Re: [Qemu-discuss] 4k Disk I/O issue

[halfdog]

Hi,

by morphin writes:
> Hello.
>
> I'm getting too much bad return from my customers at 4k random speeds. They
> claim that SQL databases works much slower than usual, and I can't find a
> solution. ...

Not sure if related, but I also observed very strange IO speed
behaviour, but did not check it again for long time.

Initial observation was, that cryptoloop sometimes was much slower
when disk IO was 4k aligned, while sector-aligned but 4k-UNaligned
was FASTER. This was unexpected: I assumed to have best speed with
4k IO and 4k memory page size.

Drilling it down, it showed that the cryptography stuff did not
matter: using a loop with offset only but no cryptography operation
showed the same behavior.

Also block device IO bus speed was NOT the real cause. Mapping
two RAM disks via loop over each other, pure memory page aligned
transfer was much SLOWER than 4k-unaligned.

Therefore I decided to continue disk partitioning with odd number
of 512-byte sectors as 4k alignment reduced the speed, but did
not search for the root cause.

hd