[racket] Profiling help
Nothing comes to mind.
If you know where the runtime is supposed to start you might try adding
printfs in a few places that print out current-process-milliseconds (put a
printf right before the call to profile-thunk too) and see if you can kind
of hone in on something. You might find something surprising that would
give a clue.
Robby
On Thu, Jul 11, 2013 at 4:33 PM, Joe Gibbs Politz <joe at cs.brown.edu> wrote:
> On Thu, Jul 11, 2013 at 5:22 PM, Robby Findler
> <robby at eecs.northwestern.edu> wrote:
> > Is it possible that the time is being spent in the expander or in other
> > C-based code that would be hidden from the profiler?
>
> I don't know quite what C-based code we might be touching, but I do
> know that our parser down to a racket/base program is no more than a
> couple hundred milliseconds for the programs we're working with. I
> also tried emitting a more basic Racket module (via pretty-write and
> syntax->datum) from a parsed program and profiled that using
> `dynamic-require` rather than `eval` and got similar results.
>
> There are still macros in the emitted program, but there is no
> (intended) runtime expansion after the initial macros are dealt with.
>
> Are there other common C-based code entry points we might be running afoul
> of?
>
>
> >
> > Robby
> >
> >
> > On Thu, Jul 11, 2013 at 4:07 PM, Joe Gibbs Politz <joe at cs.brown.edu>
> wrote:
> >>
> >> We're trying to profile some slow running programs in a #lang, and
> >> having some trouble getting believable results from the profiler. In
> >> particular, it claims that most (e.g. 60-80%) of the time is spent in
> >>
> >> /usr/lib/racket/collects/profile/main.rkt:29:2
> >>
> >> We've tried on up to 15-20 second runs of programs, and get the same
> >> result. We've tried making delay shorter (down to 0.01), and turned
> >> on #:threads. We've also tried setting
> >> (compile-context-preservation-enabled) to #t, and (eval-jit-enabled)
> >> to #f, and gotten roughly the same results. Do we need to be running
> >> for minutes or hours? Does the profiler interact badly with `eval`?
> >> We do see *some* calls to the runtime of the language, so they are
> >> recognized, but the time is clearly not being spent all at the top
> >> level of the profiler.
> >>
> >> Our test script is essentially:
> >>
> >> (time (profile-thunk (lambda ()
> >> (eval-in-our-lang (file->string "some-file"))))
> >>
> >>
> >> Some sample output is below. We're confident that, for example, our
> >> implementation of `extend` is called quite often (and is attributed
> >> 12% of top-of-stack time), but I don't know if I should believe that
> >> proportion with respect to the other functions. Any thoughts on where
> >> that 76% of time is actually spent? Note that about 1/3 of the time
> >> is spent in GC, so maybe that is affecting the times that things are
> >> sampled?
> >>
> >> $ racket benchmarks.rkt
> >> Running list-creation
> >> Profiling results
> >> -----------------
> >> Total cpu time observed: 15004ms (out of 15212ms)
> >> Number of samples taken: 568 (once every 26ms)
> >>
> >> =============================================================
> >> Caller
> >> Idx Total Self Name+src Local%
> >> ms(pct) ms(pct) Callee
> >> =============================================================
> >> [1] 15004(100.0%) 0(0.0%) [running body]
> >> ...ang/src/tests/benchmarks.rkt:##f
> >> run-tests14 [4] 100.0%
> >> -------------------------------------------------------------
> >> map/stx [19] 100.0%
> >> [2] 26(0.2%) 0(0.0%) temp155 (unknown source)
> >> parse-block [18] 100.0%
> >> -------------------------------------------------------------
> >> map/stx [19] 100.0%
> >> [3] 14(0.1%) 0(0.0%) parse-case-branch
> >> ...ang/src/lang/parser.rkt:219:0
> >> parse-block [18] 100.0%
> >> -------------------------------------------------------------
> >> [running body] [1] 100.0%
> >> [4] 15004(100.0%) 0(0.0%) run-tests14
> >> ...collects/rackunit/text-ui.rkt:243:0
> >> apply-test-suite [5] 100.0%
> >> -------------------------------------------------------------
> >> run-tests14 [4] 100.0%
> >> [5] 15004(100.0%) 0(0.0%) apply-test-suite
> >> ...it/private/test-suite.rkt:60:0
> >> the-tests [6] 100.0%
> >> -------------------------------------------------------------
> >> apply-test-suite [5] 100.0%
> >> [6] 15004(100.0%) 0(0.0%) the-tests (unknown source)
> >> profile-thunk12 [7] 100.0%
> >> -------------------------------------------------------------
> >> the-tests [6] 100.0%
> >> [7] 15004(100.0%) 0(0.0%) profile-thunk12
> >> ...t/collects/profile/main.rkt:9:0
> >> run [8] 100.0%
> >> -------------------------------------------------------------
> >> profile-thunk12 [7] 100.0%
> >> [8] 15004(100.0%) 11402(76.0%) run
> >> /usr/lib/racket/collects/profile/main.rkt:29:2
> >> extend [9] 12.8%
> >> add-brand [10] 4.2%
> >> string-map-ref5 [14] 4.1%
> >> get-raw-field [11] 2.5%
> >> eval-pyret/check [12] 0.2%
> >> get-field [13] 0.2%
> >> -------------------------------------------------------------
> >> run [8] 100.0%
> >> [9] 1926(12.8%) 1798(12.0%) extend
> >> ...ts/pyret-lang/src/lang/runtime.rkt:545:0
> >> foldr [15] 6.6%
> >> -------------------------------------------------------------
> >> run [8] 100.0%
> >> [10] 636(4.2%) 636(4.2%) add-brand
> >> ...pyret-lang/src/lang/runtime.rkt:524:0
> >> -------------------------------------------------------------
> >> run [8] 100.0%
> >> [11] 372(2.5%) 372(2.5%) get-raw-field
> >> ...t-lang/src/lang/runtime.rkt:439:0
> >> -------------------------------------------------------------
> >> run [8] 100.0%
> >> [12] 26(0.2%) 0(0.0%) eval-pyret/check
> >> .../src/tests/test-utils.rkt:60:0
> >> pyret->racket7 [16] 100.0%
> >> -------------------------------------------------------------
> >> run [8] 100.0%
> >> [13] 24(0.2%) 0(0.0%) get-field
> >> ...pyret-lang/src/lang/runtime.rkt:445:0
> >> string-map-ref5 [14] 100.0%
> >> -------------------------------------------------------------
> >> get-field [13] 3.7%
> >> run [8] 96.3%
> >> [14] 642(4.3%) 642(4.3%) string-map-ref5
> >> ...ang/src/lang/string-map.rkt:8:0
> >> -------------------------------------------------------------
> >> extend [9] 100.0%
> >> [15] 128(0.9%) 128(0.9%) foldr
> >> ...ong/whalesong/lang/private/list.rkt:251:2
> >> -------------------------------------------------------------
> >> eval-pyret/check [12]100.0%
> >> [16] 26(0.2%) 0(0.0%) pyret->racket7
> >> ...yret-lang/src/lang/eval.rkt:34:0
> >> parse-program [17] 100.0%
> >> -------------------------------------------------------------
> >> pyret->racket7 [16] 100.0%
> >> [17] 26(0.2%) 0(0.0%) parse-program
> >> ...ret-lang/src/lang/parser.rkt:33:0
> >> parse-block [18] 100.0%
> >> -------------------------------------------------------------
> >> parse-case-branch [3] 17.9%
> >> temp155 [2] 41.0%
> >> parse-program [17] 41.0%
> >> [18] 26(0.2%) 0(0.0%) parse-block
> >> ...pyret-lang/src/lang/parser.rkt:56:0
> >> map/stx [19] 100.0%
> >> -------------------------------------------------------------
> >> map/stx [19] 10.8%
> >> temp733 [21] 10.8%
> >> parse-block [18] 78.5%
> >> [19] 26(0.2%) 0(0.0%) map/stx
> >> ...uts/pyret-lang/src/lang/parser.rkt:28:0
> >> temp155 [2] 33.8%
> >> temp152 [22] 23.1%
> >> temp733 [21] 10.8%
> >> map/stx [19] 10.8%
> >> parse-case-branch [3] 10.8%
> >> temp596 [20] 10.8%
> >> -------------------------------------------------------------
> >> map/stx [19] 100.0%
> >> [20] 14(0.1%) 0(0.0%) temp596 (unknown source)
> >> dots-loop [23] 100.0%
> >> -------------------------------------------------------------
> >> map/stx [19] 100.0%
> >> [21] 14(0.1%) 0(0.0%) temp733 (unknown source)
> >> map/stx [19] 100.0%
> >> -------------------------------------------------------------
> >> map/stx [19] 100.0%
> >> [22] 12(0.1%) 12(0.1%) temp152 (unknown source)
> >> -------------------------------------------------------------
> >> temp596 [20] 100.0%
> >> [23] 14(0.1%) 14(0.1%) dots-loop (unknown source)
> >> -------------------------------------------------------------
> >> cpu time: 15228 real time: 15247 gc time: 5208
> >> ____________________
> >> Racket Users list:
> >> http://lists.racket-lang.org/users
> >
> >
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.racket-lang.org/users/archive/attachments/20130711/b785f341/attachment-0001.html>