Hannu Krosing:
> did you also test this with fsync=off ?
Yes. No significant difference.
> I suspect that what you are seeing is the effect of randomly writing to
> the index files. While sequential write performance can be up to
> 80MB/sec on modern drives, sequential writes are an order of magnitude
> slower. And at your data sizes you are very likely to hit a
> CHECKPOINT, which needs to do some random writes.
Yes, from the server log I noticed that I hit checkpoints
too early and too often. I tried the astronomical value of
1000 for checkpoint_segments to not hit a single one for the
whole test run (copying 800 MB) -- even though that is no
good idea in practice of course.
It took even longer then. Probably because the server
created a lot of 16 MB log files (about 300 in my case)
which is presumly more costy (at least for the first run?)
than overwriting existing files. I am not too much into
that, though, since this is not a solution anyway on the
long run IMHO.
Thanks again.
Felix
