Why is boxed mutable array so slow?

Branimir Maksimovic

1 Dec 2012 1 Dec '12

11:09 a.m.

I have made benchmark test inspired by http://lemire.me/blog/archives/2012/07/23/is-cc-worth-it/ What surprised me is that unboxed array is much faster than boxed array.Actually boxed array performance is on par with standard Haskell listwhich is very slow.All in all even unboxed array is about 10 times slower than Java version.I don't understand why is even unboxed array so slow.But! unboxed array consumes least amount of RAM.(warning, program consumes more than 3gb of ram) bmaxa@maxa:~/examples$ time ./Cumulboxed arraylast 262486571 seconds 4.972unboxed arraylast 262486571 seconds 0.776listlast 262486571 seconds 6.812 real 0m13.086suser 0m11.996ssys 0m1.080s -------------------------------------------------------------------------{-# LANGUAGE CPP, BangPatterns #-}import System.CPUTimeimport Text.Printfimport Data.Array.IOimport Data.Array.Baseimport Data.Intimport Control.DeepSeqimport System.Mem main :: IO()main = do (newArray_ (0,n'-1) :: IO(A)) >>= test "boxed array" performGC (newArray_ (0,n'-1) :: IO(B)) >>= test "unboxed array" performGC begin <- getCPUTime printf "list\nlast %d" $ last $ force $ take n' $ sum' data' end <- getCPUTime let diff = (fromIntegral (end - begin)) / (10^12) printf " seconds %.3f\n" (diff::Double) test s a = do putStrLn s begin <- getCPUTime init' a partial_sum a end <- getCPUTime let diff = (fromIntegral (end - begin)) / (10^12) last <- readArray a (n'-1) printf "last %d seconds %.3f\n" last (diff::Double) n' :: Intn' = 50 * 1000 * 1000 type A = IOArray Int Int32type B = IOUArray Int Int32 init' a = do (_,n) <- getBounds a init a 0 n where init a k n | k > n = return () | otherwise = do let !x = fromIntegral $ k + k `div` 3 unsafeWrite a k x init a (k+1) n partial_sum a = do (_,n) <- getBounds a k <- unsafeRead a 0 ps a 1 n k where ps a i n s | i > n = return () | otherwise = do k <- unsafeRead a i let !l = fromIntegral $ s + k unsafeWrite a i l ps a (i+1) n l data' :: [Int32]data' = [k + k `div` 3 | k <- [0..] ] sum' = scanl1 (+)

Attachments:

attachment.html (text/html — 5.6 KB)

Show replies by date

KC

1 Dec 1 Dec

11:30 a.m.

Boxed arrays have a wrapper (extra layer of indirection) to allow for a fully evaluated value, an unevaluated thunk, or the special value bottom (a value that can contain bottom is referred to as lifted). Unboxed arrays always have some value; that is, they cannot represent a thunk nor bottom. On Sat, Dec 1, 2012 at 8:09 AM, Branimir Maksimovic wrote:

...

I have made benchmark test inspired by http://lemire.me/blog/archives/2012/07/23/is-cc-worth-it/

What surprised me is that unboxed array is much faster than boxed array. Actually boxed array performance is on par with standard Haskell list which is very slow. All in all even unboxed array is about 10 times slower than Java version. I don't understand why is even unboxed array so slow. But! unboxed array consumes least amount of RAM. (warning, program consumes more than 3gb of ram)

bmaxa@maxa:~/examples$ time ./Cumul boxed array last 262486571 seconds 4.972 unboxed array last 262486571 seconds 0.776 list last 262486571 seconds 6.812

real 0m13.086s user 0m11.996s sys 0m1.080s

------------------------------------------------------------------------- {-# LANGUAGE CPP, BangPatterns #-} import System.CPUTime import Text.Printf import Data.Array.IO import Data.Array.Base import Data.Int import Control.DeepSeq import System.Mem

main :: IO() main = do (newArray_ (0,n'-1) :: IO(A)) >>= test "boxed array" performGC (newArray_ (0,n'-1) :: IO(B)) >>= test "unboxed array" performGC begin <- getCPUTime printf "list\nlast %d" $ last $ force $ take n' $ sum' data' end <- getCPUTime let diff = (fromIntegral (end - begin)) / (10^12) printf " seconds %.3f\n" (diff::Double)

test s a = do putStrLn s begin <- getCPUTime init' a partial_sum a end <- getCPUTime let diff = (fromIntegral (end - begin)) / (10^12) last <- readArray a (n'-1) printf "last %d seconds %.3f\n" last (diff::Double)

n' :: Int n' = 50 * 1000 * 1000

type A = IOArray Int Int32 type B = IOUArray Int Int32

init' a = do (_,n) <- getBounds a init a 0 n where init a k n | k > n = return () | otherwise = do let !x = fromIntegral $ k + k `div` 3 unsafeWrite a k x init a (k+1) n

partial_sum a = do (_,n) <- getBounds a k <- unsafeRead a 0 ps a 1 n k where ps a i n s | i > n = return () | otherwise = do k <- unsafeRead a i let !l = fromIntegral $ s + k unsafeWrite a i l ps a (i+1) n l

data' :: [Int32] data' = [k + k `div` 3 | k <- [0..] ]

sum' = scanl1 (+)

_______________________________________________ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe

-- -- Regards, KC

Don Stewart

12:20 p.m.

The obvious difference between boxed and unboxed arrays is that the boxed arrays are full of pointers to heap allocated objects. This means you pay indirection to access the values, much more time in GC spent chasing pointers (though card marking helps), and generally do more allocation. Compare the GC stats below, for * Boxed vector: 88M bytes copied; 75% of time in GC, 0.472s * Unboxed vector: 11k bytes copied, 1.3% of time in GC, 0.077s So there's your main answer. The increased data density of unboxed arrays also helps a too. Now, you can help out the GC signifcantly by hinting at how much you're going to allocated in the youngest generation (see the ghc-gc-tune app for a methodical approach to this, though it needs updating to ghc 7 -- http://donsbot.wordpress.com/2010/07/05/ghc-gc-tune-tuning-haskell-gc-settin... and http://stackoverflow.com/questions/3171922/ghcs-rts-options-for-garbage-coll... ). Use the +RTS -A flag to set an initial youngest generation heap size to the size of your array, and watch the GC cost disappear. For our boxed vector, we'd use +RTS -A50M, resulting in: * Boxed vector: 8k copied, 1% of time in GC, 0.157s So not bad. 3x speedup through a RTS flag. -A is very useful if you are working with boxed, mutable arrays. For reference, there's a generic version below that specializes based on the vector type parameter. --------------------------------- {-# LANGUAGE BangPatterns #-} import System.CPUTime import Text.Printf import Data.Int import Control.DeepSeq import System.Mem import qualified Data.Vector.Mutable as V import qualified Data.Vector.Unboxed.Mutable as U import qualified Data.Vector.Generic.Mutable as G main :: IO() main = do -- (G.new n' :: IO (V.IOVector Int32)) >>= test' "boxed vector" -- performGC (G.new n' :: IO (U.IOVector Int32)) >>= test' "unboxed vector" performGC test' s a = do putStrLn s begin <- getCPUTime init'' a partial_sum' a end <- getCPUTime let diff = (fromIntegral (end - begin)) / (10**12) last <- G.read a (n'-1) printf "last %d seconds %.3f\n" last (diff::Double) n' :: Int n' = 1000 * 1000 init'' !a = init 0 (n'-1) where init :: Int -> Int -> IO () init !k !n | k > n = return () | otherwise = do let !x = fromIntegral $ k + k `div` 3 G.write a k x init (k+1) n partial_sum' !a = do k <- G.read a 0 ps 1 (n'-1) k where ps :: Int -> Int -> Int32 -> IO () ps i n s | i > n = return () | otherwise = do k <- G.read a i let !l = fromIntegral $ s + k G.write a i l ps (i+1) n l --------------------------------- $ time ./A +RTS -s boxed vector last 945735787 seconds 0.420 40,121,448 bytes allocated in the heap 88,355,272 bytes copied during GC 24,036,456 bytes maximum residency (6 sample(s)) 380,632 bytes maximum slop 54 MB total memory in use (0 MB lost due to fragmentation) %GC time 75.2% (75.9% elapsed) Alloc rate 359,655,602 bytes per MUT second ./A +RTS -s 0.40s user 0.07s system 98% cpu 0.475 total $ time ./A +RTS -s unboxed vector last 945735787 seconds 0.080 4,113,568 bytes allocated in the heap 11,288 bytes copied during GC 4,003,256 bytes maximum residency (3 sample(s)) 182,856 bytes maximum slop 5 MB total memory in use (0 MB lost due to fragmentation) %GC time 1.3% (1.3% elapsed) Alloc rate 51,416,660 bytes per MUT second ./A +RTS -s 0.08s user 0.01s system 98% cpu 0.088 total $ time ./A +RTS -A50M -s boxed vector last 945735787 seconds 0.127 40,121,504 bytes allocated in the heap 8,032 bytes copied during GC 44,704 bytes maximum residency (2 sample(s)) 20,832 bytes maximum slop 59 MB total memory in use (0 MB lost due to fragmentation) %GC time 1.0% (1.0% elapsed) Productivity 97.4% of total user, 99.6% of total elapsed ./A +RTS -A50M -s 0.10s user 0.05s system 97% cpu 0.157 total --------------------------------- On Sat, Dec 1, 2012 at 11:09 AM, Branimir Maksimovic wrote:

...

I have made benchmark test inspired by http://lemire.me/blog/archives/2012/07/23/is-cc-worth-it/

What surprised me is that unboxed array is much faster than boxed array. Actually boxed array performance is on par with standard Haskell list which is very slow. All in all even unboxed array is about 10 times slower than Java version. I don't understand why is even unboxed array so slow. But! unboxed array consumes least amount of RAM. (warning, program consumes more than 3gb of ram)

bmaxa@maxa:~/examples$ time ./Cumul boxed array last 262486571 seconds 4.972 unboxed array last 262486571 seconds 0.776 list last 262486571 seconds 6.812

real 0m13.086s user 0m11.996s sys 0m1.080s

------------------------------------------------------------------------- {-# LANGUAGE CPP, BangPatterns #-} import System.CPUTime import Text.Printf import Data.Array.IO import Data.Array.Base import Data.Int import Control.DeepSeq import System.Mem

main :: IO() main = do (newArray_ (0,n'-1) :: IO(A)) >>= test "boxed array" performGC (newArray_ (0,n'-1) :: IO(B)) >>= test "unboxed array" performGC begin <- getCPUTime printf "list\nlast %d" $ last $ force $ take n' $ sum' data' end <- getCPUTime let diff = (fromIntegral (end - begin)) / (10^12) printf " seconds %.3f\n" (diff::Double)

test s a = do putStrLn s begin <- getCPUTime init' a partial_sum a end <- getCPUTime let diff = (fromIntegral (end - begin)) / (10^12) last <- readArray a (n'-1) printf "last %d seconds %.3f\n" last (diff::Double)

n' :: Int n' = 50 * 1000 * 1000

type A = IOArray Int Int32 type B = IOUArray Int Int32

init' a = do (_,n) <- getBounds a init a 0 n where init a k n | k > n = return () | otherwise = do let !x = fromIntegral $ k + k `div` 3 unsafeWrite a k x init a (k+1) n

partial_sum a = do (_,n) <- getBounds a k <- unsafeRead a 0 ps a 1 n k where ps a i n s | i > n = return () | otherwise = do k <- unsafeRead a i let !l = fromIntegral $ s + k unsafeWrite a i l ps a (i+1) n l

data' :: [Int32] data' = [k + k `div` 3 | k <- [0..] ]

sum' = scanl1 (+)

_______________________________________________ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe

Branimir Maksimovic

1:04 p.m.

Wow that sped it up 5 times.I see that boxed Vector is 25% faster than IOArray.What is the difference and when to use Vector,when IOArray?Thanks! bmaxa@maxa:~/examples$ time ./Cumul +RTS -A1600Mboxed arraylast 262486571 seconds 1.196unboxed arraylast 262486571 seconds 0.748boxed vectorlast 262486571 seconds 0.908unboxed vectorlast 262486571 seconds 0.720 real 0m3.805suser 0m3.428ssys 0m0.372s

...

Date: Sat, 1 Dec 2012 12:20:37 -0500 Subject: Re: [Haskell-cafe] Why is boxed mutable array so slow? From: dons00@gmail.com To: bmaxa@hotmail.com CC: haskell-cafe@haskell.org

The obvious difference between boxed and unboxed arrays is that the boxed arrays are full of pointers to heap allocated objects. This means you pay indirection to access the values, much more time in GC spent chasing pointers (though card marking helps), and generally do more allocation.

Compare the GC stats below, for

* Boxed vector: 88M bytes copied; 75% of time in GC, 0.472s * Unboxed vector: 11k bytes copied, 1.3% of time in GC, 0.077s

So there's your main answer. The increased data density of unboxed arrays also helps a too.

Now, you can help out the GC signifcantly by hinting at how much you're going to allocated in the youngest generation (see the ghc-gc-tune app for a methodical approach to this, though it needs updating to ghc 7 -- http://donsbot.wordpress.com/2010/07/05/ghc-gc-tune-tuning-haskell-gc-settin... and http://stackoverflow.com/questions/3171922/ghcs-rts-options-for-garbage-coll... ).

Use the +RTS -A flag to set an initial youngest generation heap size to the size of your array, and watch the GC cost disappear. For our boxed vector, we'd use +RTS -A50M, resulting in:

* Boxed vector: 8k copied, 1% of time in GC, 0.157s

So not bad. 3x speedup through a RTS flag. -A is very useful if you are working with boxed, mutable arrays.

For reference, there's a generic version below that specializes based on the vector type parameter.

---------------------------------

{-# LANGUAGE BangPatterns #-}

import System.CPUTime import Text.Printf import Data.Int import Control.DeepSeq import System.Mem

import qualified Data.Vector.Mutable as V import qualified Data.Vector.Unboxed.Mutable as U import qualified Data.Vector.Generic.Mutable as G

main :: IO() main = do

-- (G.new n' :: IO (V.IOVector Int32)) >>= test' "boxed vector" -- performGC (G.new n' :: IO (U.IOVector Int32)) >>= test' "unboxed vector" performGC

test' s a = do putStrLn s begin <- getCPUTime init'' a partial_sum' a end <- getCPUTime let diff = (fromIntegral (end - begin)) / (10**12) last <- G.read a (n'-1) printf "last %d seconds %.3f\n" last (diff::Double)

n' :: Int n' = 1000 * 1000

init'' !a = init 0 (n'-1) where init :: Int -> Int -> IO () init !k !n | k > n = return () | otherwise = do let !x = fromIntegral $ k + k `div` 3 G.write a k x init (k+1) n

partial_sum' !a = do k <- G.read a 0 ps 1 (n'-1) k where ps :: Int -> Int -> Int32 -> IO () ps i n s | i > n = return () | otherwise = do k <- G.read a i let !l = fromIntegral $ s + k G.write a i l ps (i+1) n l

---------------------------------

$ time ./A +RTS -s boxed vector last 945735787 seconds 0.420 40,121,448 bytes allocated in the heap 88,355,272 bytes copied during GC 24,036,456 bytes maximum residency (6 sample(s)) 380,632 bytes maximum slop 54 MB total memory in use (0 MB lost due to fragmentation)

%GC time 75.2% (75.9% elapsed)

Alloc rate 359,655,602 bytes per MUT second

./A +RTS -s 0.40s user 0.07s system 98% cpu 0.475 total

$ time ./A +RTS -s unboxed vector last 945735787 seconds 0.080 4,113,568 bytes allocated in the heap 11,288 bytes copied during GC 4,003,256 bytes maximum residency (3 sample(s)) 182,856 bytes maximum slop 5 MB total memory in use (0 MB lost due to fragmentation)

%GC time 1.3% (1.3% elapsed)

Alloc rate 51,416,660 bytes per MUT second

./A +RTS -s 0.08s user 0.01s system 98% cpu 0.088 total

$ time ./A +RTS -A50M -s boxed vector last 945735787 seconds 0.127 40,121,504 bytes allocated in the heap 8,032 bytes copied during GC 44,704 bytes maximum residency (2 sample(s)) 20,832 bytes maximum slop 59 MB total memory in use (0 MB lost due to fragmentation)

%GC time 1.0% (1.0% elapsed)

Productivity 97.4% of total user, 99.6% of total elapsed

./A +RTS -A50M -s 0.10s user 0.05s system 97% cpu 0.157 total

---------------------------------

On Sat, Dec 1, 2012 at 11:09 AM, Branimir Maksimovic wrote:

...
I have made benchmark test inspired by http://lemire.me/blog/archives/2012/07/23/is-cc-worth-it/

What surprised me is that unboxed array is much faster than boxed array. Actually boxed array performance is on par with standard Haskell list which is very slow. All in all even unboxed array is about 10 times slower than Java version. I don't understand why is even unboxed array so slow. But! unboxed array consumes least amount of RAM. (warning, program consumes more than 3gb of ram)

bmaxa@maxa:~/examples$ time ./Cumul boxed array last 262486571 seconds 4.972 unboxed array last 262486571 seconds 0.776 list last 262486571 seconds 6.812

real 0m13.086s user 0m11.996s sys 0m1.080s

------------------------------------------------------------------------- {-# LANGUAGE CPP, BangPatterns #-} import System.CPUTime import Text.Printf import Data.Array.IO import Data.Array.Base import Data.Int import Control.DeepSeq import System.Mem

main :: IO() main = do (newArray_ (0,n'-1) :: IO(A)) >>= test "boxed array" performGC (newArray_ (0,n'-1) :: IO(B)) >>= test "unboxed array" performGC begin <- getCPUTime printf "list\nlast %d" $ last $ force $ take n' $ sum' data' end <- getCPUTime let diff = (fromIntegral (end - begin)) / (10^12) printf " seconds %.3f\n" (diff::Double)

test s a = do putStrLn s begin <- getCPUTime init' a partial_sum a end <- getCPUTime let diff = (fromIntegral (end - begin)) / (10^12) last <- readArray a (n'-1) printf "last %d seconds %.3f\n" last (diff::Double)

n' :: Int n' = 50 * 1000 * 1000

type A = IOArray Int Int32 type B = IOUArray Int Int32

init' a = do (_,n) <- getBounds a init a 0 n where init a k n | k > n = return () | otherwise = do let !x = fromIntegral $ k + k `div` 3 unsafeWrite a k x init a (k+1) n

partial_sum a = do (_,n) <- getBounds a k <- unsafeRead a 0 ps a 1 n k where ps a i n s | i > n = return () | otherwise = do k <- unsafeRead a i let !l = fromIntegral $ s + k unsafeWrite a i l ps a (i+1) n l

data' :: [Int32] data' = [k + k `div` 3 | k <- [0..] ]

sum' = scanl1 (+)

_______________________________________________ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe

Don Stewart

1:22 p.m.

Regarding when to use mutable arrays versus vectors, I would always use vectors -- they optimize better, and have a better interface. Also, I have updated and released a new version of the tool mentioned below. You can get it on Hackage, updated to ghc 7 series. http://hackage.haskell.org/package/ghc-gc-tune-0.3 For your boxed vector program, we get results that show a clear performance peak with a -A of around 64M, about the size of the allocated array ... http://i.imgur.com/dZ2Eo.png Best settings for Running time: 0.16s: +RTS -A67108864 -H1048576 0.16s: +RTS -A67108864 -H2097152 0.16s: +RTS -A67108864 -H8388608 E.g. $ time ./A +RTS -A67M -H1M boxed vector last 945735787 seconds 0.123 -- Don On Sat, Dec 1, 2012 at 12:20 PM, Don Stewart wrote:

...

The obvious difference between boxed and unboxed arrays is that the boxed arrays are full of pointers to heap allocated objects. This means you pay indirection to access the values, much more time in GC spent chasing pointers (though card marking helps), and generally do more allocation.

Compare the GC stats below, for

* Boxed vector: 88M bytes copied; 75% of time in GC, 0.472s * Unboxed vector: 11k bytes copied, 1.3% of time in GC, 0.077s

So there's your main answer. The increased data density of unboxed arrays also helps a too.

Now, you can help out the GC signifcantly by hinting at how much you're going to allocated in the youngest generation (see the ghc-gc-tune app for a methodical approach to this, though it needs updating to ghc 7 -- http://donsbot.wordpress.com/2010/07/05/ghc-gc-tune-tuning-haskell-gc-settin... and http://stackoverflow.com/questions/3171922/ghcs-rts-options-for-garbage-coll... ).

Use the +RTS -A flag to set an initial youngest generation heap size to the size of your array, and watch the GC cost disappear. For our boxed vector, we'd use +RTS -A50M, resulting in:

* Boxed vector: 8k copied, 1% of time in GC, 0.157s

So not bad. 3x speedup through a RTS flag. -A is very useful if you are working with boxed, mutable arrays.

For reference, there's a generic version below that specializes based on the vector type parameter.

---------------------------------

{-# LANGUAGE BangPatterns #-}

import System.CPUTime import Text.Printf import Data.Int import Control.DeepSeq import System.Mem

import qualified Data.Vector.Mutable as V import qualified Data.Vector.Unboxed.Mutable as U import qualified Data.Vector.Generic.Mutable as G

main :: IO() main = do

-- (G.new n' :: IO (V.IOVector Int32)) >>= test' "boxed vector" -- performGC (G.new n' :: IO (U.IOVector Int32)) >>= test' "unboxed vector" performGC

test' s a = do putStrLn s begin <- getCPUTime init'' a partial_sum' a end <- getCPUTime let diff = (fromIntegral (end - begin)) / (10**12) last <- G.read a (n'-1) printf "last %d seconds %.3f\n" last (diff::Double)

n' :: Int n' = 1000 * 1000

init'' !a = init 0 (n'-1) where init :: Int -> Int -> IO () init !k !n | k > n = return () | otherwise = do let !x = fromIntegral $ k + k `div` 3 G.write a k x init (k+1) n

partial_sum' !a = do k <- G.read a 0 ps 1 (n'-1) k where ps :: Int -> Int -> Int32 -> IO () ps i n s | i > n = return () | otherwise = do k <- G.read a i let !l = fromIntegral $ s + k G.write a i l ps (i+1) n l

---------------------------------

$ time ./A +RTS -s boxed vector last 945735787 seconds 0.420 40,121,448 bytes allocated in the heap 88,355,272 bytes copied during GC 24,036,456 bytes maximum residency (6 sample(s)) 380,632 bytes maximum slop 54 MB total memory in use (0 MB lost due to fragmentation)

%GC time 75.2% (75.9% elapsed)

Alloc rate 359,655,602 bytes per MUT second

./A +RTS -s 0.40s user 0.07s system 98% cpu 0.475 total

$ time ./A +RTS -s unboxed vector last 945735787 seconds 0.080 4,113,568 bytes allocated in the heap 11,288 bytes copied during GC 4,003,256 bytes maximum residency (3 sample(s)) 182,856 bytes maximum slop 5 MB total memory in use (0 MB lost due to fragmentation)

%GC time 1.3% (1.3% elapsed)

Alloc rate 51,416,660 bytes per MUT second

./A +RTS -s 0.08s user 0.01s system 98% cpu 0.088 total

$ time ./A +RTS -A50M -s boxed vector last 945735787 seconds 0.127 40,121,504 bytes allocated in the heap 8,032 bytes copied during GC 44,704 bytes maximum residency (2 sample(s)) 20,832 bytes maximum slop 59 MB total memory in use (0 MB lost due to fragmentation)

%GC time 1.0% (1.0% elapsed)

Productivity 97.4% of total user, 99.6% of total elapsed

./A +RTS -A50M -s 0.10s user 0.05s system 97% cpu 0.157 total

---------------------------------

On Sat, Dec 1, 2012 at 11:09 AM, Branimir Maksimovic wrote:

...
I have made benchmark test inspired by http://lemire.me/blog/archives/2012/07/23/is-cc-worth-it/

What surprised me is that unboxed array is much faster than boxed array. Actually boxed array performance is on par with standard Haskell list which is very slow. All in all even unboxed array is about 10 times slower than Java version. I don't understand why is even unboxed array so slow. But! unboxed array consumes least amount of RAM. (warning, program consumes more than 3gb of ram)

bmaxa@maxa:~/examples$ time ./Cumul boxed array last 262486571 seconds 4.972 unboxed array last 262486571 seconds 0.776 list last 262486571 seconds 6.812

real 0m13.086s user 0m11.996s sys 0m1.080s

------------------------------------------------------------------------- {-# LANGUAGE CPP, BangPatterns #-} import System.CPUTime import Text.Printf import Data.Array.IO import Data.Array.Base import Data.Int import Control.DeepSeq import System.Mem

main :: IO() main = do (newArray_ (0,n'-1) :: IO(A)) >>= test "boxed array" performGC (newArray_ (0,n'-1) :: IO(B)) >>= test "unboxed array" performGC begin <- getCPUTime printf "list\nlast %d" $ last $ force $ take n' $ sum' data' end <- getCPUTime let diff = (fromIntegral (end - begin)) / (10^12) printf " seconds %.3f\n" (diff::Double)

test s a = do putStrLn s begin <- getCPUTime init' a partial_sum a end <- getCPUTime let diff = (fromIntegral (end - begin)) / (10^12) last <- readArray a (n'-1) printf "last %d seconds %.3f\n" last (diff::Double)

n' :: Int n' = 50 * 1000 * 1000

type A = IOArray Int Int32 type B = IOUArray Int Int32

init' a = do (_,n) <- getBounds a init a 0 n where init a k n | k > n = return () | otherwise = do let !x = fromIntegral $ k + k `div` 3 unsafeWrite a k x init a (k+1) n

partial_sum a = do (_,n) <- getBounds a k <- unsafeRead a 0 ps a 1 n k where ps a i n s | i > n = return () | otherwise = do k <- unsafeRead a i let !l = fromIntegral $ s + k unsafeWrite a i l ps a (i+1) n l

data' :: [Int32] data' = [k + k `div` 3 | k <- [0..] ]

sum' = scanl1 (+)

_______________________________________________ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe

Branimir Maksimovic

1:58 p.m.

Thanks! I have downloaded tool and playing with it.I will use boxed vectors in the future ;)

...

Date: Sat, 1 Dec 2012 13:22:47 -0500 Subject: Re: [Haskell-cafe] Why is boxed mutable array so slow? From: dons00@gmail.com To: bmaxa@hotmail.com CC: haskell-cafe@haskell.org

Regarding when to use mutable arrays versus vectors, I would always use vectors -- they optimize better, and have a better interface.

Also, I have updated and released a new version of the tool mentioned below. You can get it on Hackage, updated to ghc 7 series.

http://hackage.haskell.org/package/ghc-gc-tune-0.3

For your boxed vector program, we get results that show a clear performance peak with a -A of around 64M, about the size of the allocated array ...

http://i.imgur.com/dZ2Eo.png

Best settings for Running time: 0.16s: +RTS -A67108864 -H1048576 0.16s: +RTS -A67108864 -H2097152 0.16s: +RTS -A67108864 -H8388608

E.g.

$ time ./A +RTS -A67M -H1M boxed vector last 945735787 seconds 0.123

-- Don

On Sat, Dec 1, 2012 at 12:20 PM, Don Stewart wrote:

...
The obvious difference between boxed and unboxed arrays is that the boxed arrays are full of pointers to heap allocated objects. This means you pay indirection to access the values, much more time in GC spent chasing pointers (though card marking helps), and generally do more allocation.

Compare the GC stats below, for

* Boxed vector: 88M bytes copied; 75% of time in GC, 0.472s * Unboxed vector: 11k bytes copied, 1.3% of time in GC, 0.077s

So there's your main answer. The increased data density of unboxed arrays also helps a too.

Now, you can help out the GC signifcantly by hinting at how much you're going to allocated in the youngest generation (see the ghc-gc-tune app for a methodical approach to this, though it needs updating to ghc 7 -- http://donsbot.wordpress.com/2010/07/05/ghc-gc-tune-tuning-haskell-gc-settin... and http://stackoverflow.com/questions/3171922/ghcs-rts-options-for-garbage-coll... ).

Use the +RTS -A flag to set an initial youngest generation heap size to the size of your array, and watch the GC cost disappear. For our boxed vector, we'd use +RTS -A50M, resulting in:

* Boxed vector: 8k copied, 1% of time in GC, 0.157s

So not bad. 3x speedup through a RTS flag. -A is very useful if you are working with boxed, mutable arrays.

For reference, there's a generic version below that specializes based on the vector type parameter.

---------------------------------

{-# LANGUAGE BangPatterns #-}

import System.CPUTime import Text.Printf import Data.Int import Control.DeepSeq import System.Mem

import qualified Data.Vector.Mutable as V import qualified Data.Vector.Unboxed.Mutable as U import qualified Data.Vector.Generic.Mutable as G

main :: IO() main = do

-- (G.new n' :: IO (V.IOVector Int32)) >>= test' "boxed vector" -- performGC (G.new n' :: IO (U.IOVector Int32)) >>= test' "unboxed vector" performGC

test' s a = do putStrLn s begin <- getCPUTime init'' a partial_sum' a end <- getCPUTime let diff = (fromIntegral (end - begin)) / (10**12) last <- G.read a (n'-1) printf "last %d seconds %.3f\n" last (diff::Double)

n' :: Int n' = 1000 * 1000

init'' !a = init 0 (n'-1) where init :: Int -> Int -> IO () init !k !n | k > n = return () | otherwise = do let !x = fromIntegral $ k + k `div` 3 G.write a k x init (k+1) n

partial_sum' !a = do k <- G.read a 0 ps 1 (n'-1) k where ps :: Int -> Int -> Int32 -> IO () ps i n s | i > n = return () | otherwise = do k <- G.read a i let !l = fromIntegral $ s + k G.write a i l ps (i+1) n l

---------------------------------

$ time ./A +RTS -s boxed vector last 945735787 seconds 0.420 40,121,448 bytes allocated in the heap 88,355,272 bytes copied during GC 24,036,456 bytes maximum residency (6 sample(s)) 380,632 bytes maximum slop 54 MB total memory in use (0 MB lost due to fragmentation)

%GC time 75.2% (75.9% elapsed)

Alloc rate 359,655,602 bytes per MUT second

./A +RTS -s 0.40s user 0.07s system 98% cpu 0.475 total

$ time ./A +RTS -s unboxed vector last 945735787 seconds 0.080 4,113,568 bytes allocated in the heap 11,288 bytes copied during GC 4,003,256 bytes maximum residency (3 sample(s)) 182,856 bytes maximum slop 5 MB total memory in use (0 MB lost due to fragmentation)

%GC time 1.3% (1.3% elapsed)

Alloc rate 51,416,660 bytes per MUT second

./A +RTS -s 0.08s user 0.01s system 98% cpu 0.088 total

$ time ./A +RTS -A50M -s boxed vector last 945735787 seconds 0.127 40,121,504 bytes allocated in the heap 8,032 bytes copied during GC 44,704 bytes maximum residency (2 sample(s)) 20,832 bytes maximum slop 59 MB total memory in use (0 MB lost due to fragmentation)

%GC time 1.0% (1.0% elapsed)

Productivity 97.4% of total user, 99.6% of total elapsed

./A +RTS -A50M -s 0.10s user 0.05s system 97% cpu 0.157 total

---------------------------------

On Sat, Dec 1, 2012 at 11:09 AM, Branimir Maksimovic wrote:

...
I have made benchmark test inspired by http://lemire.me/blog/archives/2012/07/23/is-cc-worth-it/

What surprised me is that unboxed array is much faster than boxed array. Actually boxed array performance is on par with standard Haskell list which is very slow. All in all even unboxed array is about 10 times slower than Java version. I don't understand why is even unboxed array so slow. But! unboxed array consumes least amount of RAM. (warning, program consumes more than 3gb of ram)

bmaxa@maxa:~/examples$ time ./Cumul boxed array last 262486571 seconds 4.972 unboxed array last 262486571 seconds 0.776 list last 262486571 seconds 6.812

real 0m13.086s user 0m11.996s sys 0m1.080s

------------------------------------------------------------------------- {-# LANGUAGE CPP, BangPatterns #-} import System.CPUTime import Text.Printf import Data.Array.IO import Data.Array.Base import Data.Int import Control.DeepSeq import System.Mem

main :: IO() main = do (newArray_ (0,n'-1) :: IO(A)) >>= test "boxed array" performGC (newArray_ (0,n'-1) :: IO(B)) >>= test "unboxed array" performGC begin <- getCPUTime printf "list\nlast %d" $ last $ force $ take n' $ sum' data' end <- getCPUTime let diff = (fromIntegral (end - begin)) / (10^12) printf " seconds %.3f\n" (diff::Double)

test s a = do putStrLn s begin <- getCPUTime init' a partial_sum a end <- getCPUTime let diff = (fromIntegral (end - begin)) / (10^12) last <- readArray a (n'-1) printf "last %d seconds %.3f\n" last (diff::Double)

n' :: Int n' = 50 * 1000 * 1000

type A = IOArray Int Int32 type B = IOUArray Int Int32

init' a = do (_,n) <- getBounds a init a 0 n where init a k n | k > n = return () | otherwise = do let !x = fromIntegral $ k + k `div` 3 unsafeWrite a k x init a (k+1) n

partial_sum a = do (_,n) <- getBounds a k <- unsafeRead a 0 ps a 1 n k where ps a i n s | i > n = return () | otherwise = do k <- unsafeRead a i let !l = fromIntegral $ s + k unsafeWrite a i l ps a (i+1) n l

data' :: [Int32] data' = [k + k `div` 3 | k <- [0..] ]

sum' = scanl1 (+)

_______________________________________________ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe

Daniel Fischer

3:12 p.m.

On Samstag, 1. Dezember 2012, 16:09:05, Branimir Maksimovic wrote:

...

All in all even unboxed array is about 10 times slower than Java version. I don't understand why is even unboxed array so slow.

It's not the unboxed arrays that are slow. Your code has a couple of weak spots, and GHC's native code generator has a weakness that bites here. On my box, I don't quite have a 10× difference to my translation to Java, it's a bit less than 7× (0.82s vs 0.12s - I don't want to bring my box to its knees by running something that takes 3GB+ of RAM, so I run the unboxed array part only) with the LLVM backend and 8× (0.93s) with the native code generator. That's in the same ballpark, though. So what's the deal? Main.main_$s$wa1 [Occ=LoopBreaker] :: GHC.Prim.Int# -> GHC.Prim.Int# -> GHC.Prim.State# GHC.Prim.RealWorld -> GHC.Types.Int -> GHC.Types.Int -> GHC.Types.Int -> ... Your loops carry boxed Ints around, that's always a bad sign. In this case it doesn't hurt too much, however, since these values are neither read nor substituted during the loop (they're first and last index of the array and number of elements). Additionally, they carry an IOUArray constructor around. That is unnecessary. Eliminating a couple of dead parameters init' a = do (_,n) <- getBounds a let init k | k > n = return () | otherwise = do let x = fromIntegral $ k + k `div` 3 unsafeWrite a k x init (k+1) init 0 partial_sum a = do (_,n) <- getBounds a let ps i s | i > n = return () | otherwise = do k <- unsafeRead a i let l = s + k unsafeWrite a i l ps (i+1) l k <- unsafeRead a 0 ps 1 k brings the time for the native code generator down to 0.82s, and for the LLVM backend the time remains the same. Next problem, you're using `div` for the division. `div` does some checking and potentially fixup (not here, since everything is non-negative) after the machine division because `div` is specified to satisfy a = (a `div` b) * b + (a `mod` b) with 0 <= a `mod` b < abs b. That is in itself slower than the pure machine division you get with quot. So let's see what we get with `quot`. 0.65s with the native code generator, and 0.13 with the LLVM backend. Whoops, what's that? The problem is, as can be seen by manually replacing k `quot` 3 with (k *2863311531) `shiftR` 33 (requires 64-bit Ints; equivalent in Java: k*28..1L >> 33), when the native backend, the LLVM backend and Java (as well as C) all take more or less the same time [well, the NCG is a bit slower than the other two, 0.11s, 0.11s, 0.14s], that division is a **very** slow operation. Java and LLVM know how to replace the division by the constant 3 with a mulitplication, a couple of shifts and an addition (since we never have negative numbers here, just one multiplication and shift suffice, but neither Java nor LLVM can do that on their own because it's not guaranteed by the type). The native code generator doesn't - not yet. So the programme spends the majority of the time dividing. The array reads and writes are on par with Java's (and, for that matter, C's). If you make the divisor a parameter instead of a compile time constant, the NCG is not affected at all, the LLVM backend gives you equal performance (it can't optimise a division by a divisor it doesn't know). Java is at an advantage there, after a while the JIT sees that it might be a good idea to optimise the division and so its time only trebles.

Branimir Maksimovic

2 Dec 2 Dec

12:57 a.m.

Wow, now performance is on par with Java ;)So slow division was main problem, that and GC . Thanks!

...

From: daniel.is.fischer@googlemail.com To: haskell-cafe@haskell.org CC: bmaxa@hotmail.com Subject: Re: [Haskell-cafe] Why is boxed mutable array so slow? Date: Sat, 1 Dec 2012 21:12:29 +0100

On Samstag, 1. Dezember 2012, 16:09:05, Branimir Maksimovic wrote:

...
All in all even unboxed array is about 10 times slower than Java version. I don't understand why is even unboxed array so slow.

It's not the unboxed arrays that are slow.

Your code has a couple of weak spots, and GHC's native code generator has a weakness that bites here.

On my box, I don't quite have a 10× difference to my translation to Java, it's a bit less than 7× (0.82s vs 0.12s - I don't want to bring my box to its knees by running something that takes 3GB+ of RAM, so I run the unboxed array part only) with the LLVM backend and 8× (0.93s) with the native code generator. That's in the same ballpark, though.

So what's the deal?

Main.main_$s$wa1 [Occ=LoopBreaker] :: GHC.Prim.Int# -> GHC.Prim.Int# -> GHC.Prim.State# GHC.Prim.RealWorld -> GHC.Types.Int -> GHC.Types.Int -> GHC.Types.Int -> ...

Your loops carry boxed Ints around, that's always a bad sign. In this case it doesn't hurt too much, however, since these values are neither read nor substituted during the loop (they're first and last index of the array and number of elements). Additionally, they carry an IOUArray constructor around. That is unnecessary. Eliminating a couple of dead parameters

init' a = do (_,n) <- getBounds a let init k | k > n = return () | otherwise = do let x = fromIntegral $ k + k `div` 3 unsafeWrite a k x init (k+1) init 0

partial_sum a = do (_,n) <- getBounds a let ps i s | i > n = return () | otherwise = do k <- unsafeRead a i let l = s + k unsafeWrite a i l ps (i+1) l k <- unsafeRead a 0 ps 1 k

brings the time for the native code generator down to 0.82s, and for the LLVM backend the time remains the same.

Next problem, you're using `div` for the division.

`div` does some checking and potentially fixup (not here, since everything is non-negative) after the machine division because `div` is specified to satisfy

a = (a `div` b) * b + (a `mod` b)

with 0 <= a `mod` b < abs b.

That is in itself slower than the pure machine division you get with quot.

So let's see what we get with `quot`.

0.65s with the native code generator, and 0.13 with the LLVM backend.

Whoops, what's that?

The problem is, as can be seen by manually replacing k `quot` 3 with

(k *2863311531) `shiftR` 33

(requires 64-bit Ints; equivalent in Java: k*28..1L >> 33), when the native backend, the LLVM backend and Java (as well as C) all take more or less the same time [well, the NCG is a bit slower than the other two, 0.11s, 0.11s, 0.14s], that division is a **very** slow operation.

Java and LLVM know how to replace the division by the constant 3 with a mulitplication, a couple of shifts and an addition (since we never have negative numbers here, just one multiplication and shift suffice, but neither Java nor LLVM can do that on their own because it's not guaranteed by the type). The native code generator doesn't - not yet.

So the programme spends the majority of the time dividing. The array reads and writes are on par with Java's (and, for that matter, C's).

If you make the divisor a parameter instead of a compile time constant, the NCG is not affected at all, the LLVM backend gives you equal performance (it can't optimise a division by a divisor it doesn't know). Java is at an advantage there, after a while the JIT sees that it might be a good idea to optimise the division and so its time only trebles.

4552

Age (days ago)

4553

Last active (days ago)

List overview

Download

7 comments

4 participants

participants (4)

Branimir Maksimovic
Daniel Fischer
Don Stewart
KC