Почему быстрее суммировать Data.Sequence методом «разделяй и властвуй», даже без параллелизма?

Я играл с параллельным сокращением на Data.Sequence.Seq и заметил, что разделяй и властвуй дает преимущество в скорости даже без параллелизма. Кто-нибудь знает, почему?

Вот мой код:

import qualified Data.Sequence as S
import qualified Data.Foldable as F

import System.Random
import Control.DeepSeq
import Criterion.Main
import Test.QuickCheck
import Control.Exception ( evaluate )

instance (Arbitrary a) => Arbitrary (S.Seq a) where
    arbitrary = fmap S.fromList arbitrary

instance NFData a => NFData (S.Seq a) where
    rnf = F.foldr seq ()

funs :: [(String, S.Seq Int -> Int)]
funs =
    [ ("seqDirect"   , seqDirect)
    , ("seqFoldr"    , seqFoldr)
    , ("seqFoldl'"   , seqFoldl')
    , ("seqSplit  1" , (seqSplit  1))
    , ("seqSplit  2" , (seqSplit  2))
    , ("seqSplit  4" , (seqSplit  4))
    , ("seqSplit  8" , (seqSplit  8))
    , ("seqSplit 16" , (seqSplit 16))
    , ("seqSplit 32" , (seqSplit 32)) ]

main :: IO ()
main = do
    mapM_ (\(_,f) -> quickCheck (\xs -> seqDirect xs == f xs)) funs
    gen <- newStdGen
    let inpt = S.fromList . take 100000 $ randoms gen
    evaluate (rnf inpt)
    defaultMain [ bench n (nf f inpt) | (n,f) <- funs ]

seqDirect :: S.Seq Int -> Int
seqDirect v = case S.viewl v of
    S.EmptyL -> 0
    x S.:< xs -> x + seqDirect xs

seqFoldr :: S.Seq Int -> Int
seqFoldr = F.foldr (+) 0

seqFoldl' :: S.Seq Int -> Int
seqFoldl' = F.foldl' (+) 0

seqSplit :: Int -> S.Seq Int -> Int
seqSplit 1 xs = seqFoldr xs
seqSplit _ xs | S.null xs = 0
seqSplit n xs =
    let (a, b) = S.splitAt (S.length xs `div` n) xs
        sa = seqFoldr a
        sb = seqSplit (n-1) b
    in  sa + sb

И результаты:

$ ghc -V
The Glorious Glasgow Haskell Compilation System, version 7.0.4

$ ghc --make -O2 -fforce-recomp -rtsopts seq.hs
[1 of 1] Compiling Main             ( seq.hs, seq.o )
Linking seq ...

$ ./seq +RTS -s
./seq +RTS -s
+++ OK, passed 100 tests.
+++ OK, passed 100 tests.
+++ OK, passed 100 tests.
+++ OK, passed 100 tests.
+++ OK, passed 100 tests.
+++ OK, passed 100 tests.
+++ OK, passed 100 tests.
+++ OK, passed 100 tests.
+++ OK, passed 100 tests.
warming up
estimating clock resolution...
mean is 5.882556 us (160001 iterations)
found 2368 outliers among 159999 samples (1.5%)
  2185 (1.4%) high severe
estimating cost of a clock call...
mean is 85.26448 ns (44 iterations)
found 4 outliers among 44 samples (9.1%)
  3 (6.8%) high mild
  1 (2.3%) high severe

benchmarking seqDirect
mean: 23.37511 ms, lb 23.01101 ms, ub 23.77594 ms, ci 0.950
std dev: 1.953348 ms, lb 1.781578 ms, ub 2.100916 ms, ci 0.950

benchmarking seqFoldr
mean: 25.60206 ms, lb 25.39648 ms, ub 25.80034 ms, ci 0.950
std dev: 1.030794 ms, lb 926.7246 us, ub 1.156656 ms, ci 0.950

benchmarking seqFoldl'
mean: 10.65757 ms, lb 10.29087 ms, ub 10.99869 ms, ci 0.950
std dev: 1.819595 ms, lb 1.703732 ms, ub 1.922018 ms, ci 0.950

benchmarking seqSplit  1
mean: 25.50376 ms, lb 25.29045 ms, ub 25.71225 ms, ci 0.950
std dev: 1.075497 ms, lb 961.5707 us, ub 1.229739 ms, ci 0.950

benchmarking seqSplit  2
mean: 18.15032 ms, lb 17.62943 ms, ub 18.66413 ms, ci 0.950
std dev: 2.652232 ms, lb 2.288088 ms, ub 3.044585 ms, ci 0.950

benchmarking seqSplit  4
mean: 10.48334 ms, lb 10.14152 ms, ub 10.87061 ms, ci 0.950
std dev: 1.869274 ms, lb 1.690063 ms, ub 1.997915 ms, ci 0.950

benchmarking seqSplit  8
mean: 5.737956 ms, lb 5.616747 ms, ub 5.965689 ms, ci 0.950
std dev: 825.2361 us, lb 442.1652 us, ub 1.232003 ms, ci 0.950

benchmarking seqSplit 16
mean: 3.677038 ms, lb 3.669035 ms, ub 3.685547 ms, ci 0.950
std dev: 42.18741 us, lb 36.57112 us, ub 49.93574 us, ci 0.950

benchmarking seqSplit 32
mean: 2.855626 ms, lb 2.849962 ms, ub 2.862226 ms, ci 0.950
std dev: 31.25475 us, lb 26.49104 us, ub 37.18611 us, ci 0.950
  25,154,069,064 bytes allocated in the heap
   4,120,506,464 bytes copied during GC
      32,344,120 bytes maximum residency (446 sample(s))
       4,042,704 bytes maximum slop
              78 MB total memory in use (0 MB lost due to fragmentation)

  Generation 0: 42092 collections,     0 parallel,  6.57s,  6.57s elapsed
  Generation 1:   446 collections,     0 parallel,  2.62s,  2.62s elapsed

  INIT  time    0.00s  (  0.00s elapsed)

  MUT   time   18.57s  ( 18.58s elapsed)
  GC    time    9.19s  (  9.19s elapsed)
  EXIT  time    0.00s  (  0.00s elapsed)
  Total time   27.76s  ( 27.77s elapsed)

  %GC time      33.1%  (33.1% elapsed)

  Alloc rate    1,354,367,579 bytes per MUT second

  Productivity  66.9% of total user, 66.9% of total elapsed