typedarray-cache-experiment

An experiment comparing different caching strategies for typed arrays. The first strategy creates an array buffer, while the second preallocates all typed array variants.

Results

$ time node nocache_test.js

  real	0m9.860s
  user	0m9.816s
  sys	0m0.097s
  
$ time node buffer_test.js

  real	0m6.974s
  user	0m6.974s
  sys	0m0.036s

$ time node prealloc_test.js a

  real	0m0.335s
  user	0m0.327s
  sys	0m0.009s
  
$ time node array_test.js

  real	0m0.545s
  user	0m0.537s
  sys	0m0.009s
  
$ time node tapool_test.js

  real	0m0.493s
  user	0m0.485s
  sys	0m0.009s

Conclusion

Caching without preallocating typed arrays is not really worth it. So there are really only two credible choices:

• preallocating arrays in batches
• preallocating a pool for each typed array

The batch preallocation is faster, has better overall cache performance and uses less memory, but the separate pool per typed array has a simpler interface. It isn't clear to me what the best solution here is. Even though it is slower, I think that the array pool has a lot going for it AND has the advantage that it could easily be adapted to work with arrays that are allocated outside the main pool of objects (for example, caching typed arrays used for communication with web workers).