Andrew Gallatin wrote:
Garrett D'Amore writes:
>
> I'm working on this area of performance. If you aren't using jumbo
> frames, you need a card that supports LSO. The only option at the
> moment is "xge" (Neterion.)
>
Since GLDv3 will apparently never be stabilized and made available for
use by IHVs, it would be really nice if you could at least add {T,L}SO
support to GLDv2. This would get you at least one more 10GbE NIC
supporting LSO.
GLDv3 will stabilize... Please be patient. I'd rather spend the energy
working on stabilizing GLDv3 enough to be useful for IHVs than spending
it working on backporting LSO...
Btw, if there is some IHV that wants to get immediate GLDv3 support on
Solaris, they might consider looking at integrating source code for
their project into ON. Drivers that are integrated into ON can use all
GLDv3 features _now_. (This does not necessarily require the driver be
open sourced, but it does require that Sun have access to the code.)
If an IHV is interested in this course of action, please get in touch
with me.
> For all other NICs, the per-packet-overheads associated with 1500 byte
> frames are quite limiting at 10Gb. I'm spending a lot of time examining
> ways to improve this for Solaris Nevada, some of these may be backported
> to Solaris 10 in the future.
As an interesting data point, Solaris has the best 1500b *receive*
rate of any OS I've tested recently. Even without LRO, I'm able to
receive a single stream at nearly 7.5Gb/s (sent via TSO from a FreeBSD
or Linux host) on Solaris/amd64.
Single stream performance is expected to be good. Its when you saturate
multiple streams that it goes down hill fast. And of course, this only
applies to TCP.
I'm working on addressing this generally. Stay tuned.
-- Garrett
Drew
_______________________________________________
networking-discuss mailing list
[email protected]
