the bytes into strings. In a system such as Node.js, you may represent such raw bytes using a Buffer instance.

You can conveniently convert a Buffer instance into a JavaScript (mybuffer.toString()). But, maybe surprisingly, creating new strings can be a bottleneck. Thus a worthwhile optimization might be to try to recognize that your incoming bytes are one out of a list of known strings. This is not a problem unique to JavaScript.

One example of such a problem occurs when parsing HTTP headers. These headers contain common strings such as  ‘accept’, ‘accept-encoding’, ‘content-encoding’, ‘content-language’, ‘content-length’, ‘from’, ‘host’, etc. If we can recognize the bytes as one of these strings, we can just point at the existing strings. To make things more complicated, we might want to ignore the case, so that both inputs ‘Accept’ and ‘ACCEPT’ should be mapped to accept’.

This problem has been addressed recently in a library called undici. This library provides Node.js with an HTTP/1.1 client. GitHub user tsctx initially proposed solving this problem using a trie implemented with JavaScript objects. A trie is a type of data structure that is used to store and search for strings in an efficient way. In its simplest implementation (sometimes called a digital search tries), we first branch out on the first character, and each possible character becomes a new tree based on the second character and so forth. The last node of each string is marked as the end of the word, and may also store some additional information, such as a point to a desired output. The problem with such an approach is that it can use much memory. And, indeed, I estimated that tsctx‘s implementation might cost between 1 MB to 2 MB of memory. Whether that is a concern depends on your priorities. Following my comments, user tsctx opted for a new strategy that may be less time efficient, but that is significantly more economical memory-wise: a ternary. A ternary tree is similar to a binary tree, but each node can have up to three children, usually called left, mid, and right.

I think that tsctx‘s is excellent, but it is sometimes important to compare with a few competitors.

I decided to implement my own approach based the observation that it is fairly easy to quickly identify a candidate using solely the length of the input. For example, there is only one candidate string of length 2: ‘TE’. So it makes sense to write code like this:

function toLowerCase(s) {
 var len = s.length;
  switch (len) {
   case 2:
    // check that the buffer is equal to te and return it if so
    if ((s[0] == 116 || s[0] == 84) && (s[1] == 101 || s[1] == 69)) {
     return "te";

This code is a function that takes a string as an input and returns a lowercased version of it. The function works as follows: It declares a variable called len and assigns it the value of the length of the input string s. It uses a switch statement to check the value of len and execute different code blocks depending on the case.
In this example, the function only has one case, which is 2. In the case of 2, the function checks that the input string is equal to “te” or “TE” or “Te” or “tE”. It does this by comparing the ASCII codes of the characters in the string. The ASCII code of t is 116, the ASCII code of T is 84, the ASCII code of e is 101, and the ASCII code of E is 69. The function uses the logical operators || (or) and && (and) to combine the conditions. If the input string matches any of these four combinations, the function will return “te”. Here is an example of how the function works: If the input string is “te”, the function will return “te”. If the input string is “TE”, the function will return “te”. If the input string is “Te”, the function will return “te”. If the input string is “tE”, the function will return “te”. If the input string is “ta”, the function will continue.

If the buffer has length 3, then I have four possible candidate strings (age, ect, rtt, via). I can differentiate them by looking only at the first character. The logic is much the same:

case 3:
  switch (s[0]) {
   case 97: case 65:
    // check that the buffer is equal to age and return it if so
    if ((s[1] == 103 || s[1] == 71) && (s[2] == 101 || s[2] == 69)) {
      return "age";
   case 101:case 69:
    // check that the buffer is equal to ect and return it if so
    if ((s[1] == 99 || s[1] == 67) && (s[2] == 116 || s[2] == 84)) {
     return "ect";
   case 114:case 82:
    // check that the buffer is equal to rtt and return it if so
    if ((s[1] == 116 || s[1] == 84) && (s[2] == 116 || s[2] == 84)) {
      return "rtt";
   case 118:case 86:
    // check that the buffer is equal to via and return it if so
    if ((s[1] == 105 || s[1] == 73) && (s[2] == 97 || s[2] == 65)) {
     return "via";

It is easy enough to do it by hand, but it gets tedious, so I wrote a little Python script. It is not complicated… I just repeat the same logic in a loop.

Pay attention to the fact that the switch key is made of nearly continuous integers from 2 to 35… It means that a good compiler will almost surely use a jump table and not a series of comparisons.

First let us compare the memory usage of the four approaches: the original (simple) code used by undici, the naive switch-case approach, the ternary tree and the digital search trie. I use various recent versions of Node.js on a Linux server. I wrote scripts that simply include the function, and only the function, and I print the memory usage. I repeat five times and report the lowest figure. When using Node.js, I call the garbage collector and pause to try to minimize the memory usage.

Node.js 21 Node.js 20 Node.js 19 Bun 1.0
original 43.3 MB 42.4 MB 44.9 MB 20.2 MB
naive switch 43.3 MB 42.9 MB 42.9 MB 23.8 MB
ternary tree 43.5 MB 44.2 MB 45.2 MB 29.3 MB
digital search trie 45.1 MB 44.6 MB 47.0 MB 26.7 MB

Thus only the digital search trie appears to bring a substantial memory usage with Node.js 21. If you use a different version of Node.js or a different operating system, results will differ… but I verified that the conclusion is the same on my macBook.

What about performance? I use an Intel Ice Lake processor running at 3.2 GHz. I wrote a small benchmark that parses a few headers. I rely on a well-known JavaScript benchmark framework (mitata).

Node.js 21 Node.js 20 Node.js 19 Bun 1.0
original 15 µs 14 µs 15 µs 12 µs
naive switch 7.8 µs 7.9 µs 7.8 µs 8.2 µs
ternary tree 9.4 µs 9.4 µs 9.0 µs 8.8 µs
digital search trie 12 µs 12 µs 11 µs 10 µs

I am not quite certain why the digital search trie does poorly in this case. I also ran the same experiment on my 2022 macBook (Apple M2 processor). I am usually against benchmarking on laptops, but these macBooks tend to give very stable numbers.

Node.js 21 Node.js 20 Node.js 19 Bun 1.0
original 8.5 µs 9.1 µs 8.5 µs 8.2 µs
naive switch 5.0 µs 4.9 µs 4.7 µs 5.3 µs
ternary tree 5.8 µs 5.8 µs 5.6 µs 6.1 µs
digital search trie 5.3 µs 5.5 µs 5.4 µs 5.5 µs

Thus I would conclude that both the naive switch and the ternary tree are consistently faster than the original. The original implementation is about 1.8 times slower than the naive switch when using Node.js 21.

One approach I did not try is perfect hashing. I am concerned that it might be difficult to pull off because JavaScript might not compile the code efficiently. One benefit of perfect hashing is that it can be nearly branchless so it provides consistent performance. We could use the perfect hashing strategy with the switch case approach: we would have just one hash computation, and then we would end up straight to single buffer-to-target comparison. It would like a C/C++ implementation in spirit, although it would generate more code.

We rely on the fact that this function was identified as a bottleneck. We ran a microbenchmark, but it would be useful to see whether these functions make a difference in a realistic application.

My source code and benchmark is online. It can be improved, pull requests invited.

How fast can you validate UTF-8 strings in JavaScript?

When you recover textual content from the disk or from the network, you may expect it to be a Unicode string in UTF-8. It is the most common format. Unfortunately, not all sequences of bytes are valid UTF-8 and accepting invalid UTF-8 without validating it is a security risk.

How might you validate a UTF-8 string in a JavaScript runtime?

You might use the valid-8 module:

import valid8 from "valid-8";
if(!valid8(file_content)) { console.log("not UTF-8"); }

Another recommended approach is to use the fact that TextDecoder can throw an exception upon error:
new TextDecoder("utf8", { fatal: true }).decode(file_content)

Or you might use the isUtf8 function which is part of Node.js and Bun.
import { isUtf8 } from "node:buffer";
if(!isUtf8(file_content)) { console.log("not UTF-8"); }

How do they compare? Using Node.js 20 on a Linux server (Intel Ice Lake), I get the following speeds with three files representative of different languages. The Latin file is just ASCII. My benchmark is available.
Arabic Chinese Latin
valid-8 0.14 GB/s 0.17 GB/s 0.50 GB/s
TextDecoder 0.18 GB/s 0.19 GB/s 7 GB/s
node:buffer 17 GB/s 17 GB/s 44 GB/s

The current isUtf8 function in Node.js was implemented by Yagiz Nizipli. It uses the simdutf library underneath. John Keiser should be credited for the UTF-8 validation algorithm.

Parsing 8-bit integers quickly

Suppose that you want to parse quickly 8-bit integers (0, 1, 2, …, 254, 255) from an ASCII/UTF-8 string. The problem comes up in the simdzone project lead by Jeroen Koekkoek (NLnet Labs). You are given a string and its length: e.g., ’22’ and length is 2. The naive approach in C might be:

int parse_uint8_naive(const char *str, size_t len, uint8_t *num) {
  uint32_t n = 0;
  for (size_t i = 0, r = len & 0x3; i < r; i++) {
    uint8_t d = (uint8_t)(str[i] - '0');
    if (d > 9)
     return 0;
    n = n * 10 + d;
  *num = (uint8_t)n;
  return n < 256 && len && len < 4;

This code is a C function that takes a string of characters, its length, and a pointer to an unsigned 8-bit integer as input arguments. The function returns a Boolean value indicating whether the input string can be parsed into an unsigned 8-bit integer or not. It restricts the input to at most three characters but it allows leading zeros (e.g. 002 is 2).

The function first initializes a 32-bit unsigned integer n to zero, we will store our answer there. The function then iterates over the input string, extracting each digit character from the string and converting it to an unsigned 8-bit integer. The extracted digit is then added to n after being multiplied by 10. This process continues until the end of the string or until the function has processed 4 bytes of the string. Finally, the function assigns the value of n to the unsigned 8-bit integer pointed to by num. It then returns a boolean value indicating whether the parsed integer is less than 256 and the length of the input string is between 1 and 3 characters.  If the input string contains any non-digit characters or if the length of the string is greater than 3 bytes, the function returns false.

If the length of the input is predictable, then this naive function should be fast. However, if the length varies (between 1 and 3), then the processor will tend to  mispredict branches, and expensive penalty.

In C++, you could call from_chars:

int parse_uint8_fromchars(const char *str, size_t len, uint8_t *num) {
  auto [p, ec] = std::from_chars(str, str + len, *num);
  return (ec == std::errc());

The std::from_chars function takes three arguments: a pointer to the beginning of the input character sequence, a pointer to the end of the input character sequence, and a reference to the output variable that will hold the parsed integer value. The function returns a std::from_chars_result object that contains two members: a pointer to the first character that was not parsed, and an error code that indicates whether the parsing was successful or not.

In this function, the std::from_chars function is called with the input string and its length as arguments, along with a pointer to the unsigned 8-bit integer that will hold the parsed value. The function then checks whether the error code returned by std::from_chars is equal to std::errc(), which indicates that the parsing was successful. If the parsing was successful, the function returns true. Otherwise, it returns false.

Can we do better than these functions? It is not obvious that we can. A function that reads between 1 and 3 bytes is not a function you would normally try to optimize. But still: can we do it? Can we go faster?

Suppose that you can always read 4 bytes, even if the string is shorter (i.e., there is a buffer). This is often a safe assumption. If you numbers are within a larger string, then you can often check efficiently whether you are within 4 bytes of the end of the string. Even if that is not the case, reading 4 bytes is always safe as long as you do not cross a page boundary, something you may check easily.

So you can load your input into a 32-bit word and process all bytes at once, in a single register. This often called SWAR: In computer science, SWAR means SIMD within a register, which is a technique for performing parallel operations on data contained in a processor register.

Jeroen Koekkoek first came up with a valid SWAR approach, but it was only about 40% faster than the naive approach in the case where we had unpredictable inputs, and potentially slower than the naive approach given predictable inputs. Let us examine an approach that might be competitive all around:

int parse_uint8_fastswar(const char *str, size_t len, 
    uint8_t *num) {
  if(len == 0 || len > 3) { return 0; }
  union { uint8_t as_str[4]; uint32_t as_int; } digits;
  memcpy(&digits.as_int, str, sizeof(digits));
  digits.as_int ^= 0x30303030lu;
  digits.as_int <<= ((4 - len) * 8);
  uint32_t all_digits = 
    ((digits.as_int | (0x06060606 + digits.as_int)) & 0xF0F0F0F0) 
       == 0;
  *num = (uint8_t)((0x640a01 * digits.as_int) >> 24);
  return all_digits 
   & ((__builtin_bswap32(digits.as_int) <= 0x020505));

Again, this code is a C function that takes a string of characters, its length, and a pointer to an unsigned 8-bit integer as input arguments. The function returns a boolean value indicating whether the input string can be parsed into an unsigned 8-bit integer or not. We check whether the length is in range ([1,3]), if not, we return false, terminating the function. After this initial check, the function first initializes a union digits that contains an array of 4 unsigned 8-bit integers and a 32-bit unsigned integer. The function then copies the input string into the 32-bit unsigned integer using the memcpy function. The memcpy function copies the input string into the 32-bit unsigned integer. In production code where you do not know the target platform, you would want to reverse the bytes when the target is a big-endian system. Big endian systems are vanishingly rare: mostly just mainframes. Modern systems compile a byte reversal to a single fast instructions. For code on my blog post, I assume that you do not have a big-endian system which is 99.99% certain.

The function then flips the bits of the 32-bit unsigned integer using the XOR operator and the constant value 0x30303030lu. This operation converts each digit character in the input string to its corresponding decimal value. Indeed, the ASCII characters from 0 to 9 have code point values 0x30 to 0x39 in ASCII. The function then shifts the 32-bit unsigned integer to the left by a certain number of bits, depending on the length of the input string. This operation removes any trailing bytes that were not part of the input string. Technically when the length is not within the allowed range ([1,3]), the shift might be undefined behaviour, but we return a false value in this case earlier, indicating that the result of the computation is invalid.

The next part is where I contributed to the routine. First we check all characters are digits using a concise expression. The function then multiplies the 32-bit unsigned integer by the constant value 0x640a01 using a 32-bit unsigned integer. It is a concise way to do two multiplications (by 100 and by 10) and two sums at once. Observe that 0x64 is 100 and 0x0a is 10. The result of this multiplication is then shifted to the right by 24 bits. This operation extracts the most significant byte of the 32-bit unsigned integer, which represents the parsed unsigned 8-bit integer. Finally, the function assigns the value of the parsed unsigned 8-bit integer to the unsigned 8-bit integer pointed to by num. It then returns a boolean value indicating whether the parsed integer is less than 256 and made entirely of digits.

The function might compile to 20 assembly instructions or so on x64 processors:

lea rcx, [rsi - 4]
xor eax, eax
cmp rcx, -3
jb .END
mov eax, 808464432
xor eax, dword ptr [rdi]
shl esi, 3
neg sil
mov ecx, esi
shl eax, cl
lea ecx, [rax + 101058054]
or ecx, eax
test ecx, -252645136
sete cl
imul esi, eax, 6556161
shr esi, 24
mov byte ptr [rdx], sil
bswap eax
cmp eax, 132358
setb al
and al, cl
movzx eax, al
.END: # %return

To test these functions, I wrote a benchmark. The benchmark uses random inputs, or sequential inputs (0,1,…), and it ends up being very relevant. I use GCC 12 and an Ice Lake (Intel) linux server running at 3.2 GHz. I report the number of millions of numbers parsed by second.

random numbers sequential numbers
std::from_chars 145 M/s 255 M/s
naive 210 M/s 365 M/s
SWAR 425 M/s 425 M/s

So the SWAR approach is twice as fast as the naive approach when the inputs are unpredictable. Otherwise, for predictable inputs, we surpass the naive approach by about 15%. Whether it is helpful in you use case depends on your data so you should run your own benchmarks.

Importantly, the SWAR approach is entirely equivalent to the naive approach, except for the fact that it always reads 4 bytes irrespective of the length.

The from_chars results are disappointing all around. I am puzzled as to why the naive approach is so much faster than the standard library. It solves a slightly different problem but the difference is still quite large. It could be that there is room for optimization in the standard library (GCC 12).

Can you do better? The benchmark is available. As part of the benchmark, we check exhaustively that the validation is correct.

Credit: I am grateful to Jeroen Koekkoek from NLnet Labs for suggesting this problem. The approach described was improved based on proposals by Jean-Marc Bourguet.

Update: User “Perforated Bob”, proposed a version which can be faster under some compilers:

int parse_uint8_fastswar_bob(const char *str, size_t len, uint8_t *num) {
  union { uint8_t as_str[4]; uint32_t as_int; } digits;
  memcpy(&digits.as_int, str, sizeof(digits));
  digits.as_int ^= 0x303030lu;
  digits.as_int <<= (len ^ 3) * 8;
  *num = (uint8_t)((0x640a01 * digits.as_int) >> 16);
  return ((((digits.as_int + 0x767676) | digits.as_int) & 0x808080) == 0) 
   && ((len ^ 3) < 3) 
   && __builtin_bswap32(digits.as_int) <= 0x020505ff;


A simple WebSocket benchmark in Python

Modern web applications often use the http/https protocols. However, when the server and client needs to talk to each other in a symmetrical fashion, the WebSocket protocol might be preferable. For example, if you want to program a multiplayer video game, the WebSocket protocol is almost surely better than http. In A simple WebSocket benchmark in JavaScript, I showed that JavaScript (through Node.js) can produce efficient WebSocket servers, at least in the simplest cases.

My benchmark is what I call a ping-pong. Client 1 sends a message to the server.The server receives the message and broadcasts it to the second client. Client 2 receives the message from the server. Client 2 replies back to the server. Client 1 receives the message. And so forth.

I want to know how many roundtrips I can generate per second. For Node.js (JavaScript), the answer is about 20,000. If you use a faster JavaScript engine (bun), you might get twice as many.

What about Python? I wrote my client using standard code, without any tweaking:

import asyncio
import websockets
async def client1():
  async with websockets.connect('ws://localhost:8080') as websocket:
    message = 'client 1!'
    await websocket.send(message)
    while True:
      response = await websocket.recv()
      await websocket.send(message)

async def client2():
  async with websockets.connect('ws://localhost:8080') as websocket:
    message = 'client 2!'
    while True:
      response = await websocket.recv()
      await websocket.send(message)

async def main():
  task1 = asyncio.create_task(client1())
  task2 = asyncio.create_task(client2())
  await asyncio.gather(task1, task2)

Python has several different frameworks to build WebSocket servers. I picked three that looked popular and mature: sanic, blacksheep, and aiohttp. By default, a module like sanic should use good optimizations like the uvloop module.

My source code is available. I run the benchmark on a Linux machine with Python 3.9. The packets are local, they don’t go out to the Internet. There is no docker container involved.

Python client Node.js 20 client
sanic server 3700 5200
blacksheep server 3000 200
aiohttp server 3600 270
Node.js 20 server 6000 19,000

Mixing the blacksheep and aiohttp servers I wrote with the Node.js server gives terrible results: I have not investigated the cause, but it would be interesting to see if others can reproduce it, or diagnose it.

Otherwise, I get that the sanic server is nearly 4 times slower than the Node.js server. Writing the client in Python appears to cut the performance significantly (except for the blacksheep and aiohttp servers anomaly).

JavaScript shines in comparison with Python in these tests. My implementations are surely suboptimal, and there might be mistakes. Nevertheless, I believe that it is as expected: the standard Python interpreter is not very fast. Node.js has a great Just-in-Time compiler. It would be interesting to switch to faster implementation of Python such as pypy.

This being said, 3000 roundtrips per second might be quite sufficient for several applications. Yet, a real-world WebSocket server would be assuredly slower: it would go through the Internet, it would do non-trivial work, and so forth.

A simple WebSocket benchmark in JavaScript: Node.js versus Bun

Conventional web applications use the http protocol (or the https variant). The http protocol is essentially asymmetrical: a client application such as a browser issues requests and the server responds. It is not possible for the server to initiate communication with the client. Certain types of applications are therefore more difficult to design. For example, if we wanted to design a multiplayer video game using the http protocol, such as a chess game, we could have one server, and two browsers connected to the server. When one of the players moves a piece within its browser, the browser can inform the server via an http request. But how do you inform the second browser? One solution is to have the browsers make requests to the server at regular intervals. A better solution is to use another protocol, the WebSocket protocol.

WebSocket is a network protocol for creating bidirectional communication channels between browsers and web servers. Most browsers support WebSocket, although the standard is relatively recent (2011). It enables the client to be notified of a change in server status, without having to make a request.

You expect WebSocket to be relatively efficient. I wrote an elementary WebSocket benchmark in JavaScript. I use the standard module ws. In my benchmark, I have one server. The server takes whatever messages it receives, and it sends them to other clients. Meanwhile I create two clients. Both clients initiate a connection to the server, so we have two connections. The clients then engage in a continual exchange:

  1. Client 1 sends a message to the server.
  2. The server receives the message and broadcasts it to the second client.
  3. Client 2 receives the message from the server.
  4. Client 2 replies back to the server.
  5. Client 1 receives the message.

My code is as simple as possible. I do not do any trick to go faster. It is ‘textbook’ code.

Importantly, this benchmark has a strong data dependency: there is only just one connection active while the other one is stalling. So we are measuring the latency (how long the trips take) rather than how many requests we can support simultaneously.

How fast can it go? I run my benchmark locally on a Linux server with a server processor (Xeon Gold). The tests are local so that they do no go through the Internet, they do not use docker or a VM, etc. Obviously, if you run these benchmark on the Internet, you will get slower results due to the network overhead. Furthermore, my benchmark does not do any processing, it just sends simple messages. But I am interested in how low the latency can get.

I use Node.js as a runtime environment (version 20). There is an alternative JavaScript runtime environment called Bun which I also use for comparison (1.0.14). Because I have two JavaScript processes, I four possibilities: the two processes may run Node.js, or they may run bun, or a mixed of those.

Can we do better? Bun has its own WebSocket API. I wrote a script specifically for it. I am keeping the clients unchanged (i.e., I am not using a bun-specific client).

I measure the number of roundtrips per second in steady state.

Node.js 20 (client) Bun 1.0 (client)
Node.js 20 (server with ws) 19,000 23,000
Bun 1.0 (server with ws) 15,000 27,000
Bun 1.0 (bun-specific server) 44,000 50,000

It seems fair to compare the pure Node.js configuration (19,000) with the pure Bun configuration (27,000) when using the ws module. At least in my tests, I am getting that Node.js clients are faster when using a Node.js server. I am not sure why that is. Bun is 40% faster than Node.js in this one test. Once you switch to the bun-specific JavaScript code, then bun is twice as fast.

In a simple http benchmark, I got that Node.js could support about 45,000 http queries per second while bun while nearly twice as capable. However, to get these high numbers, we would have multiple requests in flight at all times. So while I am not making a direct comparison, it seems likely that WebSocket is more efficient than repeatedly polling the servers from both clients.

Importantly, all of my source code is available. The benchmark should be fully reproducible.

Science and Technology links (November 12 2023)

  1. Vitamin K2 supplements might reduce the risk of myocardial infarction (heart attacks) and of all-cause death (Hasific et al. 2022). You find vitamin K2 in some Gouda cheeses and in eggs.
  2. Most of the water on Earth is salinated (in the oceans) and cannot be consumed. Fresh water is often scarce. Yet Israel is desalinating water for less than a dollar per cubic meter.
  3. People living in South America engaged in warfare for 10,000 years before the arrival of the Europeans (Standen et al. 2023).
  4. The last glacial period ended about 12,000 years ago and lasted about 100,000 years. About 26,000 years ago, all of Canada was covered by a permanent ice sheet. Thus many of us were taught in school that human beings first colonized America about 12,000 years ago by the Bering land bridge, that existed back then between modern-day Russia and modern-day Alaska. The evidence accumulates that there were human beings in America much earlier than initially thougth. They would have been present 21,000 to 23,000 years ago in New Mexico. We even have their footprints.
  5. As recently as 20,000 years ago—not long in geological terms—Britain was not, in fact, an island. Instead, the terrain that became the British Isles was linked to mainland Europe by Doggerland, a tract of now-submerged territory where early Mesolithic hunter-gatherers lived, settled and traveled (McGreevy, 2020). Correspondly, there were human beings in Ireland 31,000 years ago.
  6. Gray et al. (2023) argue that the limited freedom that children enjoy in our modern societies is leading to a rise in mental disorders.
  7. Most people cannot understand the bat and ball problem, even after the solution is given. The problem can be stated as follows: “A bat and a ball cost $1.10 in total. The bat costs $1.00 more than the ball. How much does the ball cost?” ChatGPT can solve it:
  8. When hiring, we find a slight bias in favour of females in male-dominated fields, and a strong bias in favour of females in female-dominated fields (Schaerer et al., 2023). Overall, people greatly overestimate gender biases in hiring.
  9. Retinol, a common cosmetic product, keeps one’s skin younger.
  10. Unwarranted financial optimism might be the result of low cognitive abilities.

Generating arrays at compile-time in C++ with lambdas

Suppose that you want to check whether a character in C++ belongs to a fixed set, such as ‘\0’, ‘\x09’, ‘\x0a’,’\x0d’, ‘ ‘, ‘#’, ‘/’, ‘:’, ‘<‘, ‘>’, ‘?’, ‘@’, ‘[‘, ‘\\’, ‘]’, ‘^’, ‘|’. A simple way is to generate a 256-byte array of Boolean values and lookup the value. This approach is sometimes called memoization (and not memorization!!!). You might do it as follows:

constexpr static bool is_forbidden_host_code_point_table[] = {
  1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
  0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1,
  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1, 0, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0,
  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 0,
  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
  0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
bool is_forbidden_host_code_point(char c) {
  return is_forbidden_host_code_point_table[uint8_t(c)];

It is reasonably efficient in practice. Some people might object to how the table is generated. Can you have the C++ compiler generate the array at compile-time from a function?

Using C++17, you might do it with an std::array as follows:

constexpr static std::array<uint8_t, 256> is_forbidden_array = []() {
  std::array<uint8_t, 256> result{};
  for (uint8_t c : {'\0', '\x09', '\x0a','\x0d', ' ', '#', '/', ':',
    '<', '>', '?', '@', '[', '\\', ']', '^', '|'}) {
   result[c] = true;
  return result;

bool is_forbidden_host_code_point_array(char c) {
  return is_forbidden_array[uint8_t(c)];

These two approaches should be equivalent in practice. This might compile down to a single lookup instruction, or the equivalent.

You may want to compare it against the default (without memoization) which might be…

bool is_forbidden_host_code_point_default(char c) noexcept {
  return c == '\0' || c == '\x09' || c == '\x0a'
   || c == '\x0d' || c == ' ' || c == '#'
   || c == '/'|| c == ':' || c == '<'|| c == '>'
   || c == '?' || c == '@' || c == '['|| c == '\\'
  || c == ']'|| c == '^' || c == '|';

A compiler like GCC might compile this routine to a bitset approach such as …

 cmp dil, 62
ja .L2
movabs rax, 6052978675329017345
bt rax, rdi
setc al
sub edi, 63
cmp dil, 61
ja .L4
movabs rax, 2305843013240225795
bt rax, rdi
setc al

The default approach certainly generates more instructions, and might be less efficient in some cases.

Further reading. The evolution of constexpr: compile-time lookup tables in C++

Appending to an std::string character-by-character: how does the capacity grow?

In C++, suppose that you append to a string one character at a time:

while(my_string.size() <= 10'000'000) {
  my_string += "a";

In theory, it might be possible for the C++ runtime library to implement this routine as the creation of a new string with each append: it could allocate a new memory region that contains just one extra character, and copy to the new region. It would be very slow in the worst case. Of course, the people designing the runtime libraries are aware of such potential problem. Instead of allocating memory and copying with each append, they will typically grow the memory usage in bulk. That is, every time new memory is needed, they double the memory usage (for example).

Empirically, we can measure the allocation. Starting with an empty string, we may add one character at a time. I find that GCC 12 uses capacities of size 15 × 2 k for every increasing integers k, so that the string capacities are 15, 30, 60, 120, 240, 480, 960, 1920, etc. Under macOS (LLVM 15), I get that clang doubles the capacity and add one, except for the initial doubling, so you get capacities of 22, 47, 95, 191, 383, 767, etc. So the string capacity grows by successively doubling in all cases.

If you omit the cost of writing the character, what is the cost of these allocations and copy for long strings? Assume that allocating N bytes costs you N units of work. Let us consider the GCC 12 model : they both lead to the same conclusion. To construct a string of size up to 15 × 2n, it costs you 15 + 15 × 21 + 15 × 22 + … + 15 × 2n which is a geometric series with value 15 × (2n + 1 – 1). Generally speaking, you find that this incremental doubling approach costs you no more than 2N units of work to construct a string of size N, after rounding N up to a power of two.  In computer science parlance, the complexity is linear. Insertion in a dynamic array with capacity that is expanded by a constant factor ensures that inserting an element is constant time (amortized). In common sense parlance, it scales well.

In the general case, where you replace 2 by another value (e.g., 1.5), you still get a geometric series but it is in a different basis, 1 +  b1 + b2 + … +  bn  which sums up to  (bn+1-1)/(b – 1). The ratio 2, becomes  b/(b – 1) asymptotically, so for a basis of 1.5, you get 3N units of work instead of 2N. So a smaller scaling factor b leads to more work, but it is still just a constant factor.

If I benchmark the following function for various values of ‘volume’, I get a practically constant time-per-value:

std::string my_string;
while (my_string.size() <= volume) {
  my_string += "a";

On my laptop, I get the following results. You can run my benchmark yourself.

volume time/entry (direct measure)
100 5.83 ns
1000 5.62 ns
10000 5.47 ns
100000 5.62 ns
1000000 5.68 ns
10000000 5.69 ns
100000000 5.80 ns

A consequence of how strings allocate memory is that you may find that many of your strings have excess capacity if you construct them by repeatedly appending characters. To save memory, you may call the method shrink_to_fit() to remove this excess capacity. If you are using a temporary string, it is not a concern since the memory is recovered immediately.

For processing strings, streams in C++ can be slow

The C++ library has long been organized around stream classes, at least when it comes to reading and parsing strings. But streams can be surprisingly slow. For example, if you want to parse numbers, then this C++ routine is close to being the worst possible choice for performance:

std::stringstream in(mystring);
while(in >> x) {
   sum += x;
return sum;

I recently learned that some Node.js engineers prefer stream classes when building strings, for performance reasons. I am skeptical.

Let us run an experiment. We shall take strings containing the ‘%’ character and we build new strings where the ‘%’ character is replaced by ‘%25’ but the rest of the string is otherwise unchanged.

A straight-forward string construction is as follows:

std::string string_escape(const std::string_view file_path) {
  std::string escaped_file_path;
  for (size_t i = 0; i < file_path.length(); ++i) {
    escaped_file_path += file_path[i];
    if (file_path[i] == '%')
      escaped_file_path += "25";
  return escaped_file_path;

An optimized version using streams is as follows:

std::string stream_escape(const std::string_view file_path) {
  std::ostringstream escaped_file_path;
  for (size_t i = 0; i < file_path.length(); ++i) {
    escaped_file_path << file_path[i];
    if (file_path[i] == '%')
      escaped_file_path << "25";
  return escaped_file_path.str();

I envision using these functions over strings that contain few ‘%’ characters. It is possible that most of the strings do not contain the ‘%’. In such cases, I can just search for the character and only do non-trivial work when one is found. The following code should do:

std::string find_string_escape(std::string_view file_path) {
  // Avoid unnecessary allocations.
  size_t pos = file_path.empty() ? std::string_view::npos :
  if (pos == std::string_view::npos) {
   return std::string(file_path);
  // Escape '%' characters to a temporary string.
  std::string escaped_file_path;
  do {
    escaped_file_path += file_path.substr(0, pos + 1);
    escaped_file_path += "25";
    file_path = file_path.substr(pos + 1);
    pos = file_path.empty() ? std::string_view::npos :
  } while (pos != std::string_view::npos);
  escaped_file_path += file_path;
  return escaped_file_path;

I wrote a benchmark that uses a large collection of actual file URLs as a data source. The benchmark runs under macOS and Linux. I use Linux, a recent Intel server and GCC 12:

naive strings 260 ns/string 0.45 GB/s
stream 1000 ns/string 0.12 GB/s
find 33 ns/string 3.49 GB/s

At least in this case, I find that the stream version is four times slower than  naive string processing, and it is 30 times slower than the optimized ‘find’ approach.

Your results will vary depending on your system, but I generally consider the use of streams in C++ as a hint that there might be poor performance.

Further reading: I turned this blog post into a pull request to Node.js.

How many billions of transistors in your iPhone processor?

In about 10 years, Apple has multiplied by 19 the number of transistors in its mobile processors. It corresponds roughly to a steady rate of improvement of 34% per year on the number of transistors, or a doubling every 2.5 years. In real dollars, an iPhone has roughly a constant price: the price tag of a new iPhone increases every year, but it does so while tracking the inflation. Thus you are getting ever more transistors in your iPhone for the same price.

processor release year transistors
Apple A7 2013 1 billion
Apple A8 2014 2 billions
Apple A9 2015 2 billions
Apple A10 2016 3.2 billions
Apple A11 2017 4.3 billions
Apple A12 2018 6.9 billions
Apple A13 2019 8.5 billions
Apple A14 2020 11.8 billions
Apple A15 2021 15 billions
Apple A16 2022 16 billions
Apple A17 2023 19 billions