[Haskell-cafe] Understanding GHC allocations

[Daniel Fischer]

On Thursday 17 June 2010 11:43:09, Roman Cheplyaka wrote:
> * Roman Cheplyaka <roma at ro-che.info> [2010-06-17 12:40:59+0300]
>
> > I'm trying to optimize the following program:
> > http://github.com/feuerbach/particles/blob/303c8a17c9b732e22457b5409bd
> >ce4b7520be94a/run.hs
> >
> > Of course general suggestions are welcome (BTW I'm going to give a try
> > to vector), but currently I'm concerned with two questions:
> >
> > 1. Heavy allocations in 'distance' function. Here is (part of) the
> > profile:
> >
> > COST CENTRE   MODULE    %time %alloc  ticks     bytes
> >
> > d2            Main        9.0   22.0    290 600000000
> > d             Main        8.6   65.9    278 1800000000
> > d1            Main        7.5   11.0    242 299700000
> >

I suspect the distance function is not what you intended,


distance :: Double -> Double -> Double
distance !x1 !x2 = {-# SCC "min" #-} min d1 d2
    where
    d = {-# SCC "d" #-} x1 - x2
    d1 = {-# SCC "d1" #-} abs d
    d2 = {-# SCC "d2" #-} abs $ l - d

that would give

distance 0.2 24.8 = 24.6, while the wrapping suggests that it should be 
0.4, so in d2, it should be d1 instead of d.
Either way, both d and d1 are <= 25, so the 'abs' in d2 is superfluous, 
removing that alone reduces the allocations drastically and the running 
time by ~40% (astonishingly, not in the profiling version, I suspect it's 
because profiling needs a few registers so that there aren't enough left 
for the loops on my box).

Further, if you export only main from the module, you allow GHC to be more 
aggressive with optimising. On my box, that leads to more allocation again 
because there aren't enough registers, but things become a little faster.

Also, a few more bangs here and there plus a couple of INLINE and UNPACK 
pragmas speed things up, the (on my box) fastest combination I've found is 
attached, it has the same semantics for distance as the original code, 
changing distance to what I believe it should be unfortunately slows it 
down significantly.

On my box, I get a further big speedup by compiling with

-O2 -fexcess-precision -fvia-C -optc-O3

> > >From reading core I got the impression that everything is strict &
> >
> > unboxed.

Not everything, there lurk a few boxed Doubles e.g. in average.

> > Perhaps this is related to creating some closures? How to get
> > rid of those allocations?
> >

Do you need to? Sometimes an allocating loop is faster than a non-
allocating one (of course, if you have enough registers for the allocating 
loop to run entirely in registers, it'll be much faster still).

IMO, the important criteria are time and resident memory, not allocation.

> > 2. Again from reading the core I learned that although 'l' and other
> > constants are inlined, their type is boxed Double. This makes sense
> > since CAFs are evaluated on demand, but obviously in this particular
> > case it does not make sense, so can I somehow make them unboxed?

Putting bangs in the loops where they are used likely uses the unboxed 
values; not exporting them too.

>
> Forgot to mention, I'm using ghc 6.12.1, compiling with -O2.

-------------- next part --------------
A non-text attachment was scrubbed...
Name: runNew.hs
Type: text/x-haskell
Size: 2750 bytes
Desc: not available
Url : http://www.haskell.org/pipermail/haskell-cafe/attachments/20100617/d1cef8f0/runNew.bin