[racket] typed racket slow?

They're using a shared image of the Racket run-time system, but they
have separate copies of the Racket-implemented libraries that are in
bytecode form and embedded in each executable.

The reason that `raco exe' produced slower results than `raco make' is
that the generated executable did not support on-demand parsing of its
embedded bytecode, unlike the handling of normal bytecode files.

I've changed `racket' so that an executable created by `raco exe' can
parse its bytecode on demand, and then running via `raco make' or `raco
exe' is about the same.

At Wed, 8 May 2013 08:21:23 -0400, Sean McBeth wrote:
> Perhaps the EXEs take longer because they are not using a shared image of
> Racket, thus the OS must load code that is 99% equivalent from two
> different disk locations.
> 
> 
> On Wed, May 8, 2013 at 7:55 AM, Manfred Lotz <manfred.lotz at arcor.de> wrote:
> 
> > On Wed, 8 May 2013 07:31:37 -0400
> > Carl Eastlund <cce at ccs.neu.edu> wrote:
> >
> > > On Wed, May 8, 2013 at 7:04 AM, Manfred Lotz
> > > <manfred.lotz at arcor.de> wrote:
> > >
> > > > On Wed, 8 May 2013 06:19:27 -0400
> > > > Carl Eastlund <cce at ccs.neu.edu> wrote:
> > > >
> > > > > I'm seeing similar results on my end; I timed by first running
> > > > > "raco make" on both files, then timing "racket" on both.  I think
> > > > > what we're seeing is a small startup time cost on Typed Racket.
> > > > > I ran a longer benchmark and Typed Racket edges out untyped
> > > > > Racket if I run a few million iterations (given this is such a
> > > > > short computation).  The expressions I used are:
> > > > >
> > > > > ;; utest.rkt
> > > > > (void
> > > > >   (for/list {[i (in-range (* 10 1000 1000))]}
> > > > >     (distance
> > > > >       (pt (+ i 1.2) (+ i 2.1))
> > > > >       (pt (+ i 4.3) (+ i 5.6)))))
> > > > >
> > > > > and
> > > > >
> > > > > ;; test.rkt
> > > > > (void
> > > > >   (for/list: : (Listof Float) {[i (in-range (* 10 1000 1000))]}
> > > > >     (distance
> > > > >       (pt (+ i 1.2) (+ i 2.1))
> > > > >       (pt (+ i 4.3) (+ i 5.6)))))
> > > > >
> > > > >
> > > > > I see just under 5 seconds for test.rkt and just over 5 seconds
> > > > > for utest.rkt.  So there's a fraction of a second extra startup
> > > > > time for Typed Racket, but it takes less time for each subsequent
> > > > > computation, so the difference depends on how much "real" work
> > > > > you do after startup.  I don't know what causes that startup
> > > > > cost, but hopefully this kind of benchmark will be useful to the
> > > > > Typed Racket maintainers in closing the gap for future versions.
> > > > > So, thanks for the example, Manfred!
> > > > >
> > > >
> > > > Hi Carl,
> > > > This is interesting. If I run it I have around 5 seconds for the
> > > > typed version and around 4 seconds for the untyped version. My
> > > > system is a 64bit Linux.
> > > >
> > > > --
> > > > Manfred
> > > >
> > >
> > > What I ran was:
> > >
> > >   raco make test.rkt utest.rkt && time racket test.rkt && time racket
> > > utest.rkt
> > >
> > > Just to make sure we're comparing apples to apples, does that give
> > > you the same results you saw before?
> > >
> > >
> >
> > Yep, I did something different:
> >
> > Now going your way I get 3.7 sec for typed and 4.1 sec for untyped. It
> > is interesting to note that if I do raco exe for both the executables
> > run longer: 4 sec for typed and 4.7 sec for untyped.
> >
> >
> >
> > > > > On Wed, May 8, 2013 at 5:32 AM, Manfred Lotz
> > > >  > <manfred.lotz at arcor.de> wrote:
> > > > >
> > > > > > Hi there,
> > > > > > I did a small test using typed racket.
> > > > > >
> > > > > > This is an example from the documentation:
> > > > > >
> > > > > > #lang typed/racket
> > > > > > ;; test.rkt
> > > > > >
> > > > > > (struct: pt ([x : Float] [y : Float]))
> > > > > >
> > > > > > (: distance (pt pt -> Float))
> > > > > > (define (distance p1 p2)
> > > > > >   (sqrt (+ (sqr (- (pt-x p2) (pt-x p1)))
> > > > > >            (sqr (- (pt-y p2) (pt-y p1))))))
> > > > > >
> > > > > > (distance (pt 1.2 2.1) (pt 4.3 5.6))
> > > > > >
> > > > > > This is the untyped version:
> > > > > > #lang racket
> > > > > > ;; utest.rkt
> > > > > >
> > > > > > (struct pt (x y))
> > > > > >
> > > > > > (define (distance p1 p2)
> > > > > >   (sqrt (+ (sqr (- (pt-x p2) (pt-x p1)))
> > > > > >            (sqr (- (pt-y p2) (pt-y p1))))))
> > > > > >
> > > > > > (distance (pt 1.2 2.1) (pt 4.3 5.6))
> > > > > >
> > > > > >
> > > > > > Now running both:
> > > > > > time racket test.rkt
> > > > > > 4.675467891024383
> > > > > > racket test.rkt  1.24s user 0.08s system 99% cpu 1.333 total
> > > > > >
> > > > > >
> > > > > > time racket utest.rkt
> > > > > > 4.675467891024383
> > > > > > racket utest.rkt  0.22s user 0.03s system 99% cpu 0.248 total
> > > > > >
> > > > > >
> > > > > > It seems the typed version needs a lot of time for the type
> > > > > > checking. The time for time checking could be cut mostly by:
> > > > > >
> > > > > > raco exe test.rkt
> > > > > > time ./test
> > > > > > 4.675467891024383
> > > > > > ./test  0.49s user 0.03s system 99% cpu 0.531 total
> > > > > >
> > > > > > But still runtime is more than twice as long. I could get the
> > > > > > impression that typed racket is generally slower.
> > > > > >
> > > > > >
> > > > > > Question: Is there any conclusion to be drawn from this (like
> > > > > > that typed racket is slower than 'normal' racket)? Or is my
> > > > > > test just a bad test?
> > > > > >
> > > > > >
> > > > > > --
> > > > > > Manfred
> > > >
> > >
> >
> >
> > ____________________
> >   Racket Users list:
> >   http://lists.racket-lang.org/users
> >
> ____________________
>   Racket Users list:
>   http://lists.racket-lang.org/users