On Wed, Aug 3, 2011 at 2:49 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Martijn van Oosterhout <kleptog@svana.org> writes:
>> On Wed, Aug 03, 2011 at 02:21:25PM -0400, Robert Haas wrote:
>>> It would be nice if the Linux guys would fix this problem for us, but
>>> I'm not sure whether they will. For those who may be curious, the
>>> problem is in generic_file_llseek() in fs/read-write.c. On a platform
>>> with 8-byte atomic reads, it seems like it ought to be very possible
>>> to read inode->i_size without taking a spinlock.
>
>> Interesting. There's this thread from 2003 suggesting the use of pread
>> instead, it was rejected on the argument that lseek is cheap so not a
>> problem.
>
>> http://archives.postgresql.org/pgsql-patches/2003-02/msg00197.php
>
> That seems rather unrelated. The point here is our use of lseek to find
> out the current file size --- or at least, I would hope they're not
> trying to read the inode's file size in a SEEK_CUR call.
Correct.
> The reason "-M prepared" helps is presumably that it eliminates most of
> the RelationGetNumberOfBlocks calls the planner does to check current
> table size. While we could certainly consider using a cheaper (possibly
> more stale) value there, it's a bit astonishing to think that that's the
> main cost in a parse/plan/execute cycle. Perhaps there are more hotspot
> calls than that one?
Nope.
On a straight pgbench -S test, you get four system calls per query:
recvfrom(), lseek(), lseek(), sendto(). Adding -M prepared eliminates
the two lseeks.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company