Very low disk performance on 5.x

[Poul-Henning Kamp]

In message <20050509203849.J20942 at delplex.bde.org>, Bruce Evans writes:

>> The solution is to re-engineer the way that I/O buffers pass through
>> the kernel and only assign KVA when needed (for doing software parity
>> calculations, for example).
>
>How would it yield anything except complexity and negative performance
>benefits on machines that don't need it?

The crucial point here is that disk controllers do not need it.

Most disk-controller hardware is able to do scatter/gather these
days and we cannot exploit that because we insist on disk-I/O
happening from sequential KVM addresses.

The silly thing about us insisting on KVM mapping in the first
place is that there is practically no modern hardware that needs
that.

ATA in PIO mode.  RAID5 parity calculation.  GBDE encryption and
that is about it.

Of course we need to cater for those in the new world order too,
but it seems only fair to put the overhead in front of those three
since they are already slow operations.


-- 
Poul-Henning Kamp       | UNIX since Zilog Zeus 3.20
phk at FreeBSD.ORG         | TCP/IP since RFC 956
FreeBSD committer       | BSD since 4.3-tahoe    
Never attribute to malice what can adequately be explained by incompetence.