Re: What exactly is our CRC algorithm?

At 2014-11-26 18:56:52 -0500, bruce(at)momjian(dot)us wrote:
>
> I would test it with fsync=off to remove the fsync delay. That will
> simulate an SSD or BBU controller.

The earlier tests were run with fsync=off.

I ran another round of the replay tests on the i7 server, this time
replaying 30GB of WAL segments.

master

16:08:03 16:16:23 08:20
16:19:33 16:28:03 08:30
16:32:20 16:41:17 08:57
16:42:42 16:51:22 08:40

8crc

16:52:11 16:59:48 07:37
17:01:07 17:08:30 07:23
17:22:16 17:30:11 07:55
17:31:29 17:39:06 07:37

These are just the results of the last four runs with each branch, but
the difference was consistent over many more runs.

To summarise: the slice-by-8 patch (a) increases pure CRC performance by
a large margin, (b) has a positive effect on reading WAL during replay,
(c) doesn't hurt XLogInsert in typical cases (which was the main worry).

-- Abhijit