5.6 KiB
Performance
Overall, the performance of this library is positioned roughly in the middle between the fast (effectful, eveff, etc.) and slow (polysemy, fused-effects, etc.) libraries, and can be considered average.
In all benchmarks, the speed is nearly equivalent to freer-simple, only slightly slower.
In the State benchmark, it is about 3-6x faster than polysemy and about 2-7x slower than effectful.
In the Throw/Catch benchmark, it is slightly faster than effectful in the fastest case, and at least 2x as fast as polysemy in the other cases.
In the Ask/Local benchmark, it still outperforms polysemy and fused-effects in deep effect stack cases.
Non-deterministic computations are somewhat slow;
in cases with a shallow effect stack, they are about 5x slower than fused-effects.
However, because their performance in terms of stack depth is better than that of fused-effects,
they catch up to fused-effects in speed at around a depth of 5.
Since among the practical extensible effects libraries, only fused-effects and freer-simple support non-deterministic computations,
this shouldn't be too much of a problem.
The coroutine performance is on par with freer-simple, maintaining similar speed without any noticeable degradation compared to other effects.
Moreover, the other libraries such as eff and mpeff, aside from this library and freer-simple, likely suffer from the O(n^2) issue mentioned in the "Reflection without Remorse" paper.
In comparison, these libraries are significantly slower, approximately 50x slower with n=1000 (as shown in the measured results) and about 500x slower with n=10000.
Furthermore, since there are differences among libraries in their support for higher-order effects and continuations, please note that, for example, in the Throw/Catch benchmark, only libraries that support higher-order effects are included in the comparison.
Below is a table showing the speedup factor of heftia when mtl is set to 1x (mtl's computation time divided by heftia's computation time).
Larger values indicate that heftia is faster.
| Benchmark | Speedup |
|---|---|
| state, shallow | 2.0x |
| state, deep | 6.6x |
| throw/catch, shallow | 0.3x |
| throw/catch, deep | 1.2 - 6.5x |
| nondet, shallow | 0.2x |
| nondet, deep | 0.7x |
Reproduction
The benchmark code is available at heftia-effects/bench.
To run the benchmarks, move your working directory to the root directory of the heftia repository and execute
./benchmark/bench.sh.
Benchmark Results
The code was compiled with GHC 9.8.2 and run on a Ryzen 9 3900XT.
-
State Effect Benchmark (Shallow Effect Stack):
-
State Effect Benchmark (Deep Effect Stack):
-
Throw/Catch Effect Benchmark (Shallow Effect Stack):
-
Throw/Catch Effect Benchmark (Deep Effect Stack):
Note: Here, the 5 + (5 - N) + N corresponds to the interpretation stack of catchHeftiaDeepN represented below:
catchHeftiaDeep0, catchHeftiaDeep1, catchHeftiaDeep2, catchHeftiaDeep3, catchHeftiaDeep4, catchHeftiaDeep5 :: Int -> Either () ()
catchHeftiaDeep0 n = Heftia.runPure $ run $ run $ run $ run $ run $ Heftia.runThrow $ run $ run $ run $ run $ run $ Heftia.runCatch @() $ programHeftia n
catchHeftiaDeep1 n = Heftia.runPure $ run $ run $ run $ run $ run $ Heftia.runThrow $ run $ run $ run $ run $ Heftia.runCatch @() $ run $ programHeftia n
catchHeftiaDeep2 n = Heftia.runPure $ run $ run $ run $ run $ run $ Heftia.runThrow $ run $ run $ run $ Heftia.runCatch @() $ run $ run $ programHeftia n
catchHeftiaDeep3 n = Heftia.runPure $ run $ run $ run $ run $ run $ Heftia.runThrow $ run $ run $ Heftia.runCatch @() $ run $ run $ run $ programHeftia n
catchHeftiaDeep4 n = Heftia.runPure $ run $ run $ run $ run $ run $ Heftia.runThrow $ run $ Heftia.runCatch @() $ run $ run $ run $ run $ programHeftia n
catchHeftiaDeep5 n = Heftia.runPure $ run $ run $ run $ run $ run $ Heftia.runThrow $ Heftia.runCatch @() $ run $ run $ run $ run $ run $ programHeftia n
run :: Heftia.Eff eh (Heftia.Ask () ': ef) a -> Heftia.Eff eh ef a
run = Heftia.runAsk ()
-
Ask/Local Effect Benchmark (Shallow Effect Stack):
-
Ask/Local Effect Benchmark (Deep Effect Stack):
-
NonDet Effect Benchmark (Shallow Effect Stack):
-
NonDet Effect Benchmark (Deep Effect Stack):
-
Coroutine Benchmark (Shallow Effect Stack):
-
Coroutine Benchmark (Deep Effect Stack):
