On 30 Srpen 2012, 17:53, Robert Haas wrote:
> On Fri, Aug 24, 2012 at 6:36 PM, Tomas Vondra <tv@fuzzy.cz> wrote:
>> attached is a patch that improves performance when dropping multiple
>> tables within a transaction. Instead of scanning the shared buffers for
>> each table separately, the patch removes this and evicts all the tables
>> in a single pass through shared buffers.
>>
>> Our system creates a lot of "working tables" (even 100.000) and we need
>> to perform garbage collection (dropping obsolete tables) regularly. This
>> often took ~ 1 hour, because we're using big AWS instances with lots of
>> RAM (which tends to be slower than RAM on bare hw). After applying this
>> patch and dropping tables in groups of 100, the gc runs in less than 4
>> minutes (i.e. a 15x speed-up).
>>
>> This is not likely to improve usual performance, but for systems like
>> ours, this patch is a significant improvement.
>
> Seems pretty reasonable. But instead of duplicating so much code,
> couldn't we find a way to use replace DropRelFileNodeAllBuffers with
> DropRelFileNodeAllBuffersList? Surely anyone who was planning to call
> the first one could instead call the second one with a count of one
> and a pointer to the address of the data they were planning to pass.
> I'd probably swap the order of arguments to
> DropRelFileNodeAllBuffersList as well. We could do something similar
> with smgrdounlink/smgrdounlinkall so that, again, only one copy of the
> code is needed.
Yeah, I was thinking about that too, but I simply wasn't sure which is the
best choice so I've sent the raw patch. OTOH these functions are called on
a very limited number of places, so a refactoring like this seems fine.
Tomas
