Re: Slowdown around 5.0.5-dev-0

[Sebastian Gniazdowski <sgniazdowski@xxxxxxxxx>]

Zprof says there is no change:
https://youtu.be/6A-AKesvZKM

I also attach zprof result in file. BTW, the patch didn't apply to
head or to 23f98c3, I had to make changes manually.

Best regards,
Sebastian Gniazdowski

On 11 October 2015 at 08:20, Bart Schaefer <schaefer@xxxxxxxxxxxxxxxx> wrote:
> On Oct 10,  5:06pm, Bart Schaefer wrote:
> }
> } This suggests a couple of possible fixes, but I've run out of time to
> } experiment right now.
>
> OK, this is the least intrusive change I can think of that might make
> a difference.  What this attempts to do is maintain a second pointer
> into the list of heaps that tells freeheap() where to start working.
> It's set on pushheap() and reset on popheap() so it freeheap() should
> avoid searching for free space in some arenas that were completely
> filled before pushheap() was called.  However, it won't improve the
> performance if there are a whole lot of heap arenas each with a small
> amount of space available at the time of pushheap().
>
> To improve THAT situation, I think we'd have to be willing to "leak"
> those small pieces of heap until a popheap() happens, and instead
> use at least one new arena following most calls to pushheap().  This
> might mean that zsh's memory footprint grows, but on modern hardware
> that may not be an issue.
>
> All tests pass with this change in place, let's see what it does with
> Sebastian's 89k array elements.
>
>
> diff --git a/Src/mem.c b/Src/mem.c
> index b9569ea..01072a3 100644
> --- a/Src/mem.c
> +++ b/Src/mem.c
> @@ -127,6 +127,12 @@ static Heap heaps;
>
>  static Heap fheap;
>
> +/* same as fheap, except in the preceding stack context (it will be the
> + * first heap from which space in the current context may be recovered
> + * when heap space is freed) */
> +
> +static Heap fpop;
> +
>  /**/
>  #ifdef ZSH_HEAP_DEBUG
>  /*
> @@ -293,8 +299,11 @@ pushheap(void)
>      h_push++;
>  #endif
>
> +    fpop = NULL;
>      for (h = heaps; h; h = h->next) {
>         DPUTS(!h->used, "BUG: empty heap");
> +       if (!fpop && h->used < ARENA_SIZEOF(h))
> +           fpop = h;
>         hs = (Heapstack) zalloc(sizeof(*hs));
>         hs->next = h->sp;
>         h->sp = hs;
> @@ -341,9 +350,10 @@ freeheap(void)
>       *
>       * However, if the arena to which fheap points is unused, we want to
>       * free it, so we have no choice but to do the sweep for a new fheap.
> +     * fpop is the first heap with free space following pushheap().
>       */
>      if (fheap && !fheap->sp)
> -       fheap = NULL;   /* We used to do this unconditionally */
> +       fheap = fpop;
>      /*
>       * In other cases, either fheap is already correct, or it has never
>       * been set and this loop will do it, or it'll be reset from scratch
> @@ -417,7 +427,7 @@ popheap(void)
>      h_pop++;
>  #endif
>
> -    fheap = NULL;
> +    fheap = fpop = NULL;
>      for (h = heaps; h; h = hn) {
>         hn = h->next;
>         if ((hs = h->sp)) {
> @@ -443,6 +453,8 @@ popheap(void)
>  #endif
>             if (!fheap && h->used < ARENA_SIZEOF(h))
>                 fheap = h;
> +           if (!fpop && h->sp && h->sp->used < ARENA_SIZEOF(h))
> +               fpop = h;
>             zfree(hs, sizeof(*hs));
>
>             hl = h;

patched:
# time ( npanelize < ~/lsoflsof )
num  calls                time                       self            name
-----------------------------------------------------------------------------------
 1)  161        5453,04    33,87   43,53%   5453,04    33,87   43,53%  _nlist_print_with_ansi
 2)    1       12500,28 12500,28   99,79%   4052,65  4052,65   32,35%  n-list
 3)    7        8041,91  1148,84   64,20%   2588,87   369,84   20,67%  n-list-draw
 4)    7         327,91    46,84    2,62%    322,30    46,04    2,57%  n-list-input
 5)    7          40,77     5,82    0,33%     31,97     4,57    0,26%  _nlist_setup_user_vars
 6)    7          28,91     4,13    0,23%     28,91     4,13    0,23%  _nlist_status_msg
 7)    2          24,94    12,47    0,20%     24,94    12,47    0,20%  _nlist_cursor_visibility
 8)    1          26,79    26,79    0,21%      9,97     9,97    0,08%  _nlist_exit
( n-panelize < ~/lsoflsof; )  14,33s user 0,27s system 94% cpu 15,488 total

clean:
# time ( npanelize < ~/lsoflsof )
num  calls                time                       self            name
-----------------------------------------------------------------------------------
 1)  161        5410,49    33,61   43,57%   5410,49    33,61   43,57%  _nlist_print_with_ansi
 2)    1       12391,49 12391,49   99,78%   4030,62  4030,62   32,46%  n-list
 3)    7        7961,76  1137,39   64,11%   2551,27   364,47   20,54%  n-list-draw
 4)    7         324,36    46,34    2,61%    318,36    45,48    2,56%  n-list-input
 5)    7          39,22     5,60    0,32%     31,44     4,49    0,25%  _nlist_setup_user_vars
 6)    7          27,86     3,98    0,22%     27,86     3,98    0,22%  _nlist_status_msg
 7)    2          24,35    12,18    0,20%     24,35    12,18    0,20%  _nlist_cursor_visibility
 8)    1          27,05    27,05    0,22%     10,37    10,37    0,08%  _nlist_exit
( n-panelize < ~/lsoflsof; )  14,19s user 0,29s system 94% cpu 15,319 total