While we're on this subject I thought I'd bring up something that
MikroTik are doing with some of their Routerboards. Could be a nice to
have feature on future Soekris boards too!
Their RB450G, RB750G, and some other new boards incorporate dedicated
switch controllers wired to the onboard ethernet ports. What's great
about this setup is that port routing can be set in software. Ports can
be selectively routed to the switch controller for wire speed switching,
or to the CPU for L3 routing and/or software bridging. The best of both
worlds!
Wouldn't it be nice to have something like this on our Soekris boards?
On 06/12/11 05:43, der Mouse wrote:
>>> the ports are just ethernet interfaces so all your switching is
>>> being done in software if that's what you're asking.
>> Would that mean that a pair of machines transferring at full speed on
>> eth0 and eth1 would not reduce the available bandwidth for another
>> pair on eth2 and eth3?
>
> Depends on multiple factors.
>
> - Ethernet hardware. Sometimes there's a shared piece of hardware that
> can't handle full-bore I/O on all ports at once - I don't know
> whether this is true of the 5501 or not.
>
> - Bus bandwidth (this is, strictly, a shared piece of hardware such as
> mentioned above, but it's shared with, typically, a lot more than
> just the Ethernets). However, when doing a software switch, packet
> contents must be copied across the bus into main memory, then copied
> back again on transmission. Depending on the bus(es) involved and
> how DMA works, this may or may not be a problem.
>
> - CPU. The CPU has to look at each packet at least a little (typically
> just the Ethernet header, but that means servicing a cache miss, and
> probably some kind of shootdown when the DMA happened too).
>
> - Likely other factors I haven't though of.
>
> My guess - and it's a total guess - would be that bus bandwidth will be
> your limiting factor. But I'd have to measure it to be sure. If the
> other factors are capable enough, the limiting factor will be the
> network wire speed; this is the best case from your point of view, and
> for all I know might actually obtain for you.
>
>> If not, then how can that be achieved.
>
> Dedicated switches typically have custom silicon for the switching
> fabric. If you want full wire speed on a lot of fast ports at once
> ("fast" may include 100Mb and almost certainly includes Gb), you won't
> get it without custom silicon. But if your "full speed" is slow
> enough, you will be able to do OK with a software switch (but "slow
> enough" depends on the other factors).
>
> One thing you will _not_ get with a software switch is cut-through
> switching, packet forwarding where transmission starts as soon as
> enough of the Ethernet header is received to tell where the packet
> should go (if the destination channel can handle a send at the moment).
> Each packet must be fully received by the hardware before forwarding
> decisions are made. This does not, strictly speaking, impair
> bandwidth, but it does mean forwarding latency will be higher than with
> custom dedicated silicon switching hardware, which for some kinds of
> workloads is operationally similar to bandwidth limits.
>
> In short, it's a complex question. About all that someone not working
> closely with your situation can say with confidence is "it depends,
> you'd have to try it to be sure".
>
> /~\ The ASCII Mouse
> \ / Ribbon Campaign
> X Against HTML [email protected]
> / \ Email! 7D C8 61 52 5D E7 2D 39 4E F1 31 3E E8 B3 27 4B
> _______________________________________________
> Soekris-tech mailing list
> [email protected]
> http://lists.soekris.com/mailman/listinfo/soekris-tech
_______________________________________________
Soekris-tech mailing list
[email protected]
http://lists.soekris.com/mailman/listinfo/soekris-tech
