Re: Inefficiency in SLRU stats collection

[Fujii Masao]

On 2020/05/13 11:26, Kyotaro Horiguchi wrote:

At Tue, 12 May 2020 15:50:35 -0400, Tom Lane <t...@sss.pgh.pa.us> wrote in

I happened to come across this code added by 28cac71bd:

static PgStat_MsgSLRU *
slru_entry(SlruCtl ctl)
{
     int        idx = pgstat_slru_index(ctl->shared->lwlock_tranche_name);

     Assert((idx >= 0) && (idx < SLRU_NUM_ELEMENTS));

     return &SLRUStats[idx];
}

which is invoked by all the pgstat_count_slru_XXX routines.
This seems mightily inefficient --- the count functions are
just there to increment integer counters, but first they
have to do up to half a dozen strcmp's to figure out which
counter to increment.

We could improve this by adding another integer field to
SlruSharedData (which is already big enough that no one
would notice) and recording the result of pgstat_slru_index()
there as soon as the lwlock_tranche_name is set.  (In fact,
it looks like we could stop saving the tranche name as such
altogether, thus buying back way more shmem than the integer
field would occupy.)

Sounds good to me.

Regards,

--
Fujii Masao
Advanced Computing Technology Center
Research and Development Headquarters
NTT DATA CORPORATION