Okay, with that understanding I think I can give +1 to this project.
There should be guidance somewhere (whitepaper?) that Solaris bridging
is not expected to be ultra performant. In other words, you probably
don't want to use a Solaris host to replace a 10G switch, or with
multiple aggregated gigabit ports.
- Garrett
James Carlson wrote:
> Garrett D'Amore writes:
>
>> James Carlson wrote:
>>
>>> Fishing for +1s ... has anyone read this?
>>>
>>>
>>>
>> I'll reply later today. I need to check your materials for one thing --
>> I'm a bit surprised you intend to disable crossbow polling mode. The
>> rest of your proposal (without looking at the detailed changes) looked
>> good to me.
>>
>
> This was discussed with Nicolas Droux and Venugopal Iyer at length.
> The issue is that the hooks I have in the mac layer in order to
> intercept packets at a low level for bridging purposes are in a place
> where only interrupt-based packets are processed.
>
> I investigated adding hooks for the polled mode, and got a handle on
> what might be required, but the Crossbow folks suggested that I just
> disable polling for these ports. The reasons are several:
>
> - The data paths used for polling mode are diverse, complicated, and
> very performance sensitive. Adding hooks there would likely lower
> performance for the most critical high-end features, and would be
> complex as many of the functions involved have doppelgaengers.
>
> - Polling mode is used only for high-speed high-performance
> interfaces, and enabling bridging forces us turn on promiscuous
> mode (as a basic requirement of doing bridging), which in turn
> substantially impacts performance in many ways (not just the extra
> data, but also by causing performance features to be disabled as a
> side-effect). In other words, there's a fundamental conflict
> between these two goals: if you want to sling packets fast, you
> don't want bridging, and if you want bridging, you have to be
> willing to give up the upper end (>1Gbps) of performance.
>
> - The whole issue goes away when the Crossbow team redesigns the
> classifier components so that they're able to handle the
> functionality required for bridging. (We don't currently have an
> ETA on that, though it's not expected to be immanent, which is why
> the bridging project is going forward with the existing design.)
>
>
