On Mar 26, 2007, at 2:01 AM, Galy Lee wrote:
> As AUTOVACUUM is having multiple workers now, the semantics of
> autovacuum_cost_limit also need to be redefined.
>
> Currently, autovacuum_cost_limit is the accumulated cost that will
> cause
> one single worker vacuuming process to sleep. It is used to restrict
> the I/O consumption of a single vacuum worker. When there are N
> workers,
> the I/O consumption by autovacuum workers can be increased by N times.
> This autovacuum_cost_limit semantics produces unpredictable I/O
> consumption for multiple-autovacuum-workers.
>
> One simple idea is to set cost limit for every worker to:
> autovacuum_cost_limit / max_autovacuum_workers. But for scenarios
> which
> have fewer active workers, it is obvious unfair to active workers.
> So a
> better way is to set cost limit of every active worker to:
> autovacuum_cost_limit/autovacuum_active_workers. This ensures the I/O
> consumption of autovacuum is stable.
>
> Worker can be extended to have its own cost_limit on share memory.
> When
> a worker is brought up or a worker has finished its work, launcher
> recalculates:
>
> worker_cost_limit= (autovacuum_cost_limit/
> autovacuum_active_workers)
>
> and sets new value for each active workers.
>
> The above approach requires launcher can change cost delay setting of
> workers on-the-fly. This can be achieved by forcing VACUUM refers
> to the
> cost delay setting in its worker’s share memory every
> vacuum_delay_point.
>
> Any comments or suggestions?
Well, ideally we'd set cost limit settings on a per-tablespace
basis... but I agree that what you propose is probably the best bet
for multiple daemons short of doing per-tablespace stuff.
--
Jim Nasby jim@nasby.net
EnterpriseDB http://enterprisedb.com 512.569.9461 (cell)