> -----Original Message-----
> From: Tom Lane [mailto:tgl@sss.pgh.pa.us]
> Sent: 10 May 2005 23:22
> To: Simon Riggs
> Cc: Bruce Momjian; Mark Cave-Ayland (External);
> pgsql-hackers@postgresql.org
> Subject: Re: [HACKERS] Cost of XLogInsert CRC calculations
(cut)
> That's awfully vague --- can't you give any more detail?
>
> I have seen XLogInsert eating significant amounts of time (up
> to 10% of total CPU time) on non-Intel architectures, so I
> think that dropping down to 32 bits is warranted in any case.
> But if you are correct then that might not fix the problem
> on Intel machines. We need more info.
>
> regards, tom lane
Hi Tom/Simon,
Just for the record, I found a better analysis of Adler-32 following some
links from Wikipedia. In summary, the problem with Adler-32 is that while it
is only slightly less sensitive than CRC-32, it requires roughly a 1k
"run-in" in order to attain full coverage of the bits (with respect to
sensitivity of the input). This compares to 4 bytes of "run-in" required for
CRC-32. So unless we can guarantee a minimum of 1k data per Xlog record then
Adler-32 won't be suitable. See the following two links for more
information:
http://en.wikipedia.org/wiki/Adler-32
http://www.ietf.org/rfc/rfc3309.txt
One other consideration would be that since CRC-32 calculations for Xlog
records occur so often, perhaps the CRC-32 routines could be written in
in-line assembler, falling back to C for unsupported processors. It would be
interesting to come up with some benchmarks to see if indeed this would be
faster than the current C implementation, since as the routine is called so
often it could add up to a significant saving under higher loads.
Kind regards,
Mark.
------------------------
WebBased Ltd
17 Research Way
Plymouth
PL6 8BT
T: +44 (0)1752 791021
F: +44 (0)1752 791023
W: http://www.webbased.co.uk
