[Openmcl-devel] Profiling lisp code
David L. Rager
ragerdl at cs.utexas.edu
Mon May 24 08:12:43 PDT 2010
Someone who uses this macro and can't get it to work for them might be
interested in the following two ideas:
(1) Change (get-real-time) to (get-internal-real-time)
(2) Remove the eval and let the body evaluate as normal. It may be
the case that if one really wants to use eval/apply, one would need to
create a closure first.
I should disclaim that I don't have the macro fully functioning for my
use case, but it appears that I've gotten it to a point where it's
useful. It's quite neat -- thanks for putting it together.
On Fri, Nov 20, 2009 at 9:52 AM, Taoufik Dachraoui
<taoufik.dachraoui at wanadoo.fr> wrote:
> Hi
> I added a second profiler report with hierarchical details, any suggestions
> for
> improvements or corrections are welcome:
> (defmacro with-profiler (syms &rest body)
> `(let ((*profile-hash-1* (make-hash-table))
> (*profile-hash-2* (make-hash-table :test #'equal))
> (*profile-stack* nil))
> (flet ((before-trace (f &rest vals)
> (declare (ignore vals))
> (let ((tim (get-real-time)))
> (multiple-value-bind (last count elapsed) (values-list (gethash f
> *profile-hash-1*))
> (declare (ignore last))
> (setf (gethash f *profile-hash-1*) (list tim (+ 1 count) elapsed)))
> (push f *profile-stack*)
> (multiple-value-bind (last count elapsed) (values-list (gethash
> *profile-stack* *profile-hash-2*))
> (if (null last) (setf elapsed 0 count 0))
> (setf (gethash *profile-stack* *profile-hash-2*) (list tim (+ 1 count)
> elapsed)))))
> (after-trace (f &rest vals)
> (declare (ignore vals))
> (let ((tim (get-real-time)))
> (multiple-value-bind (last count elapsed) (values-list (gethash f
> *profile-hash-1*))
> (setf (gethash f *profile-hash-1*) (list tim count (+
> elapsed (- tim last)))))
> (multiple-value-bind (last count elapsed) (values-list (gethash
> *profile-stack* *profile-hash-2*))
> (setf (gethash *profile-stack* *profile-hash-2*) (list tim
> count (+ elapsed (- tim last)))))
> (pop *profile-stack*))))
> (dolist (s ',syms)
> (format t "profiling ~S~%" s) (force-output t)
> (setf (gethash s *profile-hash-1*) '(nil 0 0))
> (CCL::%UNTRACE-0 (list s))
> (CCL::%TRACE-0 (list (list s :before #'before-trace :after
> #'after-trace))
> 'NIL))
> (format t "~%profiling ~A~%" (trace))
> (let ((total-elapsed (get-real-time)) (lst nil) (nano 1000000000))
> (eval '(progn , at body))
> (setf total-elapsed (- (get-real-time) total-elapsed))
> (format t "total elapsed time ~10,3F seconds~%" (/ total-elapsed nano))
> (dolist (s ',syms)
> (CCL::%UNTRACE-0 (list s)))
> (maphash #'(lambda (k v)
> (if (> (second v) 0)
> (push (list k (second v) (third v)) lst)))
> *profile-hash-1*)
> (format t "~35A ~8A ~14A~%" "Function" "Count" "Elapsed")
> (dolist (e (sort lst #'(lambda (e f) (< (third e) (third f)))))
> (format t "~30S ~10d ~10,3F ~5,2F%~%"
> (first e)
> (second e)
> (/ (third e) nano)
> (* 100 (/ (third e) (if (= 0 total-elapsed) 1 total-elapsed)))
> ))
> (format t "~%total elapsed time ~10,3F seconds~%" (/ total-elapsed nano))
> (setf lst nil)
> (maphash #'(lambda (k v)
> (if (> (second v) 0)
> (push (list (reverse k) (second v) (third v)) lst)))
> *profile-hash-2*)
> (labels ((compare-list (l1 l2)
> (if (null l1)
> t
> (if (null l2)
> nil
> (if (string< (string (car l1)) (string (car l2)))
> t
> (if (string> (string (car l1)) (string (car l2)))
> nil
> (compare-list (cdr l1) (cdr l2))))))))
> (let ((prefix "") (times (make-hash-table)))
> (setf (gethash 0 times) total-elapsed) ; level 0
> (dolist (e (sort lst #'(lambda (e f) (compare-list (first e) (first
> f)))))
> (setf (gethash (length (first e)) times) (third e))
> (setf prefix (make-string (* 2 (length (first e))) :initial-element
> #\.))
> (let ((elapsed (gethash (- (length (first e)) 1) times))) ; elapsed
> time of previous level
> (format t "~A~S ~10,3F ~5,2F%~%" prefix
> (car (last (first e)))
> (/ (third e) nano)
> (* 100 (/ (third e) (if (= 0 elapsed) 1 elapsed)))
> )))))))))
>
> Example of reports:
> total elapsed time 0.847 seconds
> Function Count Elapsed
> %LIST 4 0.000 0.01%
> CENTER-LAYOUT 4 0.000 0.01%
> DESTROY-CONTEXT 1 0.000 0.06%
> %GET-VFIELDS 1 0.001 0.08%
> XCREATEWINDOW 69 0.001 0.09%
> %EVAL 18 0.001 0.11%
> %LIST-SYMBOLS 12 0.001 0.11%
> %GEN-FLET-VFIELDS 4 0.001 0.17%
> %FLET-VFIELDS 1 0.001 0.18%
> %VFIELDS 1 0.002 0.18%
> COLUMN-LAYOUT 11 0.002 0.27%
> RESIZE-WINDOW 69 0.002 0.29%
> %SPLIT-ARGS 272 0.003 0.36%
> ROW-LAYOUT 16 0.003 0.37%
> SPLIT-ARGS 272 0.010 1.13%
> CLIP 27 0.013 1.52%
> GETA 2479 0.027 3.14%
> CREATE-WINDOW 69 0.037 4.41%
> CREATE-CONTEXT 69 0.046 5.43%
> total elapsed time 0.847 seconds
> ..%EVAL 0.001 0.13%
> ....%EVAL 0.001 97.22%
> ......%EVAL 0.001 88.97%
> ........%EVAL 0.001 80.29%
> ..........%EVAL 0.000 1.62%
> ..........%GET-VFIELDS 0.001 83.75%
> ............GETA 0.000 7.86%
> ............SPLIT-ARGS 0.000 60.71%
> ..............%SPLIT-ARGS 0.000 33.20%
> ..%VFIELDS 0.002 0.18%
> ....%FLET-VFIELDS 0.001 98.27%
> ......%GEN-FLET-VFIELDS 0.001 95.39%
> ........%LIST 0.000 3.62%
> ........%LIST-SYMBOLS 0.001 67.14%
> ..CENTER-LAYOUT 0.000 0.01%
> ..CLIP 0.013 1.52%
> ....GETA 0.001 6.60%
> ....RESIZE-WINDOW 0.001 8.26%
> ..COLUMN-LAYOUT 0.002 0.27%
> ....GETA 0.000 9.94%
> ..CREATE-CONTEXT 0.046 5.43%
> ....CREATE-WINDOW 0.037 81.25%
> ......GETA 0.016 43.95%
> ......XCREATEWINDOW 0.001 2.06%
> ....GETA 0.001 3.10%
> ..DESTROY-CONTEXT 0.000 0.06%
> ..GETA 0.007 0.86%
> ..RESIZE-WINDOW 0.001 0.16%
> ..ROW-LAYOUT 0.003 0.37%
> ....GETA 0.000 10.53%
> ..SPLIT-ARGS 0.009 1.08%
> ....%SPLIT-ARGS 0.003 31.65%
> NIL
> ?
>
>
> On Nov 19, 2009, at 7:41 PM, Paul Krueger wrote:
>
> Taoufik,
> I really like this idea, but unfortunately the code as written didn't work
> for me.
> I first created a couple of test functions:
> (defun test-factorial (n)
> (if (<= n 1)
> 1
> (* n (test-factorial (1- n)))))
> (defun test-profile ()
> (with-profiler (test-factorial)
> (dotimes (i 1000)
> (test-factorial (random 50)))))
> but on executing the test-profile function the macro-expansion caused the
> following:
> (test-profile)
> ;Compiler warnings :
> ; In an anonymous lambda form inside (CCL::TRACED TEST-FACTORIAL):
> Undeclared free variable *PROFILER* (2 references)
> ; In an anonymous lambda form inside (CCL::TRACED TEST-FACTORIAL):
> Undeclared free variable *PROFILER* (2 references)
> profiling (TEST-FACTORIAL)
>> Error: Unbound variable: *PROFILER*
>> While executing: (:INTERNAL (CCL::TRACED TEST-FACTORIAL)), in process
>> Listener(6).
>> Type cmd-. to abort, cmd-\ for a list of available restarts.
>> Type :? for other options.
> 1 >
> This is caused by the use of the ` form for the trace-0 argument calls
> which effectively hides the let declaration of *profiler* because you are
> passing a list that contains the unevaluated lambda form to trace rather
> than passing a list that contains the function that results from evaluation
> of that form within the scope of the let. Even if we fixed just that
> problem, this code would also create new anonymous :before and :after
> functions for each symbol being profiled when in fact you only need one for
> each hash-table.
> So I modified your code a bit to define the functions just once and to
> eliminate some of the ` forms which caused problems:
> (defmacro with-profiler (syms &rest body)
> `(let ((*profile-hash* (make-hash-table :test #'equal)))
> (flet ((before-trace (f &rest vals)
> (declare (ignore vals))
> (multiple-value-bind (last count elapsed) (values-list
> (gethash f *profile-hash*))
> (declare (ignore last))
> (setf (gethash f *profile-hash*) (list (get-real-time) (+ 1
> count) elapsed))))
> (after-trace (f &rest vals)
> (declare (ignore vals))
> (multiple-value-bind (last count elapsed) (values-list
> (gethash f *profile-hash*))
> (let ((tim (get-real-time)))
> (setf (gethash f *profile-hash*) (list tim count (+
> elapsed (- tim last))))))))
> (dolist (s ',syms)
> (setf (gethash s *profile-hash*) '(nil 0 0))
> (CCL::%UNTRACE-0 (list s))
> (CCL::%TRACE-0 (list (list s :before #'before-trace :after
> #'after-trace))
> 'NIL))
> (format t "~%profiling ~A~%" (trace))
> (let ((elapsed (get-real-time)))
> (eval '(progn , at body))
> (format t "total elapsed time ~10,3F seconds~%" (/ (-
> (get-real-time) elapsed) 1000000000)))
> (dolist (s ',syms)
> (CCL::%UNTRACE-0 (list s))
> (let ((lst nil) (total 0))
> (maphash #'(lambda (k v)
> (incf total (/ (third v) 1000000000))
> (push (list k (second v) (third v)) lst))
> *profile-hash*)
> (format t "~35A ~8A ~14A~%" "Function" "Count" "Elapsed")
> (dolist (e (sort lst #'(lambda (e f) (< (third e) (third f)))))
> (format t "~30S ~10d ~10,3F ~5,2F%~%"
> (first e) (second e) (/ (third e) 1000000000) (* 100 (/
> (/ (third e) 1000000000) total)))))))))
> Using this code everything worked fine:
> ? (test-profile)
> profiling (TEST-FACTORIAL)
> total elapsed time 0.021 seconds
> Function Count Elapsed
> TEST-FACTORIAL 1000 0.011 100.00%
> NIL
>
> Regards,
> Paul
>
> On Nov 19, 2009, at 9:54 AM, Taoufik Dachraoui wrote:
>
> The elapsed time is in nanoseconds; in the new version of with-profiler the
> times are in
> seconds.
> I modified the with-profiler macro to sort the elapsed times and added a
> column to show the percentage of the elapsed times :
> (defmacro with-profiler (syms &rest body)
> `(let ((*profiler* (make-hash-table :test #'equal)))
> (dolist (s ',syms)
> (setf (gethash `,s *profiler*) '(nil 0 0))
> (CCL::%UNTRACE-0 `(,s))
> (CCL::%TRACE-0 `((,s :before #'(lambda (f &rest vals)
> (declare (ignore vals))
> (multiple-value-bind (last count elapsed) (values-list (gethash f
> *profiler*))
> (declare (ignore last))
> (setf (gethash f *profiler*) (list (get-real-time) (+ 1 count) elapsed))))
> :after #'(lambda (f &rest vals)
> (declare (ignore vals))
> (multiple-value-bind (last count elapsed) (values-list (gethash f
> *profiler*))
> (let ((tim (get-real-time)))
> (setf (gethash f *profiler*) (list tim count (+ elapsed (- tim
> last)))))))))
> 'NIL))
> (format t "~%profiling ~A~%" (trace))
> (let ((elapsed (get-real-time)))
> (eval '(progn , at body))
> (format t "total elapsed time ~10,3F seconds~%" (/ (- (get-real-time)
> elapsed) 1000000000)))
> (dolist (s ',syms)
> (CCL::%UNTRACE-0 `(,s)))
> (let ((lst nil) (total 0))
> (maphash #'(lambda (k v)
> (incf total (/ (third v) 1000000000))
> (push (list k (second v) (third v)) lst))
> *profiler*)
> (format t "~35A ~8A ~14A~%" "Function" "Count" "Elapsed")
> (dolist (e (sort lst #'(lambda (e f) (< (third e) (third f)))))
> (format t "~30S ~10d ~10,3F ~5,2F%~%"
> (first e) (second e) (/ (third e) 1000000000) (* 100 (/ (/ (third e)
> 1000000000) total)))))
> ))
> Output example:
> total elapsed time 1.159 seconds
> Function Count Elapsed
> SEND-MESSAGE-TO-WINDOW 0 0.000 0.00
> LEFT-LAYOUT 0 0.000 0.00
> RROW-LAYOUT 0 0.000 0.00
> RIGHT-LAYOUT 0 0.000 0.00
> RCOLUMN-LAYOUT 0 0.000 0.00
> GET-EVENT-HANDLER 0 0.000 0.00
> %LIST 4 0.000 0.01
> CENTER-LAYOUT 4 0.000 0.01
> %LIST-SYMBOLS 12 0.000 0.03
> %GET-VFIELDS 1 0.000 0.06
> %GEN-FLET-VFIELDS 4 0.001 0.07
> %FLET-VFIELDS 1 0.001 0.08
> %VFIELDS 1 0.001 0.08
> DESTROY-CONTEXT 1 0.001 0.08
> XCREATEWINDOW 69 0.001 0.09
> %EVAL 20 0.001 0.09
> %SPLIT-ARGS 254 0.003 0.34
> RESIZE-WINDOW 69 0.003 0.37
> COLUMN-LAYOUT 11 0.005 0.65
> SPLIT-ARGS 254 0.008 1.05
> ROW-LAYOUT 16 0.008 1.13
> CLIP 27 0.023 3.06
> GETA 19185 0.172 23.17
> CREATE-WINDOW 69 0.253 34.02
> CREATE-CONTEXT 69 0.264 35.60
> NIL
> ?
> Taoufik
> On Nov 19, 2009, at 4:44 PM, Alexander Repenning wrote:
>
> looks great to me. It seems to use similar functions as the timer code I
> pointed to in my previous post but includes a nice profilers wrap up macro.
> I suggest to use time formatting functions similar to the code I pointed to
> because it is not clear what "elapsed=268738" really means. Just a detail.
> best, Alex
> On Nov 19, 2009, at 7:24 AM, Taoufik Dachraoui wrote:
>
> Hi,
>
> I created a macro with-profiler and would like to know the
> disadvantages of this solution
> compared to others:
>
> ; i took the following 2 functions from ccl-1.4/tools/advice-profiler/
> profiler.lisp
> ; the with-profiler can be modified to return cpu-time, user-time and
> real-time used
> ; by all profiled functions
>
> (defun mach-timespec->nanoseconds (ts)
> "Convert the given typespec structure into nanoseconds."
> (+ (* 1000000000 (pref ts :mach_timespec.tv_sec))
> (pref ts :mach_timespec.tv_nsec)))
>
> (defun get-real-time ()
> (let ((clock-port (make-record :clock_serv_t)))
> (#_host_get_clock_service (#_mach_host_self) #$REALTIME_CLOCK
> clock-port)
> (ccl:rlet ((ts :mach_timespec))
> (#_clock_get_time (%get-ptr clock-port) ts)
> (mach-timespec->nanoseconds ts))))
>
> (defmacro with-profiler (syms &rest body)
> `(let ((*profiler* (make-hash-table :test #'equal)))
> (dolist (s ',syms)
> (setf (gethash `,s *profiler*) '(nil 0 0))
> (CCL::%UNTRACE-0 `(,s))
> (CCL::%TRACE-0 `((,s :before #'(lambda (f &rest vals)
> (declare (ignore vals))
> (multiple-value-bind (last count elapsed) (values-list
> (gethash f *profiler*))
> (declare (ignore last))
> (setf (gethash f *profiler*) (list (get-real-time) (+ 1 count)
> elapsed))))
> :after #'(lambda (f &rest vals)
> (declare (ignore vals))
> (multiple-value-bind (last count elapsed) (values-list
> (gethash f *profiler*))
> (let ((tim (get-real-time)))
> (setf (gethash f *profiler*) (list tim count (+ elapsed (- tim
> last)))))))))
> 'NIL))
> (format t "~%profiling ~A~%" (trace))
> (let ((elapsed (get-real-time)))
> (eval '(progn , at body))
> (format t "total elapsed time ~S~%" (- (get-real-time) elapsed)))
> (dolist (s ',syms)
> (CCL::%UNTRACE-0 `(,s)))
> (maphash #'(lambda (k v) (format t "function=~S count=~S
> elapsed=~S~%" k (second v) (third v))) *profiler*)))
>
> Tests:
>
> ? (defun fact (n) (if (= 0 n) 1 (* n (fact (- n 1)))))
> ? (defun f () (sleep 1))
> ? (defmacro foo () (progn (f) ''foo))
> ? (with-profiler (f fact)
> (dotimes (i 10)
> (f)
> (foo)
> (fact (random 100))))
>
> profiling (FACT F)
> total elapsed time 11003452329
> function=FACT count=10 elapsed=268738
> function=F count=11 elapsed=11001194423
> NIL
> ?
>
> Any comments on this method of profiling lisp functions are welcome
>
> Kind regards
> Taoufik
>
> On Nov 18, 2009, at 5:57 PM, Matt Tenenbaum wrote:
>
>
> Norvig's classic "Paradigms of Artificial Intelligence Programming"
>
> has a section on instrumentation in chapter 9, including code for
>
> profiling in a manner analogous to TRACE. In looking quickly, this
>
> code doesn't seem to be included in the distributed source for the
>
> book (http://norvig.com/paip/README.html), but you might find that
>
> chapter helpful.
>
> Cheers,
>
> -mt
>
> On Wed, Nov 18, 2009 at 8:36 AM, Taoufik Dachraoui
>
> <taoufik.dachraoui at wanadoo.fr> wrote:
>
> Hi
>
> How do I profile lisp program to know how much time each called
>
> function takes
>
> Regards
>
> Taoufik
>
>
>
> _______________________________________________
>
> Openmcl-devel mailing list
>
> Openmcl-devel at clozure.com
>
> http://clozure.com/mailman/listinfo/openmcl-devel
>
>
>
>
>
> _______________________________________________
> Openmcl-devel mailing list
> Openmcl-devel at clozure.com
> http://clozure.com/mailman/listinfo/openmcl-devel
>
> Prof. Alexander Repenning
> University of Colorado
> Computer Science Department
> Boulder, CO 80309-430
> vCard: http://www.cs.colorado.edu/~ralex/AlexanderRepenning.vcf
>
>
> _______________________________________________
> Openmcl-devel mailing list
> Openmcl-devel at clozure.com
> http://clozure.com/mailman/listinfo/openmcl-devel
>
>
>
> _______________________________________________
> Openmcl-devel mailing list
> Openmcl-devel at clozure.com
> http://clozure.com/mailman/listinfo/openmcl-devel
>
>
More information about the Openmcl-devel
mailing list