Javascript Object vs Map/Set key lookup performance

Question

I was trying to explore how javascript Object performs in comparison to Map or Set for normal key accesses.

So i ran the below 3 codes on JSBEN.CH (here is the link to test: https://jsben.ch/q4RPK)

Objects

  const object = {};
  
  for (let i = 0; i < 10000; ++i) {
    object[`key_${i}`] = 1;
  }
  
  let result = 0;
  
  for (let i = 0; i < 10000; ++i) {
    result += object[`key_${i}`];
  }

Maps

  const map = new Map();
  
  for (let i = 0; i < 10000; ++i) {
    map.set(`key_${i}`, 1);
  }
  
  let result = 0;
  
  for (let i = 0; i < 10000; ++i) {
    result += map.get(`key_${i}`);
  }

Sets

  const set = new Set();
  
  for (let i = 0; i < 10000; ++i) {
    set.add(`key_${i}`);
  }
  
  let result = 0;
  
  for (let i = 0; i < 10000; ++i) {
    result += set.has(`key_${i}`);
  }

As you can check in the test link, map and set seem to perform almost similar however objects are much slower every time. Can someone explain what could be the reason that objects perform worse than map or set for basic key access operation?

Edit 1: As it turns out, just setting keys on objects is also slower than map/set: https://jsben.ch/al9ef

Answer 1

Looking at relative numbers only is always dangerous, here are some absolute numbers, run on NodeJS v14.14.0 on an Intel 8350U:

Iterations	Object write	Object read	Map write	Map read
100	0ms	0ms	0ms	0ms
1.000	3ms	1ms	0ms	0ms
10,000	7ms	4ms	8ms	1ms
1.000.000	1222ms	527ms	632ms	542ms

So as one can see, for 10.000 iterations the difference between objects and maps is 1 millisecond in the run above, and as that is the accuracy of the time measurement, we can't really derive any conclusion from that test. The results are absolutely random.

For 1 Million iterations one can see a clear advantage of Map writes over Object writes, the read performance is very similar. Now if we look at absolute numbers, this is still one million writes / s. So although object writes are a lot slower, this will unlikely be the bottleneck of your application.

For an accurate explanation, one would have to analyze all the steps the engine performs. For that you can run node --print-code and analyze the bytecode that gets run. I don't have the time for that, though here are some observations:

1. If the object gets constructed with Object.create(null) (having no prototype) the performance is about the same, so prototype lookup does not influence performance at all.

2. After the 20th iteration, V8 chooses the internal representation dictionary_map for object, so this is basically one hash map competing with another hashmap (one can run node --allow-natives-syntax and then use %DebugPrint(object) to get the internal representation).

For reference, here is the code used to run the benchmark:

function benchmark(TIMES) {  
  console.log("BENCHMARK ", TIMES);

  const object = Object.create(null);

    let start = Date.now();
    for (let i = 0; i < TIMES; ++i) {
        object[`key_${i}`] = 1;
    }

    console.log("Object write took", Date.now() - start);
    start = Date.now();

    let result = 0;
  
    for (let i = 0; i < TIMES; ++i) {
      result += object[`key_${i}`];
    }

    console.log("Object read took", Date.now() - start);
    start = Date.now();

  


  const map = new Map();
  
  for (let i = 0; i < TIMES; ++i) {
    map.set(`key_${i}`, 1);
  }
  
  console.log("Map write took", Date.now() - start);
  start = Date.now();

  result = 0;
  
  for (let i = 0; i < TIMES; ++i) {
    result += map.get(`key_${i}`);
  }

  console.log("Map read took", Date.now() - start);

}

benchmark(100);
benchmark(1_000);
benchmark(10_000);
benchmark(1_000_000);