From: "Qingqing Zhou" <zhouqq@cs.toronto.edu>
Subject: Re: Help in avoiding a query 'Warm-Up' period/shared buffer cache
Date: Thu, 5 Jan 2006 18:12:13 -0500
Organization: Hub.Org Networking Services
Lines: 31
Message-ID: <dpk922$slo$1@news.hub.org>
References: <200601041449.43557.mliberman@mixedsignals.com>
To: pgsql-performance@postgresql.org


"Mark Liberman" <mliberman@mixedsignals.com> wrote
>
> First run, after a night of inactivity:
>
>                     ->  Bitmap Index Scan on 1min_events_file_id_begin_idx
> (cost=0.00..37.85 rows=3670 width=0) (actual time=313.468..313.468 
> rows=11082
> loops=1)
>                           Index Cond: (file_id = 137271)
> Total runtime: 313.643 ms
>
> Second run, after that:
>
>                     ->  Bitmap Index Scan on 1min_events_file_id_begin_idx
> (cost=0.00..37.85 rows=3670 width=0) (actual time=2.106..2.106 rows=11082
> loops=1)
>                           Index Cond: (file_id = 137271)
> Total runtime: 2.276 ms

It is clear that the first query takes longer time because of the IO time of 
index 1min_events_file_id_begin_idx (see 313.468 vs. 2.106). I am afraid 
currently there is no easy solution for this situation, unless you could 
predicate which part of relation/index your query will use, then you can 
preload or "warm-up" cache for it.

Regards,
Qingqing