Re: MERGEQUERIES consumes 100% CPU ?

On Tue, May 19, 2009 at 8:25 AM, Andreas Schulze
<[email protected]> wrote:
> nobody may help ?

Sorry for not shooting you a reply.  Known problem.

http://marc.info/?l=djbdns&m=124047690620137&w=2
http://marc.info/?l=djbdns&m=124057562229449&w=2
http://marc.info/?l=djbdns&m=124065894002106&w=2

I'm sure Jeff will speak up when he gets a chance.  I believe he's
working on a solution with a better algorithm / data structure.

In the meantime, if anybody wants to take a stab at a solution, I
think these are the suspected bottlenecks:

static int try_merge(struct dns_transmit *d)
{
  int i;
  for (i = 0; i < MAXUDP; i++) {
    if (!inprogress[i]) continue;
    if (!merge_equal(d, inprogress[i])) continue;
    d->master = inprogress[i];
    inprogress[i]->slaves[inprogress[i]->nslaves++] = d;
    return 1;
  }
  return 0;
}

static void register_inprogress(struct dns_transmit *d)
{
  int i;
  for (i = 0; i < MAXUDP; i++) {
    if (!inprogress[i]) {
      inprogress[i] = d;
      return;
    }
  }
  strerr_die1x(100, "BUG: out of inprogress slots");
}

static void unregister_inprogress(struct dns_transmit *d)
{
  int i;
  for (i = 0; i < MAXUDP; i++) {
    if (inprogress[i] == d)
      inprogress[i] = 0;
  }
}

Though that really needs to be confirmed by profiling.