The charts in the Effects for Less talk shows
| library | time (μs) |
|---|---|
| eff | 125 |
| mtl | 640 |
| fused | 1162 |
So the first question is: what benchmark is Alexis talking about here? She didn’t explicitly mention it (or I didn’t catch it), but we can infer from the context that it’s likely to be the countdown benchmark. Then, what is the n she used? I tested n = 10000 on my machine and got the following results:
| library | time (μs) |
|---|---|
| mtl | 569 |
| fused | 1140 |
which is pretty close if you ask me. How, then, did eveff and my proof-of-concept sp do?
| library | time (μs) |
|---|---|
| sp | 203 |
| ev | 384 |
| mtl | 569 |
| fused | 1140 |
sp is only less than 1x slower than eff! Note that sp uses the same Ctl monad as eveff, and I only improved on the evidence-passing part. So, yes, even with Ctl you can have an 80% performance improvement, at least in this benchmark.
This is obviously not the whole story, and It’d be very interesting to see how eff performs in the other benchmarks.