Strange results when benchmarking FFI bindings using criterion library in Haskell

copy.c

#include <stdlib.h>
#include "copy.h"

double* c_copy( double* inArr, int arrLen ) {
  double* outArr = malloc( arrLen * sizeof( double ) );

  for ( int i = 0; i < arrLen; i++ ) {
    outArr[ i ] = inArr[ i ];
  }

  return outArr;
}

copy.h

#ifndef _COPY_H_
#define _COPY_H_

double* c_copy( double*, int );

#endif

ffi_crit.hs

module Main where

import Criterion.Main
import Data.Vector.Storable hiding (copy)
import Control.Monad (liftM)
import Foreign hiding (unsafePerformIO)
import Foreign.C
import System.IO.Unsafe (unsafePerformIO)


foreign import ccall unsafe "copy.h"
  c_copy  :: Ptr CDouble -> CInt -> IO (Ptr CDouble)


signal :: Vector Double
signal = fromList [1.0..16384.0]


copy :: Vector Double -> Vector Double
copy sig = unsafePerformIO $ do
    let (fpSig, _, lenSig) = unsafeToForeignPtr sig
    pLattice <- liftM castPtr $ withForeignPtr fpSig $ \ptrSig ->
                c_copy (castPtr ptrSig) (fromIntegral lenSig)
    fpLattice <- newForeignPtr finalizerFree pLattice
    return $ unsafeFromForeignPtr0 fpLattice lenSig


main :: IO ()
main = defaultMain [
         bgroup "FFI" [
           bench "C binding" $ whnf copy signal
         , bench "C binding" $ whnf copy signal
         , bench "C binding" $ whnf copy signal
         , bench "C binding" $ whnf copy signal
         , bench "C binding" $ whnf copy signal
         , bench "C binding" $ whnf copy signal
         , bench "C binding" $ whnf copy signal
         , bench "C binding" $ whnf copy signal
         , bench "C binding" $ whnf copy signal
         ]
       ]

results.txt

$ ghc -O2 -Wall -optc -std=c99 ffi_crit.hs copy.c
$ ./ffi_crit -g
warming up
estimating clock resolution...
mean is 1.431904 us (640001 iterations)
found 4750 outliers among 639999 samples (0.7%)
  3581 (0.6%) high severe
estimating cost of a clock call...
mean is 45.00523 ns (10 iterations)
found 1 outliers among 10 samples (10.0%)
  1 (10.0%) high severe

benchmarking FFI/C binding
mean: 17.44777 us, lb 16.82549 us, ub 19.84387 us, ci 0.950
std dev: 5.627304 us, lb 968.1911 ns, ub 13.18222 us, ci 0.950
found 5 outliers among 100 samples (5.0%)
  5 (5.0%) high severe
variance introduced by outliers: 97.856%
variance is severely inflated by outliers

benchmarking FFI/C binding
mean: 45.46269 us, lb 45.17545 us, ub 46.01435 us, ci 0.950
std dev: 1.950915 us, lb 1.169448 us, ub 3.201935 us, ci 0.950
found 8 outliers among 100 samples (8.0%)
  4 (4.0%) high mild
  4 (4.0%) high severe
variance introduced by outliers: 40.503%
variance is moderately inflated by outliers

benchmarking FFI/C binding
mean: 45.79727 us, lb 45.55681 us, ub 46.26911 us, ci 0.950
std dev: 1.669191 us, lb 1.029116 us, ub 3.098384 us, ci 0.950
found 5 outliers among 100 samples (5.0%)
  2 (2.0%) high mild
  3 (3.0%) high severe
variance introduced by outliers: 32.640%
variance is moderately inflated by outliers

benchmarking FFI/C binding
mean: 46.71757 us, lb 46.32679 us, ub 47.29715 us, ci 0.950
std dev: 2.393483 us, lb 1.821208 us, ub 3.336483 us, ci 0.950
found 10 outliers among 100 samples (10.0%)
  6 (6.0%) high mild
  4 (4.0%) high severe
variance introduced by outliers: 49.452%
variance is moderately inflated by outliers

benchmarking FFI/C binding
mean: 46.39661 us, lb 45.99686 us, ub 47.02149 us, ci 0.950
std dev: 2.526114 us, lb 1.830455 us, ub 3.693685 us, ci 0.950
found 9 outliers among 100 samples (9.0%)
  5 (5.0%) high mild
  4 (4.0%) high severe
variance introduced by outliers: 52.463%
variance is severely inflated by outliers

benchmarking FFI/C binding
mean: 46.60653 us, lb 46.20746 us, ub 47.35528 us, ci 0.950
std dev: 2.708197 us, lb 1.660032 us, ub 4.443404 us, ci 0.950
found 8 outliers among 100 samples (8.0%)
  4 (4.0%) high mild
  4 (4.0%) high severe
variance introduced by outliers: 55.497%
variance is severely inflated by outliers

benchmarking FFI/C binding
mean: 47.10585 us, lb 46.75674 us, ub 47.87401 us, ci 0.950
std dev: 2.536477 us, lb 1.191372 us, ub 4.499381 us, ci 0.950
found 5 outliers among 100 samples (5.0%)
  2 (2.0%) high mild
  3 (3.0%) high severe
variance introduced by outliers: 51.486%
variance is severely inflated by outliers

benchmarking FFI/C binding
mean: 46.13946 us, lb 45.84257 us, ub 46.66114 us, ci 0.950
std dev: 1.955241 us, lb 1.280196 us, ub 3.142932 us, ci 0.950
found 5 outliers among 100 samples (5.0%)
  4 (4.0%) high severe
variance introduced by outliers: 39.537%
variance is moderately inflated by outliers

benchmarking FFI/C binding
mean: 46.33848 us, lb 46.08088 us, ub 46.83814 us, ci 0.950
std dev: 1.784367 us, lb 1.018966 us, ub 2.972920 us, ci 0.950
found 6 outliers among 100 samples (6.0%)
  3 (3.0%) high mild
  3 (3.0%) high severe
variance introduced by outliers: 35.557%
variance is moderately inflated by outliers

Notice that first run of a benchmark takes about 2,5 times shorter than others. This happens most of the times, though sometimes all benchmarks take about 45 us. Where does this extra time come from?