Re: Fwd: Re: Severe performance regression with latest CVS sources?

Paolo Amoroso <amoroso@xxxxxxxxx> writes:

>> twice that many hash tables with (make-hash-table)?  The other thing
>
> CL-USER> (time (dotimes (i (* 2 824))
>                  (make-hash-table)))
> ; Compiling LAMBDA NIL:
> ; Compiling Top-Level Form:
>
> ; Evaluation took:
> ;   6,648 CPU cycles


It's elided the creation, I'm afraid.  Can you do
  (defvar *foo*)
  (time (dotimes (i (* 2 824)) (setq *foo* (make-hash-table))))
?

> ;
> NIL
> CL-USER>
>
>
>> to look at is for collisions in your hash tables; check the EDGES-FROM
>> slot of something with a lot of direct subclasses, say STRUCTURE-CLASS
>> or CONDITION or NUMBER, and see if all of the entries are in the same
>> bucket of the hash table)
>
> I'm not sure I understand what you mean.  I used the SLIME inspector
> to inspect the output of, say, (find-class 'structure-class).  I
> include the output below, which does not seem to have any EDGES-FROM
> slot.

I think the edges-from should be in the graph node output record.  I'm
not entirely sure how to get at one of those, but you could start by
inspecting clim:*application-frame* from the slime inspector, finding
the listener interactor pane, finding its output history, and looking
at the output records in there.

> 588,259,376 |   1,111 | 15.158 |  0.01364 |  529,486 | 
> CLIM-INTERNALS::LAYOUT-EDGE-1
>                   1,111: #<Code Object "LAMBDA (FWRAPPER)" {5A58260F}>

>   1,470,608 |   2,222 |  0.056 |  0.00003 |      662 | 
> CLIM-INTERNALS::ENSURE-EDGE-RECORD
>                   1,111: #<Code Object "DEFUN LAYOUT-EDGE" {59718A4F}>
>                   1,111: #<Code Object "DEFUN LAYOUT-EDGE-1" {5A75B007}>


Well, that's moderately conclusive: the cost seems to be in
LAYOUT-EDGE-1, not in ENSURE-EDGE-RECORD.  Which is weird, because I
can't really see what LAYOUT-EDGE-1 is doing that is so ludicrously
expensive.  Can you do the same thing, but with the CLIM package
profiled additionally?  (And then maybe one with CLIM but not
CLIM-INTERNALS profiled?)  The first 10 or so entries in each profile
will probably suffice.  (Maybe CLEAR-OUTPUT-RECORD is expensive?)

Cheers,

Christophe