[Haskell-cafe] Measuring memory usage
Vlatko Basic
vlatko.basic at gmail.com
Wed Jul 4 18:06:32 UTC 2018
Hi Li-yao,
Thanks for taking the time to test and explain in so much details. It is much
clearer to me now. :-)
I changed 'b' to "b = YUid $ Uid (i * 2)" so they are different now.
Thanks to your correction, I added 6W to sizeOf X and calculated that its mem
usage is now 216, but I do not understand which text field are you talking
about. I added 8 (i.e. 2N) to both Text fields (c and e). I checked xs30 and
both text fields are 4 chars long.
data X = X -- = 1W + 5W
{ a :: XUid -- = 2W -- Int
, b :: YUid -- = 2W -- Int
, c :: Id -- = 6W + 8B -- Text len 4
, d :: Either Dir Dir -- = W + 2W
, e :: Text -- = 6W + 8B -- Text len 4
} deriving (Generic, D.Data) -- = 25W + 16B = 200 + 16 = 216
so xs1 would be 216 + 4W = 248
The diff is now minimal, except for xs1.
What ds Expected Diff
xs1: 272 248 0.88% -- overhead 24B - 3W
xs30: 7,232 7,208 0.04% -- overhead 24B - 3W
xs600: 144,032 144,008 0.00% -- overhead 24B - 3W
Seems like a constant induced by measurement. For recursiveSizeNF it is also
constant, 64B - 8W.
I'm very satisfied with the proof that recursiveSize(NF) shows the correct size,
less 24 (or 64).
I'm on GHC 8.2.2 and saw no diff with different levels of optimization, or
without any.
Have a nice day,
vlatko
-------- Original Message --------
Subject: Re: [Haskell-cafe] Measuring memory usage
From: Li-yao Xia <lysxia at gmail.com>
To: vlatko.basic at gmail.com
Cc: haskell-cafe <Haskell-cafe at Haskell.org>
Date: 04/07/18 15:12
> Hi Vlatko,
>
> I don't get the same numbers with ghc-datasize
>
> recursiveSize $! xs1 -- 240 (list of length 1)
> recursiveSize $! xs30 -- 6296
>
> compiled on GHC 8.0.2 with optimizations (-O)
>
> whereas recursiveSizeNF gives me the size of the thunk (force xs1), which
> depends on whether xs1 is evaluated or not.
>
> The size of X is missing:
>
> - 1 word for the X constructor
> - 5 words for the fields (they are not unpacked, this must be explicitly
> required with the {-# UNPACK #-} pragma)
> - 8B for one of the text fields (which for some reason takes 64B instead of 56B)
>
> That's 56B extra, for a total of 224B. Add in 4W for a singleton list and that's
> 256B. Why are 16B missing in the above 240B figure? Note that in mkX, there are
> two fields equal to (Uid i), so they get shared (recursiveSize doesn't recount
> shared structures). (It's hardly an optimization: if we take out the newtypes,
> the code looks like X {a = i, b = i, ...}, so the compiler simply puts the same
> pointer in the two fields.)
>
> Notice also that nullary constructors (in particular Yes/No) will always be
> shared. The compiler may also float out the whole "Left Yes" and "Right No" to
> the toplevel, further reducing the size of longer lists.
>
> gsize counts constructors. In particular, each newtype constructor counts as
> one, and primitives like Int and Char also count as one (the value they box is
> not visible to Data), and Text has a dummy Data instance to make it seem like a
> newtype around [Char].
>
> You can use this snippet to see a trace of the generic traversal:
>
> everywhereM (\x -> print (dataTypeOf x) >> return x) xs1
>
> Finally, "weigh" counts allocations, which are generally a superset of the
> actual space taken by a data structure once it is fully computed.
>
> Li-yao
>
>
> On 07/04/2018 07:12 AM, Vlatko Basic wrote:
>> Hello again,
>>
>> I'm still trying to find some method to predict memory usage and came upon
>> this page: https://wiki.haskell.org/GHC/Memory_Footprint
>>
>> I tried to manually calculate how much memory will the record below consume
>> (64-bit, 1W = 8B):
>>
>> newtype Id = Id Text deriving (Generic, Data, NFData) -- = 6W
>> newtype Uid = Uid Int deriving (Generic, Data, NFData) -- = 2W
>> newtype XUid = XUid Uid deriving (Generic, Data, NFData) -- = 2W
>> newtype YUid = YUid Uid deriving (Generic, Data, NFData) -- = 2W
>> data Dir = Yes | No deriving (Generic, Data) -- = 2W
>>
>> data X = X
>> { a :: XUid -- = 2W -- Int
>> , b :: YUid -- = 2W -- Int
>> , c :: Id -- = 6W + 8B -- Text len 4
>> , d :: Either Dir Dir -- = 1W + 2W -- Either + Dir + No/Yes
>> , e :: Text -- = 6W + 8B -- Text len 4
>> } deriving (Generic, Data) -- = 19W + 16B = 152 + 16 = 168B
>>
>> and calculated the assumed sizes of few lists with different number of elements:
>>
>> Expected list sizes ([v] = (1 + 3N) words + N * sizeof(v))
>> 30: 1 + 3W * 30 + (19W * 30 + 16 * 30) = 5,761 B
>> 600: 1 + 3W * 600 + (19W * 600 + 16 * 600) = 115,201 B
>> 5000: 1 + 3W * 5000 + (19W * 5000 + 16 * 5000) = 960,001 B
>>
>> I also compared these sizes with three libs (Data.Generics.Schemes.gsize,
>> GHC.DataSize.recursiveSizeNF, Weigh) and the results were:
>>
>> #items recursiveSizeNF gSize Weigh Expected Diff/recursiveSizeNF
>> 1: 1,416 18 696 168 -
>> 30: 8,008 931 20,880 5,761 28%
>> 600: 135,688 18,601 417,600 115,201 15%
>> 5000: 1,121,288 155,001 3,480,000 960,001 14%
>>
>> As you can see, the results are more than surprising (to me), with
>> recursiveSizeNF coming closest. They all measure the same variable.
>>
>>
>> What am I missing?
>>
>>
>> For completeness, here are relevant parts of code for creating elements (with
>> excessive forcing):
>>
>> let mkX i = force X{ a = XUid $ Uid i
>> , b = YUid $ Uid i
>> , c = Id $ tshow i
>> , d = if even i then (Left Yes) else (Right No)
>> , e = T.reverse (tshow i)
>> }
>> xs30 = force . map mkX $ take 30 $ randomRs (1000,1030) (mkStdGen 0)
>> xs600 = force . map mkX $ take 600 $ randomRs (1000,1600) (mkStdGen 0)
>> xs5K = force . map mkX $ take 5000 $ randomRs (1000,5000) (mkStdGen 0)
>>
>> dataSize <- recursiveSizeNF $!! {a}
>> let gSize = gsize $!! mkX 0
>>
>>
>>
>>> -------- Original Message --------
>>> Subject: Measuring memory usage
>>> From: Vlatko Basic <vlatko.basic at gmail.com>
>>> To: haskell-cafe <Haskell-cafe at Haskell.org>
>>> Date: 29/06/18 14:31
>>>
>>>
>>> Hello,
>>>
>>> I've come to some strange results using Weigh package.
>>>
>>> It shows that HashMap inside 'data' is using much, much more memory.
>>>
>>> The strange thing is that I'm seeing too large mem usage in my app as well
>>> (several "MapData" like in records), and trying to figure out with 'weigh'
>>> what's keeping the mem.
>>>
>>> Noticed that when I change the code to use HashMap directly (not inside
>>> 'data', that's the only change), the mem usage observed with top drops down
>>> for ~60M, from 850M to 790M.
>>>
>>>
>>> These are the test results for 10K, 5K and 3.3K items for "data MapData k v =
>>> MapData (HashMap k v)" (at the end is the full runnable example.)
>>>
>>> Case Allocated GCs
>>> HashMap 262,824 0
>>> HashMap half 58,536 0
>>> HashMap third 17,064 0
>>> MapData 4,242,208 4
>>>
>>> I tested by changing the order, disabling all but one etc., and the results
>>> were the same. Same 'weigh' behaviour with IntMap and Map.
>>>
>>>
>>> So, if anyone knows and has some experience with such issues, my questions are:
>>>
>>> 1. Is 'weigh' package reliable/usable, at least to some extent? (the results
>>> do show diff between full, half and third)
>>>
>>> 2. How do you measure mem consumptions of your large data/records?
>>>
>>> 3. If the results are even approximately valid, what could cause such large
>>> discrepancies with 'data'?
>>>
>>> 4. Is there a way to see if some record has been freed from memory, GCed?
>>>
>>>
>>>
>>> module Main where
>>>
>>> import Prelude
>>>
>>> import Control.DeepSeq (NFData)
>>> import Data.HashMap.Strict (HashMap, fromList)
>>> import GHC.Generics (Generic)
>>> import Weigh (mainWith, value)
>>>
>>>
>>> data MapData k v = MapData (HashMap k v) deriving Generic
>>> instance (NFData k, NFData v) => NFData (MapData k v)
>>>
>>> full, half, third :: Int
>>> full = 10000
>>> half = 5000
>>> third = 3333
>>>
>>> main :: IO ()
>>> main = mainWith $ do
>>> value "HashMap" ( mkHMList full)
>>> value "HashMap half" ( mkHMList half)
>>> value "HashMap third" ( mkHMList third)
>>> value "MapData" (MapData $ mkHMList full)
>>>
>>> mkHMList :: Int -> HashMap Int String
>>> mkHMList n = fromList . zip [1..n] $ replicate n "some text"
>>>
>>>
>>>
>>
>>
>>
>> _______________________________________________
>> Haskell-Cafe mailing list
>> To (un)subscribe, modify options or view archives go to:
>> http://mail.haskell.org/cgi-bin/mailman/listinfo/haskell-cafe
>> Only members subscribed via the mailman list are allowed to post.
>>
More information about the Haskell-Cafe
mailing list