On Wed, Apr 23, 2008 at 7:29 PM, erik quanstrom <[EMAIL PROTECTED]> wrote:
> >> just put it up on a tee: why not use aoe?
> >>
> >
> > The problems of disk I/O are largely a focus issue -- all this stuff
> > is pretty new and they focused on the network mechanisms first because
> > those were the ones where the competition has published the most
> > compelling benchmarks. The disk stuff will get tuned out and will
> > likely outperform network for I/O. As an example, 9P directly over
> > virtio beats NFS/TCP/virtio-net by 70% without cacheing or
> > optimization in 9P (which is usually the opposite case on
> > unvirtualized hardware due to cacheing and what not).
>
> i wouldn't think that you could tune out rotational latency. 8.4ms is
> pretty much forever when you're counting nanoseconds.
>
> since aoe can do wirespeed (120ms/s) on typical physical gige
> chipsets i would think it would have no trouble keeping up with
> spinning media. especially when not handicapped by having to
> actually stuff bits through a phy.
>
You on the wrong portion of the problem -- the disk solution they have
is effectively AOV (ATA over Virtio), you aren't going to do better by
putting a virtual network driver in between. They just have to tune
their userspace gateway for disk access -- they put a lot of work into
making the virtio<->tun/tap gateway really efficient and I think they
are just using the crappy Qemu block device at the moment. Once they
short-out the gateway between the guest-virtio channel and the
in-kernel block driver it'll be much faster than tunneling AOE over
the network device to the host.
Now - if you are talking about supporting an off-server CORAID storage
array -- then you should absolutely go AOE, but I think he was talking
about talking between guest and host partitions on his laptop in which
case you are adding extra layers for nothing.
-eric
