Awesome

Exploring Ruby FFI and SIMD instructions

This repository explores how to do some basic FFI in ruby using Ruby FFI, as well as how SIMD instructions work. Using some basic descriptive statistics as a test case, the results are pretty interesting!

• Ruby FFI makes calling into C code relatively simple. FFI lets you write C/C++ code which exports the desired interface without really caring that it is called from ruby. Wrapping the native interface with FFI objects using the FFI DSL is a lot nicer compared to doing it directly in C with the ruby C API. This also applies to calling any arbitrary C/C++ libraries. In addition, being able to "enrich" arbitrary C structs with ruby methods directly in ruby is also really useful (although this is equally possible with the C API).
• On the statistics computation example here, ruby can perform reasonably well with well written code that can calculate multuiple statistics on every loop through the data. Libraries like descriptive_statistics don't really reuse previously computed statistics, and so have to re-calculate them in repeated loops which makes them orders of magnitude slower.
• Native code with -O3 performing the same loop as ruby is much much faster, as expected. The speed up is ~25x!
• Explicitly writing the computations out using 128-bit SIMD intrinsics squeezes out an extra ~30% performance from the native code. This is great, but not really close to the theoretical 4x that you might expect. Presumably -O3 is already doing quite well to optimise the generated code, so the explicit intrinsics don't make that big of a difference in the end. Still, 30% is a decent return for not too much work.

Benchmark Results

Comparing calculated statistics with 10 values...
Ruby (Custom):
+----------+----------+----------+----------+--------------------+
| mean     | min      | max      | variance | standard_deviation |
+----------+----------+----------+----------+--------------------+
| 0.657200 | 0.018774 | 0.999315 | 0.098239 | 0.313432           |
| 0.494228 | 0.103677 | 0.922196 | 0.093787 | 0.306247           |
| 0.435724 | 0.071016 | 0.859478 | 0.060833 | 0.246644           |
| 0.399854 | 0.212832 | 0.711103 | 0.033990 | 0.184363           |
+----------+----------+----------+----------+--------------------+
Ruby (Desc Stats):
+----------+----------+----------+----------+--------------------+
| mean     | min      | max      | variance | standard_deviation |
+----------+----------+----------+----------+--------------------+
| 0.657200 | 0.018774 | 0.999315 | 0.098239 | 0.313432           |
| 0.494228 | 0.103677 | 0.922196 | 0.093787 | 0.306247           |
| 0.435724 | 0.071016 | 0.859478 | 0.060833 | 0.246644           |
| 0.399854 | 0.212832 | 0.711103 | 0.033990 | 0.184363           |
+----------+----------+----------+----------+--------------------+
Native:
+----------+----------+----------+----------+--------------------+
| mean     | min      | max      | variance | standard_deviation |
+----------+----------+----------+----------+--------------------+
| 0.657200 | 0.018774 | 0.999315 | 0.098239 | 0.313432           |
| 0.494228 | 0.103677 | 0.922196 | 0.093787 | 0.306247           |
| 0.435724 | 0.071016 | 0.859478 | 0.060833 | 0.246644           |
| 0.399854 | 0.212832 | 0.711103 | 0.033990 | 0.184363           |
+----------+----------+----------+----------+--------------------+
Native (Simd):
+----------+----------+----------+----------+--------------------+
| mean     | min      | max      | variance | standard_deviation |
+----------+----------+----------+----------+--------------------+
| 0.657200 | 0.018774 | 0.999315 | 0.098239 | 0.313432           |
| 0.494228 | 0.103677 | 0.922196 | 0.093787 | 0.306247           |
| 0.435724 | 0.071016 | 0.859478 | 0.060833 | 0.246644           |
| 0.399854 | 0.212832 | 0.711103 | 0.033990 | 0.184363           |
+----------+----------+----------+----------+--------------------+
Test passed, results are equal to 6 decimal places!

Benchmarking with 100,000 values...
Warming up --------------------------------------
   Ruby (Desc Stats)     1.000  i/100ms
       Ruby (Custom)     1.000  i/100ms
              Native    39.000  i/100ms
       Native (Simd)    48.000  i/100ms
Calculating -------------------------------------
   Ruby (Desc Stats)      3.404  (± 0.0%) i/s -     17.000  in   5.005342s
       Ruby (Custom)     15.779  (± 6.3%) i/s -     79.000  in   5.024171s
              Native    356.612  (± 5.6%) i/s -      1.794k in   5.046749s
       Native (Simd)    461.957  (± 5.6%) i/s -      2.304k in   5.003756s

Comparison:
       Native (Simd):      462.0 i/s
              Native:      356.6 i/s - 1.30x  (± 0.00) slower
       Ruby (Custom):       15.8 i/s - 29.28x  (± 0.00) slower
   Ruby (Desc Stats):        3.4 i/s - 135.72x  (± 0.00) slower

Running the benchmark

Make sure to have gcc installed. Then you can run bundle install to install the ruby dependecies, and make to run the benchmark.