Re: COMMIT NOWAIT Performance Option - Mailing list pgsql-hackers

On 2/28/07, Gregory Stark <stark@enterprisedb.com> wrote:
> But we've already seen that CRC checks can be expensive. Not everyone will
> want to take the cpu hit. Storing a byte counter in every block is cheap.

CRC checking a page is most certainly the simplest.  And, I disagree
that it would be worse than either a sequence counter or the full page
write.  Block checksumming is done at read/write time... which is
something that needs to be improved anyway.  With a properly tuned
bgwriter, the write itself should barely be noticeable.  How fast is a
CRC of 8K?  Last time I checked it was something on the scale of ~95
usec for CRC32 and ~33 usec for sb8.

> And the idea came from what someone said MSSQL does, so "like everyone else"
> -- which isn't a very compelling argument to begin with -- doesn't argue
> against it.

Rather than basing designs on poor second-hand information, maybe you
and the person who mentioned this idea should get up-to-date and read
the SQL Server storage engine architecture.

As of SQL Server 2005, blocks *are* checksummed with CRC32.  And, just
for the record, previous versions of SQL server performed a bit
flipping technique for every 512 bytes in the page header; it did
*not* waste a byte for every 512 bytes written.

> I think the way you would work is to have the smgr note the sequential value
> it found when it read in a page and then when it writes it out increment that
> value by one. Conveniently the pages would be 16 bytes shorter than an 8kb
> page so you have 16 bytes available with every buffer to note information like
> the last sequential tag the buffer used.

This proposed design is overcomplicated and a waste of space.  I mean,
we reduce storage overhead using phantom command id and variable
varlena, but let's just fill it up again with unnecessary junk bytes.

> That seems pretty unlikely. CRC checks are expensive cpu-wise, we're already
> suffering a copy due to our use of read/write the difference between
> read/write of 8192 bytes and readv/writev of 511b*16+1*6 is going to be
> non-zero but very small. Thousands of times quicker than the CRC.

Prove it.

-- 
Jonah H. Harris, Software Architect | phone: 732.331.1324
EnterpriseDB Corporation            | fax: 732.331.1301
33 Wood Ave S, 3rd Floor            | jharris@enterprisedb.com
Iselin, New Jersey 08830            | http://www.enterprisedb.com/